Sage Journals: Discover world-class research

Abstract

Cancer care often involves making complex medical decisions within a challenging environment: a balkanized medical system of many specialists, information overload and obsolescence, limited time with doctors, and siloed data. New AI tools can enable patients and caregivers to more actively participate in treatment decisions. For example, consider receiving a complex scan report in a patient portal. While today, a patient may face confusion in interpreting a complex diagnostic report, a personalized generative AI agent could help by translating the scan report into language a patient can understand, and even contextualizing it within a patient’s personal health history and clinical evidence/guidelines. By providing an understandable version of the report and the clinical context of their test results, patients and their caregivers can engage more fully in decision-making with their oncology team. As described by Clayton Christensen’s “The Innovator’s Dilemma,” industry incumbents typically do not adopt disruptive technologies for fear of cannibalizing existing revenue streams (such as the case of Kodak and digital photography). This leads incumbents to serve their existing customers with the same value proposition, while ignoring “disruptive innovations” that offer a new value proposition to underserved customers (such as the transistor radio, which enabled teenagers to take their music with them). This theory predicts that institutional health care will focus on adopting AI for incremental operational improvements (e.g., patient scheduling, scan interpretation, claims processing). In this review, we argue that a positive disruptive impact of AI in oncology can come if AI-enabled software agents are used to support patients and caregivers in seeking better outcomes through personalized care. We review existing gaps and challenges that patients face as they go through receiving a cancer diagnosis, selecting a treatment plan, and then living with cancer. We use this process view to map uses of novel AI technologies that can assist patients and caregivers. This allows us to describe how current AI technologies/research and policies need to evolve to support patients and caregivers.

Introduction: Engaged Patients and Caregivers Are Critical to Achieve the Transition to Personalized Medicine

Precision medicine has transformed oncology practice, especially in the advanced setting. In precision medicine, advanced diagnostic testing is used to stratify characteristics of tumors at a population level, which are used to match patients to optimized care guidelines (e.g., matching all patients with advanced Non-Small Cell Lung Cancer [NSCLC] and an EGFR mutation to an EGFR tyrosine kinase inhibitor [TKI]). While precision oncology has given us many new therapeutic options that benefit subsets of patients, these therapeutic options may provide patients with limited survival improvement,¹ while clinical implementation challenges have limited the number of patients who successfully start a medication that is matched to their tumor’s molecular signature.^2,3 Implementing existing guideline-based genomically driven care has exponentially increased clinician workload,⁴ and contributed to gaps in patient education that inhibit empowered and informed patient-driven decision-making.⁵

This combinatorial complexity of precision medicine contributes to the daunting mental load that cancer patients and caregivers face. The empowerment of patients and caregivers with software agents and information can create a sea change in the current medical system, placing patients and caregivers first and protecting them from supply-side failures, such as the following:

Cost: Cancer can be a long and costly journey for patients, with significant gaps in the end-to-end process, such as remission and aftercare. Currently, it is difficult to provide full price and efficacy transparency across the lifetime of the disease,^6,7 contributing to increased financial burden to patients.

Clinical Complexity: Oncology is a field that is becoming increasingly complex, and needs better communication. With multiple sources of information, it is crucial that all the relevant information is brought to the patients and their care team in a timely and understandable manner. Gaps and errors in the diagnostic process and the application of clinical guidelines can lead to significant numbers of patients not receiving optimized therapy.^2,3

Systemic Complexity: Care is steered by clinicians, who face increasing complexity of practice driven by growing care guidelines,⁸ the use of labor-intensive health IT systems, and complex payment models.⁹ These cognitive loads increase the challenge of balancing patient wants and needs during care.

By focusing on a tumor’s molecular profile instead of a patient’s holistic health and personal treatment goals, health care providers run the risk of placing patients on medicines that provide marginal improvements in response but that have high out-of-pocket costs and provide negligible overall survival or quality-of-life improvements.² Escalating out-of-pocket health care costs pose a financial burden to patients, especially those patients who are managing complex or chronic diseases and conditions: when patients do not understand the role of their medications or are unable to finance their own care,^9,10 they are unlikely to adhere to their medication plans. Similarly, adding complex precision medicine decision-making tasks (such as considering how therapy order or combination may improve overall survival) can help patients to preserve treatment options or avoid patients’ developing treatment resistance¹¹ at the cost of increased physician workload and strain on the system. While some molecular tumor boards and clinical trials have attempted these approaches,¹² they may be economically challenging to support at the scale of the entire American health care system.

We see the imperative in moving from population-level precision medicine to patient-centric personalized medicine as a critical step in closing trust and access barriers that patients face, and that severely impact the day-to-day lives of patients and clinicians. While new reimbursement pathways aimed at improving access to care navigation could help to make this level of care financially viable within well-resourced health systems,¹³ successfully addressing patient needs across the cancer care continuum would necessitate growing a large de novo workforce of navigators with specialized skills to assist with financial⁹ or clinical trial counseling.¹⁴ Rigorously constructed AI systems can be deployed at scale with a lower cost than human workers.¹⁵ If applied to care navigation tasks, agentic AI has the potential to bridge long-standing informational gaps that contribute to poor outcomes for patients. Examples of this include improving adherence to treatment protocols and the management of side effects by educating patients about their disease and treatment,⁵ and helping to address information gaps that degrade patient trust in the clinical trial system¹⁶ by helping patients to seek care that is aligned to their personal risk tolerance and survival and quality-of-life goals.

In this review article, we review how patients experience cancer care today, with a lens toward biomarker-driven precision medicine. By reviewing the research literature on clinical implementation gaps and patient/caregiver sentiment, we identify four areas where patients and caregivers routinely experience navigational gaps. We use these gaps to identify key navigational workflows where an AI-powered agent could support empowered patients and caregivers, and review existing computational tools (both AI and non-AI) that provide solutions to these workflow steps. These workflows include logistical care navigation (finding the best health care experts, coordinating communication between patients, caregivers, their medical team, and friends and family), clinical monitoring and educational support (test report interpretation, education about their condition and testing and treatment options, monitoring and guiding the patient through therapy to help avoid care gaps), and addressing care access barriers (supporting reimbursement, cross-specialty care, and trial access).

We conclude by reviewing the trust/safety, regulatory, and workflow integration challenges that patient-centric AI must address for successful deployment. The success of AI-enabled software agents will depend on both technical solutions to current weaknesses (such as bias, knowledge cutoff where large language models [LLMs] cannot accurately answer questions that depend on knowledge that has changed since they were trained,¹⁷ and hallucinations where LLMs generate unsupported “factual statements”¹⁸) and the thoughtful integration of software agents into existing patient support workflows (navigation, care coordination, and patient education) through patient-centered design and workforce education. The concerted action of patient advocates and policy lobbyists has driven disruptive changes in the American health care through actions such as the National Cancer Act in 1971, the Americans with Disabilities Act of 1990, the National Cancer Survivorship Act of 1996, ACT UP/AIDS Coalition to Unleash Power, and the Affordable Care Act. As health care recovers from the COVID pandemic and starts to invest in AI technologies, we have the opportunity to use lessons learned from past patient advocacy movements and the COVID pandemic to address the key gaps that patients encounter with precision medicine.

A Significant Opportunity for AI in Medicine is to Address the Challenges That Patients and Caregivers Face

Flaws in the structure of health care markets lead to financial returns (and thus power) accruing to large enterprises (such as health insurance companies, pharmaceutical companies, and hospital systems), at the expense of patients and caregivers.¹⁹ This creates powerful systemic incentives that can conflict with patients’ best interests. An example of this is the increasing cost of cancer medications. Surveys of patients indicate that patients are skeptical of the value of extremely high-priced oncology medications that provide marginal survival benefits.²⁰ However, the approval and pricing of these medications have direct and indirect systemic effects on patients:

Example of direct impact: To protect against escalating costs driven by high-priced medications, payers have adopted cumbersome prior authorization processes.²¹ These processes contribute to the oncology practitioner burnout²² and impact patients through the denial or delay of care.²³

Example of indirect impact: The financial rewards associated with bringing a new drug to the market, or expanding the market for a drug, incentivize the design of trials that will succeed. While this is superficially aligned with a patient’s interests by increasing the number of treatment options, trial sponsors may be incentivized to design trials that use primary endpoints that fall short of patient’s outcome preferences²⁴ or that contribute to financial toxicity.^25,26

Existing oncology AI tools have predominantly been developed by researchers working within the health care enterprise, who experience organization pressures to develop innovations that “sustain” their existing market. While the term “disruption” is used colloquially to describe new technical innovations and business models, Clayton Christensen’s Theory of Disruptive Innovation provides a more formal definition.²⁷ This theory distinguishes between sustaining and disruptive innovations, as shown in Figure 1: a sustaining innovation is a technological innovation that allows an incumbent technology to sustain its advantage in an existing market (with the same processes, value proposition, and business model), while a disruptive innovation is a technological innovation that allows a technology to enter new, previously inaccessible markets for an underserved customer (typically with a new value proposition and business model). Disruptive innovations typically reduce the cost of providing a service/technology by more than an order of magnitude (enabling new market entry), but may not have the performance at introduction necessary to be used in existing “high-end” applications.

Fig. 1.

While the term “disruption” is often used colloquially to describe new market entrants, Christensen’s theory of disruptive innovation²⁷ distinguishes between “sustaining innovations” (which drive improvements in existing markets, e.g., improving the efficiency of working in an EHR) and “disruptive innovations,” which create new markets through innovations that reduce the cost of delivering a service to currently un- or underserved consumers. (Fig. is reproduced from Kalbach, 2020,²⁸ under a Creative Commons CC-BY 2.0 license.).

The idea of using artificial intelligence to “disrupt” health care is not new; many oncology clinicians and researchers remember both the promises of how the IBM Watson platform could change medicine²⁹ and the limited reality of what it achieved.³⁰ While IBM Watson never achieved the accuracy and usability necessary to be a clinical companion, market forces also limit how institutional health care can adopt disruptive innovations. This can be seen in the ongoing dialogue about how health care will pay for AI, which focuses on fitting ambient AI applications into the existing “fee-for-service” reimbursement model.^31,32 In this model, the profitability of a hospital or health care practice is effectively determined by the productivity of billable procedures (the mix of procedure quantity and value) that a hospital/practice completes, and the overhead costs of the hospital/practice (what is the cost of running shared services, such as the electronic health record [EHR] or hospital call center). This model inhibits the adoption of technologies that improve patient experiences and outcomes,³³ unless the technologies are attached to reimbursement codes or governmental (dis)incentives.

Early experiences with generative AI have encountered this issue, since existing reimbursement pathways typically mean that an AI tool has to look similar to a “procedure” or “medical device.”^32,34 As such, early adoption has primarily accrued to diagnostic AI, or to “sustaining technologies” that provide a clear productivity-based financial benefit to hospitals such as AI scribes that improve physician productivity. While value-based care (VBC) reimbursement models, such as the Center for Medicare and Medicaid Services’ “Enhancing Oncology Model” can provide additional leverage for the adoption of technologies that improve patient outcomes, initial patient experiences have been mixed to slightly negative,³⁵ as VBC models are still centered on institutional health care and have not historically comprehensively incorporated patient’s experienced outcomes.³⁶ Ultimately, the current reimbursement model externalizes these costs to patients, in both high out-of-pocket costs and poor care experiences.

Given the challenges of health care reimbursement and health IT incumbency, is the most viable path for disruptive innovation to use generative AI to address patient’s information needs? If AI continues to improve and it becomes cost effective for patients to have on-demand 24/7 access to a doctor with generalist knowledge across all medical specialties, we can enable a number of patient education tasks that go underserved due to health care staffing pressures that have reduced the amount of time that oncology care providers can allocate to patient education.³⁷ The information needs of a person living with cancer change dramatically over the course of treatment, and vary given their personal and family health histories. Several studies^38,39 have used patient interviews to identify common phases that patients experience; Figure 2, reproduced from Hall et al.,³⁹ shows how patients’ support and information needs change over time. Patients with advanced or recurrent/refractory cancer will pass through these phases multiple times, with differing needs over time: if their cancer is staying well managed, they may seek education that can help them mitigate quality-of-life issues related to treatment,⁴⁰ while patients with progressive disease may be seeking education about treatment beyond the standard of care or clinical trials.¹⁴

Fig. 2.

Cancer experience map, reproduced from Hall et al.,³⁹ under a Creative Commons CC-BY 2.0 license. This map shows how cancer patient’s day-to-day experience, needs, and level of knowledge change over time. After an initial diagnosis, patients face a challenging learning process driven by the complexity of oncology, and the shock of receiving a diagnosis. Patient engagement typically increases over time, as patients’ knowledge grows, but can be hampered by factors such as adverse events or other life stressors.

Currently, the vast majority of patients turn to the internet to educate themselves about their conditions and care options: a 2021 survey that was conducted by the American Health Information Management Association Foundation⁴¹ found that up to 94% of patients use Google to answer health-related questions. As their condition changes and their learning needs evolve, patients and caregivers may move from generalist resources such as Google or WebMD’s Cancer Resource Center to more specialized web resources that educate them about genomics (including general resources such as MyHeritage’s Promethease or the Wellcome Genome Center’s YourGenome.org, or gene/biomarker-focused resources such as the Global Alliance for Genomics and Health's [GA4GH] BRCA Exchange⁴² or International Cancer Advocacy Network's [ICAN] Biomarker Collaborative), their medication options (such as the National Comprehensive Cancer Network's [NCCN] Guidelines for Patients), or potential clinical trial options.

However, searching for information about oncology care on the internet is time-intensive, prone to error and misinformation,⁴³ and is often inaccessible to patients and caregivers with limited English proficiency⁴⁴ or disabilities.⁴⁵ In addition, common oncology terminology such as “salvage chemotherapy” or “palliative treatment intent” can have negative connotations that may dissuade patients from pursuing care that they may benefit from⁴⁶ and can contribute to worse outcomes through the “nocebo” effect.^47,48

Early research applying LLMs to patient communication has shown that LLMs have the potential to communicate with patients in a highly empathetic manner.⁴⁹ However, specialization is needed for LLMs to perform in a trustworthy manner when interacting with patients. Moor et al.⁵⁰ proposed three necessary capabilities for “generalist” medical foundation models; here, we reframe these capabilities to focus on patient-facing navigation workflows:

Reasoning over complex medical knowledge: Oncology is a field defined by rapidly moving evidence⁸ that is communicated in a specialized language that can create a barrier to patient understanding.⁵ Generalist AIs can internally represent medical knowledge, and be trained or fine-tuned to communicate in either a clinician or patient-friendly manner.¹⁵

Need for multimodal reasoning: Foundation models are increasingly able to reason across multiple modalities of data; examples in medicine include MedFlamingo,⁵¹ which can reason over text and images, or Polaris, which can reason over text and speech.¹⁵ Patients routinely need to link multiple modalities of data, such as jointly contextualizing a pathology report and genomic testing results, or using their medical history to seek financial approval from an insurer.

Ability to flexibly learn and compose tasks into novel workflows: Although several task-specific AI models have been developed for oncology,^12,52,53 patients experience complex, multistep workloads. For instance, successfully starting a given treatment requires correct interpretation of diagnostics, identifying the optimal treatment (standard-of-care/trial), completing prior authorization/trial enrollment, and orchestrating numerous medical appointments. For a patient, success requires that all of these steps happened correctly; thus, we should be wary of point solutions that perform well on task-specific benchmarks but that do not succeed at end-to-end workflows.

In the rest of this section, we use the patient journey map as a guide and identify four thematic areas where patients and caregivers commonly encounter challenges in making fully informed decisions. While generative foundation models face challenges with hallucinations that must be resolved (which we discuss later in this article), generative AI technologies may provide an opportunity to meet navigation needs in a way that improves patient trust and reduces the complexity of navigating oncology care.

Treatment volatility and complexity

The rate of oncology drug development has accelerated dramatically in the precision medicine era. In the decade following the approval of the initial “precision oncology” medicine imatinib (2002–2012), the FDA approved 65 new oncology medicines.²⁴ In 2018 alone, the FDA issued 46 new oncology medicine approvals, which applied to more than 400,000 cancer patients in the United States.⁵⁴ There are now over 180 precision oncology treatments with more than 90 companion biomarkers.⁵⁵

While having more treatment choices benefits patients and oncologists, it also increases the complexity of decision-making. As a proxy, we see evidence of growing complexity in oncology practice guidelines: from 1996 to 2019, the mean length of site-specific NCCN guidelines grew by more than 700%, from 26 to 198 pages.⁸ A survey of physicians across specialties identified that 60% of physicians identify the length and complexity of practice guidelines as an inhibitor to both using guidelines in care⁵⁶ and patient education. A study by Kaupp et al.⁵⁷ identified that fewer than 50% of oncology patients are satisfied with the education they receive about their medication options, and a related study outlined how educational gaps contribute to reduced patient empowerment in the community hospital setting.⁵

Similar issues inhibit enrollment in clinical trials. While cancer patients who participate in a clinical trial are likely to have improved outcomes,⁵⁸ recent estimates show that only 7% of U.S. cancer patients participated in a trial⁵⁹ and 20% of oncology clinical trials fail due to low enrollment.⁶⁰ Some of these problems are because structural issues in trial design contribute to access disparities that bias trials toward enrolling patients healthier than average⁶¹ while excluding women⁶² and ethnically underrepresented patients.⁶³ While approaches that use AI and real-world data can help to design trials that are more likely to enroll,^63,64 patients encounter information and workflow gaps that contribute to them not enrolling in a trial:

Patient understanding of clinical research: The oncology clinical trial landscape varies widely, as some trials test a novel medicine for the first time,⁶⁵ while other trials aim to enable the use of mature medicines in earlier treatment settings⁶⁶ or in broader indications.⁶⁷ There is evidence that patients frequently do not understand the nuances of clinical trial design,¹⁶ which may contribute to patients abstaining from trials due to treatment preference or fear of randomization.⁶⁸

Trial matching: The growing complexity of trial inclusion/exclusion criteria⁶⁰ exacerbates existing issues that patients face when searching for a trial. Existing online resources for trial matching require a high degree of knowledge of medical jargon, computer skills (searching and data entry), and a comprehensive understanding of a patient’s entire medical history.^69,70

Trial enrollment: Even if a patient is willing to participate in clinical research, and is able to find a trial, they may fail to enroll. The process of connecting a patient with the correct staff who support the trial and completing all enrollment steps is highly laborious and adds inertia at a critical step in the process.¹⁴

Figure 3 depicts the decision-making process that patients go through when they are considering a clinical trial. At each step of the flowchart, we outline typical questions that a patient is likely to consider, and give an example reason that a patient may not proceed to enroll after that step.

Fig. 3.

Even when patients have trials that are a match, they face numerous barriers that can lead to them not enrolling. While technology solutions often focus on the problem of matching patients to trials,^52,71 it is critical to also focus on financial gaps and technology measures that can help patients to better understand the clinical research process and the pros/cons of participating in a trial.

Testing access and test interpretation

While the rapid growth in the number of therapeutics available gives new choices to patients and clinicians, taking advantage of these choices is far more difficult. Deciding on the correct therapy has caused an exponential increase in clinician workload, as it requires the integration of insights about both molecular biology and a patient’s treatment and overall health history.^4,73 As a consequence, real-world usage of precision medicines remains low, with estimates showing that up to 60% of eligible patients fail to receive optimized treatment due to the failure to both test biomarkers and to interpret those biomarker results correctly.²

While the suboptimal usage of precision medicines is partially driven by workflow issues that can be addressed through clinical informatic enhancements,⁷⁴ interpretation gaps cause approximately 25% of patients who receive biomarker testing to fail to start an optimized treatment.² Even if testing is properly ordered and returned on time, the correct understanding of biomarker testing results depends on having a comprehensive understanding of both pathology and molecular biology.⁷⁵ Biomarker testing results are typically explained to patients by their oncologist, who is unlikely to have received training in genetics/genomics.^76,77 This training gap is complicated further by the design of typical clinical genomic reports: these reports are written to give diagnosticians accurate technical information while complying with applicable regulation. While this is critically important, conventional diagnostic report formats are hard for patients to understand⁷⁸ and create workflow problems for oncologists, as reports are often not integrated into the EHR.⁷⁹

Even when correct biomarkers are tested, discordant test results can result when biomarker test results are generated by orthogonal methods (e.g., PD-L1 quantification via FISH vs. bulk RNA-seq⁸⁰) or next-generation sequencing (NGS) tissue sources.⁸¹ While the FDA’s companion diagnostic (CDx) pathway is used to jointly approve a biomarker test alongside a medication,⁸² there may be multiple—potentially discordant—CDxs for a single biomarker, or non-CDx laboratory developed tests (LDTs) that diagnosticians may choose. For an example of how this can cause biomarker status discordance that can impact decision-making, we can look at the biomarkers associated with the two pan-solid tumor approvals of pembrolizumab,^83,84 which can be tested with orthogonal diagnostic tests:

Tumor mutational burden (TMB): TMB is a measurement of the frequency of mutations in a tumor. In the initial pan-solid tumor approval (KEYNOTE-158, NCT02628067) for the use of pembrolizumab, the FoundationOne NGS assay was used as a CDx to assess TMB, with a cutoff of “High” set at 10 mutations per megabase⁸³ (mut/mb). TMB measurement is prone to variation across assays, due to sample preparation, panel size, and bioinformatic differences; in a study comparing TMB calculation across 16 NGS assays,⁸⁵ the interquartile spread of TMB measurements was between 3 and 8 mut/mb for samples with a median TMB near the TMB-high/low cut point (0 to 20 mut/Mb).

Mismatch repair deficiency (dMMR) and microsatellite instability (MSI): MSI is a form of DNA error where errors in the DNA mismatch repair pathway (MLH1, MSH2, MSH6, and PMS2 genes) lead to repeat expansion.⁸⁶ While dMMR was originally associated with germline mutations in these genes (Lynch syndrome), it is now understood that somatic mutations can also lead to a similar MMR-deficient tumor phenotype.⁸⁷ MMR can be assessed by immunohistochemistry (IHC) stains (and potentially NGS⁸⁷) while MSI can be assessed through multiple PCR tests and through NGS, with accuracy increasing with an increased number of microsatellites tested.⁸⁸ Despite the pan-solid tumor approval of pembrolizumab for MSI-high/dMMR tumors, the College of American Pathologists/American Society for Clinical Oncology (CAP/ASCO) guidelines for dMMR/MSI testing currently only cover colorectal, gastroesophageal, small bowel, and endometrial cancer.^89,90 When combined with cost and tissue availability, dMMR/MSI testing is uncommon in cancers not typically associated with Lynch syndrome, even though patients in these indications may benefit from pembrolizumab therapy.⁹¹

Continuous biomarkers such as TMB or PD-L1 are likely to pose a greater concordance challenge than categorical variants. Biomarker testing concordance is likely to be high in well-characterized genes (such as BRCA1/2⁹²) and in genes where there are single-locus changes that have therapy associations (e.g., KRAS G12C⁹³). However, biological factors (such as clonal hematopoiesis of indeterminate potential, CHIP⁹⁴), sample collection differences (such as sequencing a solid tissue biopsy vs. a “liquid biopsy”^81,95), and technical factors (such as bioinformatic filtering⁹⁶) can lead to discordance in the determination of categorical biomarker status.

Care fragmentation and time pressures

Patients are likely to see multiple clinical practitioners within their treatment, either within a single institution or across institutions. Cancer patients regularly need cross-specialty care for comorbidities,⁹⁷ and are likely to work with both oncologists and oncology-specialized advanced practice providers.⁹⁸ Coordination errors within a patient’s care team are a primary reason that treatment gets delayed,⁹⁹ and can contribute to worse survival outcomes for cancer patients with multiple chronic conditions.¹⁰⁰ While the broader adoption of EHRs has made it easier to share health data between practices, patients encounter integration gaps driven by emerging health information exchange technologies¹⁰¹ and conflicting data standards.^102–104

Process and clinician-to-clinician communication gaps also present an issue, especially during transitions in care between primary care providers (PCPs) and oncologists at the beginning and end of treatment. Gaps at the start of care can contribute to worse outcomes for patients with chronic conditions. Managing drug–drug and drug–comorbidity interactions becomes both more complex and more important as the therapeutic landscape in oncology shifts to therapies targeted at specific molecular pathways and immune signatures, requiring increased review by pharmacists and primary care practitioners.¹⁰⁵ Gaps at the end of active treatment can delay transitioning to survivorship⁴⁰ or palliative¹⁰⁶ care.

Time pressure exacerbates this problem. Cancer patients, especially those with advanced tumors, face an unusual paradox: while they spend approximately three hours every four days obtaining care,¹⁰⁷ they typically have only an average of 23 minutes for office visits with their oncologist in the United States.¹⁰⁸ While dedicated navigation supports can help to mitigate these problems and improve patient outcomes,¹⁰⁹ tools for shared decision-making are essential to help patients, caregivers, and all of the clinicians involved in their care—which may include oncologists, oncology nurses, surgeons, pathologists, psychiatrists—make informed decisions together.

Reimbursement and paying for care

Uncertainties about the cost of care influence patients’ choices throughout their journey. While American regulators have issued regulations in support of health care price transparency, their impact on cancer care has been limited because cancer centers are often not in compliance with these statutes⁶ and many common oncology therapies are not included in pricing lists.⁷ In addition, price transparency laws cover hospitals but not care given in ambulatory settings, or diagnostic tests processed by nonhospital laboratories (such as commercial genomic panels).

Patients and oncologists may choose to forgo diagnostic testing due to concerns about reimbursement¹¹⁰ or overall cost,⁹¹ as the cost of large clinical NGS panels can approach $5,000.¹¹¹ This cost becomes a significant consideration to a patient who is choosing between an array of diagnostic testing panels (e.g., selecting between multiple providers of a tumor/normal matched sequencing panel). Patients and oncologists may struggle to accurately assess the value of a test, given that the value depends on both the sensitivity of the test for detecting key biomarkers coupled with the likelihood that a patient will be matched to optimized treatment after testing.¹¹²

The cost of care also has a significant impact on the treatments that patients pursue. Recent estimates place the monthly out-of-pocket costs of an American cancer patient at $180–2,600 per month.¹¹³ These costs can rise significantly if a patient is on a medication with a high copay; the FDA’s Project Optimus has recently drawn attention to sotorasib^114,115: sotorasib costs approximately $21,000 per month, with patients reporting out-of-pocket costs of up to $3,000 per month. While patients may be able to access programs that reduce out-of-pocket costs, such as free medication and medication copay assistance programs, these programs require additional administrative work, which some hospitals support through financial navigation and counseling programs.⁹ However, these programs are most likely to be deployed in well-resourced academic hospitals, which are less likely to serve patients facing financial distress.¹¹⁶ Patients considering clinical trials face similar concerns, driven by uncertainty over insurance coverage of clinical trials.¹⁴

The Solution Concept: Applying AI Through Software Agents to Enable Patients and Caregivers to Be Copilots

If we apply Clayton Christensen’s theory of disruptive innovation²⁷ and Moor et al.’s⁵⁰ three common patterns for generalist medical AIs to the patient challenges articulated earlier, we can envision an AI-enabled patient-centric oncology model where well-informed personal software agents will work on behalf of patients and caregivers to improve patients’ and caregivers’ experiences and patients’ outcomes. Low-cost software agents could help patients and caregivers to both set and achieve well-informed care goals (including outcome, quality of life, and financial factors) by providing the information and workflow assistance needed to help patients choose and integrate health care and supportive services. In this way, AI-powered consumer-facing information and advisory services, can augment and support a patient’s existing medical team.

It is useful and important work to pursue AI-enabled solutions that will try to fix the many systemic breakdowns and opportunities for process improvement within the incumbent health care industry institutions. Since the health care system is composed of many humans working at a vast scale, human factors contribute to the challenge of reforming health care. Beyond workload and education issues, factors such as preference/attitude asymmetry,^16,40 bias,¹¹⁷ provider training gaps and discomfort,¹¹⁸ and relationship comfort,¹¹⁹ make it difficult to redesign the clinical trial, care delivery, and reimbursement processes in a patient-centric manner. The careful design of health care information technology that centers patients’ and caregivers’ needs can improve care outcomes.¹²⁰

To date, the development of benchmarks for medical question/answer tasks^121–123 has driven rapid improvement in the accuracy of knowledge of generalist medical AIs. However, these benchmarks do not necessarily translate to clinical workflows, which may involve multiple steps and involve reasoning over private data stored in an EHR.¹²⁴ To maximize the utility and safety of a generalist AI that patients might use, we need to develop a good understanding of the tasks the generalist AI might be used for, and characterize what a good answer might look like. To further flesh out this solution, we use the “jobs to be done” framework,¹²⁵ which was developed by Christensen as a complement to the theory of disruptive innovation.²⁷ Roughly, this framework describes situations where a customer “hires” a product to complete a task.

As an example, consider a cancer patient getting a scan to determine the progression of the disease. Patients commonly experience “scanxiety,” which is distress or anxiety occurring before, during, and after cancer-related imaging or scans.¹²⁶ Since the “information blocking” prohibitions enacted by 2021's 21st Century Cures Act require that patient notes (including imaging reports) are released promptly to patients via a patient portal, there is evidence of a significant increase in the number of patients who review their imaging reports.¹²⁷ While current guidelines for imaging reports focus extensively on the technical detail and structure of the reports, they typically do not focus on how patients may interpret the report, which can lead to additional patient anxiety.¹²⁸

Earlier this year, one of the authors received a pathology report that he struggled to interpret due to the complex anatomical terms and cancer-specific shorthand. He wanted to test whether generative AI could help. He “hired an LLM” by copying and pasting his full pathology report into ChatGPT with a prompt to “translate into seventh-grade English,” as has been described in the literature.¹²⁹ The LLM helped to make the report understandable, and a review of the transliteration by several doctor friends confirmed that the AI translation was accurate. The author was put at ease: he understood what the report said. That said, he still needed more expert interpretation from his oncologist to know how good or bad he should feel about the results, and the process involved a manual step of copying his medical data into ChatGPT.

While large technology companies are developing high performing general purpose generative AI models such as Microsoft/OpenAI GPT family, Google’s Gemini, the Anthropic Claude model, and Meta’s Llama, their business models will steer them to offer services that are viable as products for a billion or more customers. Thus, cancer care is likely to be too narrow of a market for them to develop specific tools, limiting accuracy and thus utility.¹³⁰ Health care-specific LLMs have advanced rapidly in both capability and accuracy, through the progression from early text-focused models such as ClinicalBERT,¹³¹ BioGPT,¹³² and GatorTron,¹³³ to recent multimodal models such as MedFlamingo,⁵¹ Polaris,¹⁵ and MedGemini.¹³⁴ However, they will still require customization to converse with patients of disparate knowledge levels, and integrate with medical systems of record (e.g., the EHR).

In the remainder of this section, we map the challenges that patients face to “jobs to be done,” and highlight existing point solutions that map to these tasks. We aim to augment and extend other recent reviews (e.g., Bitterman et al.¹³⁵) by highlighting the potential benefits to patients of using AI to help automate these tasks, the prior art (both models, services, and evaluations), the data sources that these algorithms might be trained on, the health care technology integrations that these tools may need, and the task-specific risks that AI operating in these areas may encounter. A more general discussion of the regulation and validation is included in the following section of this article.

Enable 24/7 support for patients through AI-supplemented care navigation services

Cancer patients need multidisciplinary care: a core care team will likely include a surgeon, a radiation oncologist, a medical oncologist, a pathologist, and a radiologist. Patients can also benefit from working with a social worker,¹³⁶ a mental health specialist,¹³⁷ and a pharmacist.¹³⁸ Working with a large multidisciplinary team can cause care fragmentation, so software agents and AI tools can help patients to set treatment goals, share insights about their care with their medical team, and advocate for services that help improve their psychosocial state and minimize time toxicity (e.g., access telemedicine services wherever appropriate¹³⁹). Table 1 describes example jobs-to-be-done that assist patients with care navigation.

Table 1.

Patient Navigation “Jobs-to-Be-Done” That Can Benefit from Agentic AI

Jobs to be done	AI enablement
Find a doctor: When I am searching for a care team, I want to select clinicians aligned with my care (Pigott et al., 2019) and financial (Doherty et al.¹¹⁶) goals	Websites that provide this service include Anova Enterprises, MediFind, ExpertScape, Chaim Medical Resources, and Zocdoc
Communicate with friends and family: Help me coordinate my interactions with my team of family, friends, and health care practitioners to maintain a positive attitude and get help for day-to-day living (Hohmann et al.¹⁰⁰)	Models: Polaris¹⁵ Talk2Care¹⁴⁰ Website examples: PostHope, CaringBridge, Inspire, WiTT (We’re in This Together-put out requests, make a wish), Facebook private groups Evaluations: LLMs for patient/provider messaging at Stanford Hospital¹⁴¹
Communicate with my medical team: When I have a doctor’s appointment, help me maximize my time by guiding me and my caregivers with information and the questions to ask, and capturing important information¹⁶	Models: AIME¹⁴² Evaluations: Ambient scribes at Kaiser Permanente medical group¹⁴³ Commercial products (see discussion in Bongurala et al., 2024): Abridge¹⁴⁴ Amazon Web Services HealthScribe Ambience Healthcare AugMedix DeepScribe Nabla Nuance Communications Dragon Ambient eXperience (DAX) Suki
Communicate with peers: Help me connect with peers who have been through what I’m going through to give me peace of mind and receive support, love, encouragement, and hope¹⁴⁵	No AI tools for peer-to-peer navigation have been reported. The importance of peer support being well integrated into information systems for cancer patients is discussed by Gustafson et al.¹⁴⁶ Website examples: Facebook private groups
Help me educate my caregivers on key roles (e.g., proxies and backups) by creating, sharing, and managing the role descriptions¹⁴⁷	website example: Triage Cancer

We expect that tools in this space are most likely to have the “form factor” of a personalized software agent, a la ChatGPT. While tools in this category will rely on curated medical knowledge, we defer a detailed discussion of this to the following subsection. The successful implementation of these tools is likely to depend on the following integration points:

Communication channels: If properly integrated with secured connections between institutions, AI agents can enable more fluid communications between patients, caregivers, and clinicians (e.g., oncologist, pathologist) and clinical services (e.g., diagnostic testing laboratories). Early experiments have shown that AI agents have the potential to reduce patient-to-doctor communication latency, but careful implementation is necessary to ensure that AI-in-the-loop does not increase clinician message review burden.¹⁴¹

System navigator role: Existing navigation services help patients and caregivers to assess information quality, interpret information, and communicate with their medical team and advocate for their needs.¹⁰⁹ AI implementations can help to implement a “system navigator” role by using techniques such as chain-of-thought prompting (asking a model to explain how it arrived to a result) or ensembling (integrating multiple responses) when integrating LLMs into patient-facing workflows. These techniques can improve the accuracy of model response,¹⁴⁸ while also providing personalized feedback to patients on how to restructure their questions to avoid confusion and frustration.¹⁴⁹

Personalized communication preferences: Evaluations have shown that LLMs can achieve higher perceived empathy than clinicians.⁴⁹ By building off of existing work and developing novel prompting/ensembling/fine-tuning strategies (e.g., as demonstrated by Mukherjee et al.¹⁵), it is possible for patients to choose from a menu of AI agents aligned with varying communication preferences, or to even potentially design the personality of their software agent. This will allow for information, matching, and presentation to be customized to the patient/caregiver, which can improve outcomes related to patient education and satisfaction.⁵ Implementations will need to be conscious of language/dialect-centered biases that can drive biased responses that contribute to poor patient experiences and outcomes.¹¹⁷

Engagement: While much attention to date has focused on web-based AI services, socioeconomic and computer literacy barriers may limit access to web-based tools. To mitigate this, AI can be integrated with other communication channels such as audio,¹⁴⁰ video,¹⁵ or text messaging. While using additional communication channels can help to mitigate biases that restrict equal access to medical AI, these approaches are imperfect and should be augmented with additional communication approaches and social support.¹⁵⁰

Real-time patient support tools face interesting safety and accuracy challenges. While some workloads are unlikely to embody significant safety risks (e.g., searching for doctors, health proxy planning), patients and caregivers may use care monitoring and care-team communications tools during emergent situations, such as a patient experiencing a treatment-related adverse event. While there is evidence that digital platforms can improve the management of critical health events such as adverse events,¹⁵¹ if an AI tool provides incorrect guidance or neglects to inform a health care practitioner of the patient’s adverse event concerns, this would cause clear risk to patients. Tool development and assessment are starting to place a deeper emphasis on safety and trust that goes beyond model accuracy,¹⁵ but a further characterization of failure modes for AI-enabled patient communication platforms is necessary.

Empower patients in evidence-based treatment decision-making through the use of AI and health care interoperability technologies

Some of the most lauded example use cases for AI in medicine are minimizing diagnostic errors¹⁵² and improving treatment selection performance.¹² AI can examine a patient’s tumor DNA and pinpoint the genetic mutations causing the cancer with the goal of suggesting a personalized, evidence-based treatment strategy directed at the patient’s molecular profile.¹⁵³ Furthermore, generative AI can help doctors and health care providers to accurately predict patient outcomes.¹⁵⁴ Learning from large data sets of patients, it can uncover patterns linked to specific outcomes and help make informed decisions. Table 2 describes example “jobs-to-be-done” that support patients in understanding their diagnostic and treatment options.

Table 2.

Clinical Decision and Evidence Interpretation “Jobs-to-Be-Done” That Can Benefit from Agentic AI

Jobs to be done	AI enablement
Manage my health data: Help me capture my health data in a central place for convenience of access, and help me to compare it with reference data sets to enable customization of treatment uniquely to me and progression monitoring^5,78	No AI-enabled patient health records have yet been described. For a discussion of patient health records, see Brands et al.¹²⁰ and Cahill et al.,¹⁵⁵ For a discussion of data interchange via SMART-on-FHIR, see Mandel et al.¹⁵⁶
Monitor my care: When I am selecting therapy, monitor my care to make sure I’m consistently following the best-known practices¹⁵³ and proactively managing side effects	AI Systems: SAGE¹⁵⁷ Commercial products: OpenEvidence / ClinicalKey AI Atropos Health / ChatRWD Evaluations: Diagnostic accuracy of symptom monitoring apps^158,159 Accuracy of GPT vs. NCCN guidelines^130,160 Answering real-world clinical questions using LLMs¹⁶¹
Select and access tests: Help me understand and access the diagnostic tests that are most valuable to me across the full range of test options¹¹¹	Evaluations: ChatGPT for radiology referral¹⁶²
Interpret test results: Help me interpret test report treatment options by translating them into language personalized to my language skills⁵	AI models / systems: TEMED-LLM¹⁶³ VarChat¹⁶⁴ Mastermind¹⁶⁵ Evaluations: Assessing ChatGPT for explaining radiology reports^166,167 GPT prompt design for extracting data from pathology and ultrasound reports¹⁶⁸
Develop my treatment strategy: Help me build an overall treatment strategy that can evolve as my condition evolves by providing a prioritized list of testing, treatment, and clinical trial options with a transparent and causal logic behind treatment options¹⁵⁴	AI models / systems: PACMIDS¹² xDECIDE⁵⁴ CaML¹⁵⁴ LORIS¹⁶⁹ PurIST¹⁷⁰ Evaluations: LLMs for treatment decision support¹⁷¹

Use cases in this category rely heavily on the ability to accurately reason about complex medical tasks. This drives a strong dependence on having up-to-date medical data: Patients and caregivers will access continuously updated medical knowledge bases (such as WebMD, but curated to meet the specific needs of the rapidly evolving field of oncology) to make it easy to answer questions (including quality-of-life side effects), and provide the sources and experts to bring to their clinicians. These use cases are likely to harness integrations focused on:

Evidence databases and generation: While generalist models have reached accuracies above 60% on clinical Q&A benchmarks, techniques such as retrieval augmented generation (RAG) can improve accuracy in complex fields such as oncology by allowing models to use clinical guidelines when answering questions.¹⁷² While RAG is useful when high-quality clinical evidence exists, a recent benchmarking article demonstrated that evidence generation techniques are necessary to accurately answer questions when evidence is ambiguous.¹⁶¹ Some avenues for integration include the following:

-Omic databases: As clinical practice shifts from single biomarker tests to many-gene panels/whole exome/whole-genome sequencing, large genomic databases (e.g., The Cancer Genome Atlas [TCGA] or American Association for Cancer Research's Genomics Evidence Neoplasia Information Exchange [AACR GENIE]) can help to provide contextualization of -omic findings (e.g., gene upregulation), including normalizing for various biases or helping to identify pharmacogenomic signals.¹⁷⁰

Observational registries with patient-reported outcomes: Patients and caregivers will access observational registries to see the experiences of “patients like me” who are taking their treatment or tests they are considering. Each patient is an experiment, and we should be running small experiments and scale what works.

Information notifications: Given the integration into evidence knowledge bases and evidence generation tools, agentic services have the potential to continuously read all the medical research literature. This can be leveraged to provide patients with updates, including alerts, on new medical evidence, clinical guidelines, diagnostic tests, treatments, and trials that are important for them, and summarize it at a level they can understand. Approaches such as these can be helpful for mitigating care gaps following changes in the standard-of-care for an indication,² careful deployment is needed to avoid alert fatigue and the degradation of patient/physician trust.

EHR data cleaning: Many clinical/clinic-adjacent tasks require the accurate extraction of diagnosis, treatment plan/history, and biomarker status from the EHR. While demographic and laboratory values are often straightforward to extract from the EHR, many values are stored in textual records. AI-based services will need to perform well at working with data that are stored textually in the EHR; early research shows that AI services can clean and normalize data from the EHR at accuracies that are similar to manual curation/chart review by experts.¹⁷³

Applications in this category face the strongest safety risk, as the decisions surfaced by an AI tool are likely to directly impact the treatment that a patient receives. Multiple different regulatory pathways cover AI algorithms in this area. The integration of the AI tool into health care will influence whether it is likely to be considered an AI Clinical Decision Support tool,¹⁷⁴ AI-powered software-as-a-medical-device,¹⁷⁵ or be covered under non-software pathways, such as the LDT regulations. While these three pathways are all regulated by the FDA in the United States, they are seen as presenting different risk levels and thus have differing validation requirements before clinical use.

Streamline reimbursement, care coordination, and trial access by leveraging agent technologies

Access to financial assistance programs and clinical trials often involves complex processes and additional paperwork. By using agents to help automate these processes,¹⁷⁶ we can reduce the time and financial burden that patients face, while also closing understanding gaps that stand in the way of truly informed consent.¹⁶ This will help to reduce financial and time toxicity, while also enabling more patients to access clinical trial.¹⁴ Table 3 shows example “jobs-to-be-done” that can support patients with reimbursement and care access workflows.

Table 3.

Reimbursement and Access “Jobs-to-Be-Done” That Can Benefit from Agentic AI

Jobs to be done	AI enablement
Manage consent forms: Help me complete institution-provided consent forms, with a focus on highlighting possible implications and helping me to negotiate patient-friendly alternatives¹⁶	Evaluations: “Consent-GPT”¹⁷⁷ GPT for trial consent readability¹⁷⁸
Learn about and access emerging treatments: Help me to be educated about novel tests, clinical trials, and expanded access, and if I choose to participate, help me navigate test/trial/study enrollment¹⁴	AI models and systems: Zero-shot system for matching⁷¹ PRISM¹⁷⁹ GPT for trial I/E criteria structuring¹⁸⁰ TrialGPT¹⁸¹ Evaluations: Commercial LLMs used for trial matching⁵²
Manage finances: Help me minimize out-of-pocket costs by accessing patient assistance programs¹⁸²	AI models and systems: ECLAIR¹⁸³
Share my health data: Help me with data sharing across institutions⁴⁰ and provide options to contribute my data to research that will benefit others¹⁸⁴ Also see discussion of “Manage my health data” in Table 3	No AI-specific models/evaluations yet reported Commercial patient health records: Andaman 7 Cancer Commons/xCures Ciitizen/Invitae Consuli Navio PicnicHealth

Since these use cases are primarily nontreatment and aim to reduce time/financial toxicity to patients, these use cases generally present a low patient safety risk (with the notable exception of the clinical trial matching use case, which is similar in risk to the diagnostic/treatment selection workloads). We expect that use cases in this category will place an emphasis on data normalization (e.g., helping to avoid the overhead of repeated patient data entry into consent forms) and cross-system integration. Given that many of these workflows may lead to health data being sent to a third party (e.g., records may be sent to a patient assistance program, a clinical trial site, or a third-party patient health record), they face a unique risk of inadvertent release of a patient’s health data. As such, it is critical that these tools manage patient data in a way that both complies with HIPAA’s Privacy Rule by ensuring that patients are adequately informed and consent to health data sharing, and also ensuring that patients can access their health data in an unencumbered manner under the scope of both HIPAA law and the 21st Century Cures Act’s “Information Blocking” prohibitions.¹⁸⁵

Call to Action: A Roadmap for Using Data and AI to Enable Democratized, Patient-Led Care

Health care AI is at a critical movement. Using AI as an automation tool has the potential to dramatically reduce the cost of providing accurate, well-contextualized health care information and advice. As the cost of providing personalized health care services drops, new business and engagement models become feasible and sustainable. In health care, this can shift power to consumers (patients and caregivers), who can receive better outcomes as a result of care that is personalized to their goals, values, and history, while creating a learning health system where care improvement opportunities are identified proactively and patients and doctors can learn from the experiences of “patients like me.”

However, previous solutions focused on using data and AI to bring forward personalized care have encountered obstacles. These obstacles still need to be addressed for new health care AI:

Data privacy and legal responsibility for AI: The regulation around health data privacy is complex and patients frequently have misunderstandings about their rights under existing law such as HIPAA.¹⁸⁶ Likewise, it is currently unclear whether physicians or AI providers will be liable for lawsuits relating to clinical actions taken in response to the output of an AI model.¹⁸⁷ Given that HIPAA breach and malpractice costs add significantly to the cost and risk of care, it is necessary to achieve solutions that balance the privacy and malpractice risk of AI.

Value/business models: Health care reimbursement in the United States tends to favor IT services (whose adoption may be incentivized through regulation¹⁸⁸) and clinical procedures that map to the fee-for-service reimbursement model. While diagnostic AI tools may map well to the fee-for-service model, ambient and navigatory AI services are a less obvious fit. It is possible for existing VBC pathways to grow to enable the reimbursement of AI offerings, but novel reimbursement schemes may be needed.

Interpretability and trust: Sometimes the content generated can be challenging to interpret. The difficulties in understanding an algorithm’s decision-making process will impact patients’ and providers’ trust and decision-making. People are concerned about relying on AI systems that may hallucinate¹⁸ or otherwise generate incorrect results. Ensuring transparency is crucial to addressing biases and inaccuracies, fostering trust among providers and patients.

Given the magnitude of the benefits of personalized care—especially better patient outcomes, a health care revolution leveraging AI-enabled software agents to personalize care is possible. We believe that advocacy along two avenues is necessary to accelerate this revolution: (1) developing collaborations that leverage the unique strengths of health care providers, new technology entrants, and technology platforms; and (2) a political movement that ensures that health AI regulation meets both patients’ safety and access needs.

Achieving ubiquitous distribution of AI-enabled consumer health services through cooperation between health care, technology startups, and technology platforms

Much investment in the development of novel AI technologies is driven through the “hyperscaler” technology platform companies, such as Google/Alphabet, Amazon, Apple, Meta, and Microsoft. The products of these companies are data and services, which they sell through their various platforms. In health care, the platform services developed by a technology platform provider are typically not used at the point-of-care: a health care institution must build its own clinical informatics offering using the hyperscaler’s platform.¹⁸⁹ Therefore, the technology giants are likely to develop generalist tools for health care that serve a billion or more customers. Given their existing investments in AI, coupled with their key resource advantages (capital, technical staff, and computational), they can develop state-of-the-art health care foundation models, both in specific modalities (e.g., GigaPath¹⁹⁰) and for generalist tasks (e.g., MedGemini, Saab et al.¹³⁴). Emerging startups will build focused AI-enabled tools using the generalist AI platforms.

While platform companies can benefit from scale, the converse is also true: developing a patient/clinician facing tool to provide information and advice on a specific cancer will be too small of a market for them to want to serve themselves,¹⁸⁹ especially if most of the value accrues to the organization delivering the health care service, instead of the technology.¹⁹¹ As a consequence, integrating AI into the point-of-care will need to be done by other parties. For instance, implementing LLM-based trial matching has been done by hospital-affiliated academics⁷¹ and technology startups (e.g., Triomics OncoLLM for trial matching⁵²). This exposes a safety/access dichotomy that is similar to the discretion provided to hospital laboratories under the new FDA final rule regarding the regulation of “LDTs”¹⁹² : while analyses developed within a specific hospital may be safer because they are developed with greater understanding of the hospital environment, this also can restrict access to necessary developments, especially for patients with rare diseases, where there may not be local clinical expertise.¹⁹³

To allow all patients to benefit from health AI, regardless of resources or clinical setting, we will need to identify approaches that allow these tools to become ubiquitous. This could take numerous forms, between open science collaborations with the goal of openly developing and disseminating best practice methods standards (e.g., GA4GH¹⁹⁴) reforming health care reimbursement to lower the barrier-to-entry for health care technology startups,^31,32 or coming up with novel partnership models between hospitals, technology companies, and potentially third parties such as pharmaceutical or diagnostic companies. Another form could be building on the consumer services of the technology hyperscalers: they are likely to support new service providers who build on top of their platforms, for example, for supporting patients and caregivers with a particular cancer, such as lymphoma. Since these new service providers will use the hyperscaler’s cloud services, the hyperscalers may want to invest in these emerging service providers to increase use of their IT services, a dynamic seen in the 2010s with bioinformatic service providers building on the major clouds.

Ensuring robust and rapid adoption of AI-enabled software tools that touch patients directly through sensible regulation

The adoption of personal AI-enabled software agents brings the opportunity to dramatically improve the accessibility of high-quality health care, especially in lower resource settings, where there are fewer human experts and more language and health literacy barriers.¹⁹⁵ However, the current regulatory environment appears to favor the creation of regulatory regimes that may slow the adoption of patient-facing AI applications,¹⁹⁶ by either favoring “lower risk” back office applications¹⁹⁷ or by enacting practice guidelines that may create barriers that may limit the adoption of AI technologies.¹⁹⁸

It is worth drawing comparisons between the current moment for AI in medicine and the recent history of the adoption of biomarker testing in oncology. A recent survey of oncologist attitudes toward AI conducted by Hantel et al.¹⁸⁷ indicated that 53% of oncologists surveyed had not received training on the use of AI in health care and that 93% felt that they needed additional education. These numbers are remarkably close to the results of a 2023 survey covering UK oncologists’ attitudes toward NGS⁷⁷ that indicated that 39% of oncologists had not received training in interpreting NGS results and that 93% felt that they needed additional training in genomics and genetic counseling.

Ultimately, the health care industry and regulators have an ethical imperative to proceed wisely while integrating the perspectives of the patients and caregivers who consume the health care system’s services. Current regulatory proposals put a very strong emphasis on safety.¹⁹⁹ While validating the safety of novel AI tools is critical, regulation should balance the harm of inaccuracy with the harms of delaying patient access to helpful tools,²⁰⁰ and be mindful of how regulation can create barriers to market entry at the cost of consumers, such as the Epic/Cerner EHR duopoly resulting from the HITECH act.¹⁸⁸ To maximally deploy health AI in a way that improves patient-centricity through personalization, the health care system needs to chart a pragmatic course to deploying AI:

Assert patient power as partners: Incumbents in the health care industrial complex (pharmaceutical companies, provider, payers) have a lot of power to steer the regulation and thus adoption of AI tools. When clinician attitudes diverge from patient attitudes, this can create harm or perpetuate existing challenges. A concrete example of this can be seen in calls for “transparency” in AI models: in the aforementioned 2024 survey of oncologist attitudes toward AI,¹⁸⁷ 84% of oncologists felt that AI models needed to be explainable to oncologists, but a drastically smaller percentage (23%) of oncologists felt that AI models needed to be explainable to patients. Moving forward with a definition of transparency that privileges clinicians over patients runs the risk of perpetuating existing information gaps that contribute to poor patient experiences and outcomes.^5,14,57,77 We can mitigate these issues by including patients when defining practice guidelines and regulatory processes.

Define validation processes that center meaningful patient outcomes and populations: To reduce accuracy and bias concerns with medical AI, researchers have advocated for the use of clinical trials²⁰¹ and implementation science centers²⁰² to evaluate the effectiveness of AI models in clinical use. These approaches are pragmatic and will help with the safe adoption of clinical AI. However, these processes do not automatically protect against bias or ensure good-quality evidence. It is important that these approaches take care to protect against bias by defining representative study populations⁶³ and using study endpoints that go beyond correctness and focus on meaningful outcomes to patients.²⁴

Be mindful to avoid creating regulation that inadvertently limits access: Earlier in this section, we noted how the HITECH act contributed to the rise of an Epic/Cerner EHR duopoly.¹⁸⁸ Regulation can create barriers to market entry in a number of ways, with two common modes being regulatory capture that privileges existing solutions in the market, or through compliance requirements that drastically increase the cost of introducing a new solution. Taking the example of clinical trials and implementation science centers, there are several plausible cost-driven failure modes to consider:

Pricing new model developers out of market entry: The cost of running a phase 2 or 3 clinical trial for a therapeutic can range from 7 to >50 million U.S. dollars and is driven primarily by procedure costs, administrative staff costs, and site monitoring costs.²⁰³ For therapeutics, these large upfront costs are balanced by the many millions to billions of dollars of revenue that a pharmaceutical company can derive from a new drug. While AI trials are unlikely to require significant procedure costs, they could face similar monitoring and staff costs. Given that the reimbursement climate for medical AI remains uncertain, how will this impact organizations’ willingness to validate their solutions via trials? This could lead to smaller trials that may generate lower quality evidence or focus on more limited patient populations, or could lead to more trials being sponsored by large corporations.

Validation costs prevent validated models from migrating into community hospitals: One of the ambitions of implementation science centers²⁰² is to protect against the deployment of models that are biased by being trained on data collected at a nonrepresentative hospital. While this is an important goal, assessing whether implementation science centers will serve this goal requires further details. For instance, what happens if an AI model is shown to be effective in hospital A but not hospital B? There are many conceivable outcomes: we could halt deployment of the model, we could deploy the model with caveats, we could mandate that the developer fixes the model, and so on. What if hospital B is less resourced than hospital A, and the disparity in the model performance is attributed to hospital-level resource differences (e.g., investment in clinical informatic capabilities); would hospital B need to face the financial burden of investing in clinical informatics before they could benefit from the model? What if the converse is true, and hospital A is lower resourced and sees disproportionate benefits because the model provides outsized benefits focused on their patient population? These are nontrivial ethical questions that should be addressed ahead of the establishment of novel regulatory approaches such as implementation science centers.

The political and regulatory environment will focus on the needs of patients and caregivers if institutional health care puts in the will to work closely with patient-centric groups. For example, the National Cancer Act in 1971, the Americans with Disabilities Act of 1990, the National Cancer Survivorship Act of 1996, ACT UP (AIDS Coalition to Unleash Power), and the Affordable Care Act have all driven profound changes in health care after people pushed for them and lawmakers responded.^204–207 As we move forward with approaches that aim to validate medical AI, we must get granular about both the benefits and possible harms of AI in health care, and proceed with pragmatic validation schemes that enable us collectively to achieve highly useful and validated medical AI.

Conclusion

In this review article, we have identified key use cases where AI has the potential to dramatically improve how patients and caregivers experience cancer care. AI-enabled services have the potential to address challenges patients and caregivers face as they go through receiving and understanding a cancer diagnosis,⁵ selecting a treatment plan,^4,73 and then living with cancer, enabling them to become copilots with their medical team. As a contrast to “sustaining” innovations that make “back office” services of health care operationally more efficient, using AI-enabled software agents to support personalized care and the patient and caregiver experience has the potential to disruptively impact oncology care by allowing all patients and caregivers to access personalized care and navigation services, regardless of health literacy or resource barriers.

Footnotes

Authors’ Contributions

F.N.: Conceptualization, Writing—original draft, Writing—review & editing. B.P.: Conceptualization, Writing—original draft, Writing—review & editing.

Author Disclosure Statement

Brad is a founding member of ennov1; an advisor to 4DPath, Alva10, Cancer Commons, Consuli, Rabble Health, and Travera; and is an active contributor to the Personalized Medicine Coalition. Frank: Employment and stock ownership: Translating Science PBC, (former) Tempus AI, (former) Databricks.

Funding Information

No funding was received for this article.

References

Haslam

, Kim

, Prasad

. Overall survival for oncology drugs approved for genomic indications. Eur J Cancer, 2022; 160:175–179.

Sadik

, Pritchard

, Keeling

, et al. Impact of clinical practice gaps on the implementation of personalized medicine in advanced non–small-cell lung cancer. JCO Precis Oncol, 2022; 6:e2200246.

Scott

, Lennerz

, Johnson

, et al. Compromised outcomes in stage IV non–small-cell lung cancer with actionable mutations initially treated without tyrosine kinase inhibitors: A retrospective analysis of real-world data. JCO Oncol Pract, 2024; 20(1):145–153.

Hamilton

, Banerjee

, Carlsson

, et al. Clinician perspectives on communication and implementation challenges in precision oncology. Per Med, 2021; 18(6):559–572.

Lambourne

, Minard

, Deal

, et al. Optimizing patient education of oncology medications: A patient perspective. J Cancer Educ, 2019; 34(5):1024–1030.

Chino

, Johnson

, Moss

. Compliance with price transparency rules at US National Cancer Institute–designated cancer centers. JAMA Oncol, 2021; 7(12):1903–1904.

Prasad

, Royce

, Palmer

. Do federal price transparency regulations neglect oncology patients? JCO Oncol Pract, 2022; 18(6):413–416.

Kann

, Johnson

, Aerts

, et al. Changes in length and complexity of clinical practice guidelines in oncology, 1996-2019. JAMA Netw Open, 2020; 3(3):e200841.

Yezefski

, Steelquist

, Watabayashi

, et al. Impact of trained oncology financial navigators on patient out-of-pocket spending. Am J Manag Care, 2018; 24(5 Suppl):S74–S79.

10.

Haverfield

, Singer

, Gray

, et al. Implementing routine communication about costs of cancer treatment: Perspectives of providers, patients, and caregivers. Support Care Cancer, 2020; 28(9):4255–4262.

11.

Nussinov

, Jang

, Nir

, et al. A new precision medicine initiative at the dawn of exascale computing. Signal Transduct Target Ther, 2021; 6(1):3.

12.

Kowalski

, Xu

, Ylagan

, et al. PACMIDS: Patient-centered, ai-enabled, molecular-informed data fabric for clinical treatment decision support. AI Precision Oncol, 2024; 1(2):127–144.

13.

Finer

, Pannone

, Bard

, et al. 75. Clinical implementation of a precision medicine consultation service. Cancer Genetics, 2022; 268-269:24–25.

14.

Sae-Hau

, Disare

, Michaels

, et al. Overcoming barriers to clinical trial participation: Outcomes of a national clinical trial matching and navigation service for patients with a blood cancer. JCO Oncol Pract, 2021; 17(12):e1866-78–e1878.

15.

Mukherjee

, Gamble

, Ausin

, et al. Polaris: A Safety-focused LLM Constellation Architecture for Healthcare. arXiv, 2024 preprint arXiv:2403.13313.

16.

Christine

, Daniel

, Florian

, et al. “Uninformed consent” in clinical trials with cancer patients: A qualitative analysis of patients’ and support persons’ communication experiences and needs. Patient Educ Couns, 2024; 122:108144.

17.

Longwell

, Hirsch

, Binder

, et al. Performance of large language models on medical oncology examination questions. JAMA Netw Open, 2024; 7(6):e2417641.

18.

Hegselmann

, Shen

, Gierse

, et al. Medical expert annotations of unsupported facts in doctor-written and LLM-generated patient summaries. Physionet, 2024.

19.

Yeoh

, Gray

. Health economics and cancer care. Clin Oncol (R Coll Radiol), 2022; 34(9):e377-82–e382.

20.

Rajguru

, Carmichael

, Cetnar

. Determining value of high cost cancer therapies and discussing cost of cancer care: The patient perspective.

21.

Wahlster

, Scahill

, Lu

, et al. Barriers to access and use of high cost medicines: A review. Health Policy Technol, 2015; 4(3):191–214.

22.

Trapani

, Kraemer

, Rugo

, et al. Impact of prior authorization on patient access to cancer care. Am Soc Clin Oncol Educ Book, 2023; 43:e100036.

23.

Kyle

, Keating

. Prior authorization and association with delayed or discontinued prescription fills. J Clin Oncol, 2024; 42(8):951–960.

24.

Shea

, Roberts

, Walrath

, et al. Use of multiple endpoints and approval paths depicts a decade of FDA oncology drug approvals. Clin Cancer Res, 2013; 19(14):3722–3731.

25.

Olivier

, Haslam

, Prasad

. Is financial toxicity captured in quality of life assessments in oncology randomized clinical trials?. J Cancer Policy, 2023; 36:100423.

26.

Gharzai

, Jagsi

. Incorporating financial toxicity considerations into clinical trial design to facilitate patient‐centered decision‐making in oncology. Cancer, 2023; 129(8):1143–1148.

27.

Christensen

. The innovator’s dilemma: When new technologies cause great firms to fail. Harvard Business Review Press; 2013.

28.

Kalbach

. The jobs to be done playbook: Align your markets, organization, and strategy around customer needs. Rosenfeld Media; 2020.

29.

Chen

, Argentinis

, Weber

. IBM Watson: How cognitive computing can be applied to big data challenges in life sciences research. Clin Ther, 2016; 38(4):688–701.

30.

Strickland

. IBM Watson, heal thyself: How IBM overpromised and underdelivered on AI health care. IEEE Spectr, 2019; 56(4):24–31.

31.

Abràmoff

, Roehrenbeck

, Trujillo

, et al. A reimbursement framework for artificial intelligence in healthcare. NPJ Digit Med, 2022; 5(1):72.

32.

Abramoff

, Dai

, Zou

. Scaling adoption of medical AI—Reimbursement from value-based care and fee-for-service perspectives. NEJM AI, 2024; 1(5):AIpc2400083.

33.

Poon

, Jha

, Christino

, et al. Assessing the level of healthcare information technology adoption in the United States: A snapshot. BMC Med Inform Decis Mak, 2006; 6(1):1–9.

34.

Chen

, Golding

, Nicola

. Who will pay for AI? Radiol Artif Intell, 2021; 3(3):e210030.

35.

Patel

, Jain

, Parikh

. The enhancing oncology model: Leveraging improvement science to increase health equity in value-based care. J Natl Cancer Inst, 2023; 115(2):125–130.

36.

Basch

, Wilfong

, Schrag

. Adding patient-reported outcomes to Medicare’s oncology value-based payment model. JAMA, 2020; 323(3):213–214.

37.

Anderson

, Klemm

. The Internet: Friend or foe when providing patient education? Clin J Oncol Nurs, 2008; 12(1):55–63.

38.

Duimel

, Linn

, Smets

, et al. Profiling cancer patients based on their motives for seeking informational and emotional support online. Health Commun, 2023; 38(14):3223–3237.

39.

Hall

, Kunz

, Davis

, et al. The cancer experience map: An approach to including the patient voice in supportive care solutions. J Med Internet Res, 2015; 17(5):e132.

40.

Adler Jaffe

, Kano

, Rieder

, et al. “Care needs to be integrated” Patient and provider perspectives on a cancer shared-care model. J Cancer Surviv, 2023:1–8.

41.

American Health Information Management Association Foundation. Understanding, Access and Use of Health Information in America. 2021. Available from: https://ahimafoundation.org/media/ngfbggsk/oct2021_understanding_access_use_health_information_america_ahima_foundation.pdf

42.

Cline

, Liao

, Parsons

, et al. BRCA Challenge: BRCA exchange as a global resource for variants in BRCA1 and BRCA2. PLoS Genet, 2018; 14(12):e1007752.

43.

Swire-Thompson

, Johnson

. Cancer: A model topic for misinformation researchers. Curr Opin Psychol, 2023; 56:101775.

44.

Roy

, Purington

, Liu

, et al. Limited English proficiency and disparities in health care engagement among patients with breast cancer. JCO Oncol Pract, 2021; 17(12):e1837-45–e1845.

45.

Krupa

, Roark

, Barrett

. The Critical Role of Web Accessibility in Health Information Access, Understanding, and Use. Mathematica Policy Research; 2022.

46.

Freedman-Cass

, Fischer

, Alpert

, et al. The value and process of inclusion: Using sensitive, respectful, and inclusive language and images in NCCN content. J Natl Compr Canc Netw, 2023; 21(5):434–441.

47.

Garg

. Nocebo side-effects in cancer treatment. Lancet Oncol, 2011; 12(13):1181–1182.

48.

Wojtukiewicz

, Politynska

, Skalij

, et al. It is not just the drugs that matter: The nocebo effect. Cancer Metastasis Rev, 2019; 38(1–2):315–326.

49.

Sorin

, Brin

, Barash

, et al. Large language models (LLMs) and empathy-a systematic review. medRxiv, 2023:2023–2008.

50.

Moor

, Banerjee

, Abad

, et al. Foundation models for generalist medical artificial intelligence. Nature, 2023; 616(7956):259–265.

51.

Moor

, Huang

, Wu

, et al. Med-flamingo: a multimodal medical few-shot learner. In: Machine learning for health (ML4H). PMLR; 2023, pp. 353–367.

52.

Nievas

, Basu

, Wang

, et al. Distilling large language models for matching patients to clinical trials. JAMIA. 2024:ocae073.

53.

Shapiro

, Stuhlmiller

, Wasserman

, et al. AI-augmented clinical decision support in a patient-centric precision oncology registry. AI in Precision Oncol, 2023; 1(1):27–37.

54.

DeMartino

, Miljković

, Prasad

. Potential cost implications for all US Food and Drug Administration oncology drug approvals in 2018. JAMA Intern Med, 2021; 181(2):162–167.

55.

FDA. Table of Pharmacogenomic Biomarkers in Drug Labeling. 2024. Available from: https://www.fda.gov/drugs/science-and-research-drugs/table-pharmacogenomic-biomarkers-drug-labeling, 2024

56.

Qumseya

, Goddard

, Qumseya

, et al. Barriers to clinical practice guideline implementation among physicians: A physician survey. Int J Gen Med, 2021; 14:7591–7598.

57.

Kaupp

, Scott

, Minard

, et al. Optimizing patient education of oncology medications: A quantitative analysis of the patient perspective. J Oncol Pharm Pract, 2019; 25(6):1445–1455.

58.

Chow

, Habermann

, Abraham

, et al. Does enrollment in cancer trials improve survival? J Am Coll Surg, 2013; 216(4):774–780.

59.

Unger

, Shulman

, Facktor

, et al. National estimates of the participation of patients with cancer in clinical research studies based on Commission on cancer accreditation data. JCO, 2024; 42(18):2139–2148.

60.

Peterson

, Plana

, Bitterman

, et al. Growth in eligibility criteria content and failure to accrue among National Cancer Institute (NCI)‐affiliated clinical trials. Cancer Med, 2023; 12(4):4715–4724.

61.

Iskander

, Moyer

, Vigneault

, et al. Survival benefit associated with participation in clinical trials of anticancer drugs: A systematic review and meta-analysis. JAMA, 2024; 331(24):2105–2113.

62.

Marks

, Sridhar

, Ai

, et al. Precision immuno-oncology in NSCLC through gender equity lenses. Cancers (Basel), 2024; 16(7):1413.

63.

De La Vega

, Pouliot

, Rhead

. Incorporating Real-World Clinico-Genomic Insights to Inform Diversity Enrollment Targets in Oncology Trials. medRxiv, 2024:2024–2004.

64.

Liu

, Rizzo

, Whipple

, et al. Evaluating eligibility criteria of oncology trials using real-world data and AI. Nature, 2021; 592(7855):629–633.

65.

Fukuda

, Huang

, Finnigan

, et al. Risks and benefits of phase 1 oncology trials, 2001 through 2012.

66.

Shah

, Vaishampayan

. Therapy of advanced prostate cancer: Targeting the androgen receptor axis in earlier lines of treatment. Target Oncol, 2018; 13(6):679–689.

67.

Haslam

, Prasad

. The response rate by tumour type for tissue-agnostic approved drugs. Eur J Cancer, 2023; 184:117–119.

68.

Unger

, Hershman

, Albain

, et al. Patient income level and cancer clinical trial participation. J Clin Oncol, 2013; 31(5):536–542.

69.

Atkinson

, Massett

, Mylks

, et al. User-centered research on breast cancer patient needs and preferences of an Internet-based clinical trial matching system. J Med Internet Res, 2007; 9(2):e621.

70.

Atkinson

, Saperstein

, Massett

, et al. Using the Internet to search for cancer clinical trials: A comparative audit of clinical trial search tools. Contemp Clin Trials, 2008; 29(4):555–564.

71.

Wornow

, Lozano

, Dash

, et al. Zero-shot clinical trial patient matching with LLMs. arXiv, 2024 preprint arXiv:2402.05125.

72.

Zarinshenas

, Amini

, Mambetsariev

, et al. Assessment of barriers and challenges to screening, diagnosis, and biomarker testing in early-stage lung cancer. Cancers (Basel), 2023; 15(5):1595.

73.

Waterhouse

, Ward

, Arnal

, et al. Closing the testing gap: Standardization of comprehensive biomarker testing for metastatic non–small-cell lung cancer in a large community oncology practice. JCO Oncol Pract, 2023; 19(6):e951-6–e956.

74.

Brown

, Aisner

, Oxnard

. Precision medicine in non–small cell lung cancer: Current standards in pathology and biomarker interpretation. Am Soc Clin Oncol Educ Book, 2018; 38:708–715.

75.

Hall

, D’Avanzo

, Chertock

, et al. Oncologists’ perceptions of tumor genomic profiling and the communication of test results and risks. Public Health Genomics, 2021; 24(5–6):304–309.

76.

Tutika

, Bennett

, Abraham

, et al. Mainstreaming of genomics in oncology: A nationwide survey of the genomics training needs of UK oncologists. Clin Med (Lond), 2023; 23(1):9–15.

77.

Haga

, Mills

, Pollak

, et al. Developing patient-friendly genetic and genomic test reports: Formats to promote patient engagement and understanding. Genome Med, 2014; 6(7):58.

78.

Schilsky

, Longo

. Closing the gap in cancer genomic testing. N Engl J Med, 2022; 387(23):2107–2110.

79.

Zerdes

, Karafousia

, Mezheyeuski

, et al. Discordance of PD-L1 expression at the protein and RNA levels in early breast cancer. Cancers (Basel), 2021; 13(18):4655.

80.

Iams

, Mackay

, Ben-Shachar

, et al. Concurrent tissue and circulating tumor DNA Molecular profiling to detect guideline-based targeted mutations in a multicancer cohort. JAMA Netw Open, 2024; 7(1):e2351700.

81.

Mansfield

. FDA perspective on companion diagnostics: An evolving paradigm. Clin Cancer Res, 2014; 20(6):1453–1457.

82.

Marcus

, Fashoyin-Aje

, Donoghue

, et al. FDA approval summary: Pembrolizumab for the treatment of tumor mutational burden–high solid tumors. Clin Cancer Res, 2021; 27(17):4685–4689.

83.

Casak

, Marcus

, Fashoyin-Aje

, et al. FDA approval summary: Pembrolizumab for the first-line treatment of patients with MSI-H/dMMR advanced unresectable or metastatic colorectal carcinoma. Clin Cancer Res, 2021; 27(17):4680–4684.

84.

Vega

, Yee

, McShane

, et al. Aligning Tumor Mutational Burden (TMB) quantification across diagnostic platforms: Phase II of the friends of cancer research TMB harmonization project. Ann Oncol, 2021; 32(12):1626–1636.

85.

Umar

, Boland

, Terdiman

, et al. Revised bethesda guidelines for hereditary nonpolyposis colorectal cancer (Lynch syndrome) and microsatellite instability. J Natl Cancer Inst, 2004; 96(4):261–268.

86.

Shirts

, Konnick

, Upham

, et al. Using somatic mutations from tumors to classify variants in mismatch repair genes. Am J Hum Genet, 2018; 103(1):19–29.

87.

Hempelmann

, Lockwood

, Konnick

, et al. Microsatellite instability in prostate cancer by PCR or next-generation sequencing. J Immunother Cancer, 2018; 6(1):29–27.

88.

Bartley

, Mills

, Konnick

, et al. Mismatch repair and microsatellite instability testing for immune checkpoint inhibitor therapy: Guideline from the College of American Pathologists in collaboration with the Association for Molecular Pathology and Fight Colorectal Cancer. Arch Pathol Lab Med, 2022; 146(10):1194–1210.

89.

Vikas

, Messersmith

, Compton

, et al. Mismatch repair and microsatellite instability testing for immune checkpoint inhibitor therapy: ASCO endorsement of College of American Pathologists guideline. J Clin Oncol, 2023; 41(10):1943–1948.

90.

, Rizvi

, Schau

, et al. Development and validation of a deep learning-based microsatellite instability predictor from prostate cancer whole-slide images. NPJ Precis Oncol, 2024; 8(1):88.

91.

Lincoln

, Yang

, Cline

, et al. Consistency of BRCA1 and BRCA2 variant classifications among clinical diagnostic laboratories. JCO Precis Oncol, 2017; 1:1–0.

92.

Sherwood

, Brown

, Rettino

, et al. Key differences between 13 KRAS mutation detection technologies and their relevance for clinical practice. ESMO Open, 2017; 2(4):e000235.

93.

Marshall

, Gondek

, Luo

, et al. Clonal hematopoiesis of indeterminate potential in patients with solid tumor malignancies. Cancer Res, 2022; 82(22):4107–4113.

94.

Pishvaian

, Bender

, Matrisian

, et al. A pilot study evaluating concordance between blood-based and patient-matched tumor molecular testing within pancreatic cancer patients participating in the Know Your Tumor (KYT) initiative. Oncotarget, 2017; 8(48):83446–83456.

95.

Stetson

, Ahmed

, Xu

, et al. Orthogonal comparison of four plasma NGS tests with tumor suggests technical factors are a major source of assay discordance. JCO Precis Oncol, 2019; 3:1–9.

96.

Panigrahi

, Ambs

. How comorbidities shape cancer biology and survival. Trends Cancer, 2021; 7(6):488–495.

97.

Bruinooge

, Pickard

, Vogel

, et al. Understanding the role of advanced practice providers in oncology in the United States. J Oncol Pract, 2018; 14(9):e518-32–e532.

98.

Lafferty

, Fauer

, Wright

, et al. Causes and consequences of chemotherapy delays in ambulatory oncology practices: A Multi-Site Qualitative Study. Oncol Nurs Forum, 2020; 47(4):417–427. NIH Public Access.

99.

Hohmann

, McDaniel

, Mason

, et al. Patient perspectives on primary care and oncology care coordination in the context of multiple chronic conditions: A systematic review. Res Social Adm Pharm, 2020; 16(8):1003–1016.

100.

Parker

, Kazemi

, Ahmed

, et al. Exploring the impact of primary care utilization and health information exchange upon treatment patterns and clinical outcomes of glioblastoma patients. J Neurooncol, 2024; 168(2):345–353.

101.

Osterman

, Terry

, Miller

. Improving cancer data interoperability: The promise of the Minimal Common Oncology Data Elements (mCODE) initiative. JCO Clin Cancer Inform, 2020; 4:993–1001.

102.

Guerin

, Laizet

, Le Texier

, et al. OSIRIS: A minimum data set for data sharing and interoperability in oncology. JCO Clin Cancer Inform, 2021; 5:256–265.

103.

Belenkaya

, Gurley

, Golozar

, et al. Extending the OMOP common data model and standardized vocabularies to support observational cancer research. JCO Clin Cancer Inform, 2021; 5:12–20.

104.

Mackler

, Azar

, Johengen

, et al. Feasibility of a comprehensive medication review to improve medication use for patients with cancer and comorbid conditions. Support Care Cancer, 2022; 30(12):10111–10116.

105.

Villalobos

, Korezelidou

, Unsöld

, et al. Challenges of general practitioner-oncologist interaction in end-of-life communication: A Qualitative Study. 2023.

106.

Kagalwalla

, Tsai

, George

, et al. Consuming patients’ days: Time spent on ambulatory appointments by people with cancer. Oncologist, 2024; 29(5):400–406.

107.

Guy

Jr , Richardson

. Visit duration for outpatient physician office visits among patients with cancer. J Oncol Pract, 2012; 8(3 Suppl):2s–8s.

108.

Gorin

, Haggstrom

, Han

, et al. Cancer care coordination: A systematic review and meta-analysis of over 30 years of empirical studies. Ann Behav Med, 2017; 51(4):532–546.

109.

Kurian

, Abrahamse

, Furgal

, et al. Germline genetic testing after cancer diagnosis. JAMA, 2023; 330(1):43–51.

110.

Ortendahl

, Cuyun Carter

, Thakkar

, et al. Value of next generation sequencing (NGS) testing in advanced cancer patients. J Med Econ, 2024; 27(1):519–530.

111.

Stenzinger

, Cuffel

, Paracha

, et al. Supporting biomarker-driven therapies in oncology: A genomic testing cost calculator. Oncologist, 2023; 28(5):e242-53–e253.

112.

Iragorri

, de Oliveira

, Fitzgerald

, et al. The out-of-pocket cost burden of cancer care—a systematic literature review. Curr Oncol, 2021; 28(2):1216–1248.

113.

Strohbehn

, Sankar

, Qin

, et al. An evaluation of sotorasib for the treatment of patients with non–small cell lung cancer with KRASG12C mutations. Expert Opin Pharmacother, 2022; 23(14):1569–1575.

114.

Strohbehn

, Ratain

. Sotorasib dosing and incremental cost ineffectiveness—implications and lessons for stakeholders. Nat Rev Clin Oncol, 2024; 21(5):331–332.

115.

Doherty

, Thom

, Gany

. Evidence of the feasibility and preliminary efficacy of oncology financial navigation: A scoping review. Cancer Epidemiol Biomarkers Prev, 2021; 30(10):1778–1784.

116.

Penner

, Dovidio

, Gonzalez

, et al. The effects of oncologist implicit racial bias in racially discordant oncology interactions. J Clin Oncol, 2016; 34(24):2874–2880.

117.

Scott

, Harrington

. Are cost-of-care conversations best practice? A qualitative study of oncologists’ attitudes and practice. JCO Oncol Pract, 2021; 17(10):e1424-32–e1432.

118.

Kubi

, Istl

, Lee

, et al. Advance care planning in cancer: Patient preferences for personnel and timing. JCO Oncol Pract, 2020; 16(9):e875-83–e883.

119.

Brands

, Gouw

, Beestrum

, et al. Patient-centered digital health records and their effects on health outcomes: Systematic review. J Med Internet Res, 2022; 24(12):e43086.

120.

Hendrycks

, Burns

, Basart

, et al. Measuring massive multitask language understanding. arXiv, 2020 preprint arXiv:2009.03300.

121.

Jin

, Pan

, Oufattole

, et al. What disease does this patient have? a large-scale open domain question answering dataset from medical exams. Applied Sciences, 2021; 11(14):6421.

122.

Kung

, Cheatham

, Medenilla

, et al. Performance of ChatGPT on USMLE: Potential for AI-assisted medical education using large language models. PLOS Digit Health, 2023; 2(2):e0000198.

123.

Fleming

, Lozano

, Haberkorn

, et al. Medalign: A clinician-generated dataset for instruction following with electronic medical records. AAAI, 2024; 38(20):22021–22030.

124.

Christensen

, Hall

, Dillon

, et al. Competing against luck. The story of innovation and customer choice. 1st edition. HarperBusiness an imprint of HarperCollins Publishers: New York, NY; 2016.

125.

Derry-Vick

, Heathcote

, Glesby

, et al. Scanxiety among adults with cancer: A scoping review to guide research and interventions. Cancers (Basel), 2023; 15(5):1381.

126.

Ellenbogen

, Patrie

, Gaskin

. Improving patient access to medical images by integrating an imaging portal with the electronic health record patient portal. J Am Coll Radiol, 2021; 18(6):864–867.

127.

Farmer

, Bourne

, O’Connor

, et al. Enhancing clinician and patient understanding of radiology reports: A scoping review of international guidelines. Insights Imaging, 2020; 11(1):62–60.

128.

Lyu

, Tan

, Zapadka

, et al. Translating radiology reports into plain language using ChatGPT and GPT-4 with prompt learning: Results, limitations, and potential. Vis Comput Ind Biomed Art, 2023; 6(1):9.

129.

Chen

, Kann

, Foote

, et al. Use of artificial intelligence chatbots for cancer treatment information. JAMA Oncol, 2023; 9(10):1459–1462.

130.

Huang

, Altosaar

, Ranganath

. ClinicalBERT: Modeling clinical notes and predicting hospital readmission. arXiv, 2019 preprint arXiv:1904.05342.

131.

Luo

, Sun

, Xia

, et al. BioGPT: Generative pre-trained transformer for biomedical text generation and mining. Brief Bioinform, 2022; 23(6):bbac409.

132.

Yang

, Chen

, PourNejatian

, et al. GatorTron: A large clinical language model to unlock patient information from unstructured electronic health records. arXiv, 2022 preprint arXiv:2203.03540.

133.

Saab

, Tu

, Weng

, et al. Capabilities of gemini models in medicine. arXiv, 2024 preprint arXiv:2404.18416.

134.

Bitterman

, Downing

, Maués

, et al. Promise and perils of large language models for cancer survivorship and supportive care. J Clin Oncol, 2024; 42(14):1607–1611.

135.

Christ

, Messner

, Behar

., editors. Handbook of oncology social work: Psychosocial care for people with cancer. Oxford University Press; 2015.

136.

Garchinski

, DiBiase

, Wong

, et al. Patient-centered care in cancer treatment programs: The future of integrative oncology through psychoeducation. Future Oncol, 2014; 10(16):2603–2614.

137.

Mackler

, Wietzke

, Schwartz

, et al. Characterization of an embedded clinical oncology pharmacy model across the State of Michigan. 2022.

138.

Shaverdian

, Gillespie

, Cha

, et al. Impact of telemedicine on patient satisfaction and perceptions of care quality in radiation oncology. J Natl Compr Canc Netw, 2021; 19(10):1174–1180.

139.

Yang

, Xu

, Yao

, et al. Talk2Care: An LLM-based voice assistant for communication between healthcare providers and older adults. Proc ACM Interact Mob Wearable Ubiquitous Technol, 2024; 8(2):1–35.

140.

Garcia

, Ma

, Shah

, et al. Artificial intelligence–generated draft replies to patient inbox messages. JAMA Netw Open, 2024; 7(3):e243201.

141.

, Palepu

, Schaekermann

, et al. Towards conversational diagnostic ai. arXiv, 2024 preprint arXiv:2401.05654.

142.

Tierney

, Gayre

, Hoberman

, et al. Ambient artificial intelligence scribes to alleviate the burden of clinical documentation. NEJM Catal Innov Care Deliv, 2024; 5(3):CAT-23.

143.

Krishna

, Khosla

, Bigham

, et al. Generating SOAP notes from doctor-patient conversations using modular summarization techniques. arXiv, 2020 preprint arXiv:2005.01795.

144.

Hong

, Pena-Purcell

, Ory

. Outcomes of online support and resources for cancer survivors: A systematic literature review. Patient Educ Couns, 2012; 86(3):288–296.

145.

Gustafson

, Hawkins

, McTavish

, et al. Internet-based interactive support for cancer patients: Are integrated systems better? J Commun, 2008; 58(2):238–257.

146.

Glajchen

. The emerging role and needs of family caregivers in cancer care. J Support Oncol, 2004; 2(2):145–155.

147.

Lucas

, Yang

, Pomeroy

, et al. Reasoning with large language models for medical question answering. JAMIA, 2024:ocae131.

148.

Lee

, Pham

, Uzuner

. Enhancing consumer health question reformulation: Chain-of-thought prompting integrating focus, type, and user knowledge level. In: Proceedings of the First Workshop on Patient-Oriented Language Processing (CL4Health)@ LREC-COLING 2024. 2024; pp. 220–228.

149.

Prayaga

, Agrawal

, Nguyen

, et al. Impact of social determinants of health and demographics on refill requests by Medicare patients using a conversational artificial intelligence text messaging solution: Cross-Sectional Study. JMIR Mhealth Uhealth, 2019; 7(11):e15771.

150.

Holch

, Warrington

, Bamforth

, et al. Development of an integrated electronic platform for patient self-report and management of adverse events during cancer treatment. Ann Oncol, 2017; 28(9):2305–2311.

151.

Topol

. Toward the eradication of medical diagnostic errors. Science, 2024; 383(6681):eadn9602.

152.

Chen

, Stumpe

, Cohen

. Evolving from discrete molecular data integrations to actionable molecular insights within the electronic health record. JCO Clin Cancer Inform, 2024; 8:e2400011.

153.

Nilforoshan

, Moor

, Roohani

, et al. Zero-shot causal learning. Adv Neural Inf Process Syst, 2023; 36:6862–6901.

154.

Cahill

, Gilbert

, Armstrong

. Personal health records as portal to the electronic medical record. J Neurooncol, 2014; 117(1):1–6.

155.

Mandel

, Kreda

, Mandl

, et al. SMART on FHIR: A standards-based, interoperable apps platform for electronic health records. J Am Med Inform Assoc, 2016; 23(5):899–908.

156.

Alam

, Zou

, Vir

, et al. SAGE: System for accessible guided exploration of health information. In: AAAI-2024 Workshop on Public Sector LLMs: Algorithmic and Sociotechnical Design. 2024.

157.

Ben-Shabat

, Sharvit

, Meimis

, et al. Assessing data gathering of chatbot based symptom checkers-a clinical vignettes study. Int J Med Inform, 2022; 168:104897.

158.

Ćirković

. Evaluation of four artificial intelligence–assisted self-diagnosis apps on three diagnoses: Two-Year Follow-Up Study. J Med Internet Res, 2020; 22(12):e18097.

159.

Hamilton

, Naffakh

, Reizine

, et al. Relevance and accuracy of ChatGPT-generated NGS reports with treatment recommendations for oncogene-driven NSCLC.

160.

Low

, Jackson

, Hyde

, et al. Answering real-world clinical questions using large language model based systems. arXiv, 2024 preprint arXiv:2407.00541.

161.

Rosen

, Saban

. Evaluating the reliability of ChatGPT as a tool for imaging test referral: A comparative study with a clinical decision support system. Eur Radiol, 2024 May;34(5):2826–2837.

162.

Bisercic

, Nikolic

, van der Schaar

, et al. Interpretable medical diagnostics with structured data extraction by large language models. arXiv, 2023 preprint arXiv:2306.05052.

163.

De Paoli

, Berardelli

, Limongelli

, et al. VarChat: The generative AI assistant for the interpretation of human genomic variations. Bioinformatics, 2024; 40(4):btae183.

164.

Chunn

, Nefcy

, Scouten

, et al. Mastermind: A comprehensive genomic association search engine for empirical evidence curation and genetic variant interpretation. Front Genet, 2020; 11:577152.

165.

, Moon

, Iyer

, et al. Decoding radiology reports: Potential application of OpenAI ChatGPT to enhance patient understanding of diagnostic reports. Clin Imaging, 2023; 101:137–141.

166.

Kuckelman

, Paul

, Bui

, et al. Assessing AI-powered patient education: A Case Study In Radiology. Acad Radiol, 2024; 31(1):338–342.

167.

Choi

, Song

, Shin

, et al. Developing prompts from large language model for extracting clinical information from pathology and ultrasound reports in breast cancer. Radiat Oncol J, 2023; 41(3):209–216.

168.

Chang

, Cao

, Sfreddo

, et al. LORIS robustly predicts patient outcomes with immune checkpoint blockade therapy using common clinical, pathologic and genomic features. Nat Cancer, 2024:1–8.

169.

Rashid

, Peng

, Jin

, et al. Purity Independent Subtyping Of Tumors (PurIST), a clinically robust, single-sample classifier for tumor subtyping in pancreatic cancer. Clin Cancer Res, 2020; 26(1):82–92.

170.

Benary

, Wang

, Schmidt

, et al. Leveraging large language models for decision support in personalized oncology. JAMA Netw Open, 2023; 6(11):e2343689.

171.

Iivanainen

, Lagus

, Viertolahti

, et al. Investigating large language model (LLM) performance using In-Context Learning (ICL) for interpretation of ESMO and NCCN guidelines for lung cancer.

172.

Sushil

, Kennedy

, Miao

, et al. Extracting detailed oncologic history and treatment plan from medical oncology notes with large language models. arXiv, 2023 preprint arXiv:2308.03853.

173.

Vasey

, Nagendran

, Campbell

, et al. Reporting guideline for the early stage clinical evaluation of decision support systems driven by artificial intelligence: DECIDE-AI. BMJ, 2022; 377:e070904.

174.

Gerke

, Babic

, Evgeniou

, et al. The need for a system view to regulate artificial intelligence/machine learning-based software as medical device. NPJ Digit Med, 2020; 3(1):53.

175.

Lenert

, Lane

, Wehbe

. Could an artificial intelligence approach to prior authorization be more human? J Am Med Inform Assoc, 2023; 30(5):989–994.

176.

Allen

, Earp

, Koplin

, et al. Consent-GPT: Is it ethical to delegate procedural consent to conversational AI? J Med Ethics, 2024; 50(2):77–83.

177.

Litt

, Greenstreet Akman

, Idossa

, et al. Improving clinical trial consent form readability through artificial intelligence.

178.

Gupta

, Basu

, Nievas

, et al. PRISM: Patient Records Interpretation for Semantic Clinical Trial Matching using large language models. arXiv, 2024 preprint arXiv:2404.15549.

179.

Wong

, Zhang

, Gu

, et al. Scaling clinical trial matching using large language models: A case study in oncology. In: Machine Learning for Healthcare Conference. PMLR; 2023 pp. 846–862).

180.

Jin

, Wang

, Floudas

, et al. Matching patients to clinical trials with large language models. ArXiv, 2023.

181.

Zullig

, Wolf

, Vlastelica

, et al. The role of patient financial assistance programs in reducing costs for cancer patients. J Manag Care Spec Pharm, 2017; 23(4):407–411.

182.

Wornow

, Narayan

, Opsahl-Ong

, et al. Automating the enterprise with foundation models. arXiv, 2024 preprint arXiv:2405.03710.

183.

Heo

, Rodriguez

, McNichol

, et al. Does altruism affect participation in cancer research? A systematic review.

184.

Savage

, Savage

. Doctors routinely share health data electronically under HIPAA, and sharing with patients and patients’ third-party health apps is consistent: Interoperability and privacy analysis. J Med Internet Res, 2020; 22(9):e19818.

185.

Pearman

, Young

, Cranor

. User-friendly yet rarely read: A case study on the redesign of an online HIPAA authorization. PPETs, 2022; 2022(3):558–581.

186.

Hantel

, Walsh

, Marron

, et al. Perspectives of oncologists on the ethical implications of using artificial intelligence for cancer care. JAMA Netw Open, 2024; 7(3):e244077.

187.

Wanderer

, Mishra

, Ehrenfeld

. Innovation & market consolidation among electronic health record vendors: An acute need for regulation. J Med Syst, 2014; 38(1):8–5.

188.

Gleiss

, Kohlhagen

, Pousttchi

. An apple a day–how the platform economy impacts value creation in the healthcare market. Electron Mark, 2021; 31(4):849–876.

189.

, Usuyama

, Bagga

, et al. A whole-slide foundation model for digital pathology from real-world data. Nature, 2024; 630(8015):181–188.

190.

Rudin

, Jones

, Shekelle

, et al. The value of health information technology: Filling the knowledge gap. Am J Manag Care, 2014; 20(11 Spec No. 17):eSP1–eSP8.

191.

Smith

, Carricaburu

, Genzen

. The FDA’s Proposed Rule on Laboratory-Developed Tests: Impacts on Clinical Laboratories and Patient Care.

192.

Paniz-Mondolfi

, Ramírez

. FDA’s proposed rule and its regulatory impact on emerging and reemerging neglected tropical diseases in the United States. PLoS Negl Trop Dis, 2024; 18(5):e0012116.

193.

Rehm

, Page

, Smith

, et al. GA4GH: International policies and standards for data sharing across genomic research and healthcare. Cell Genom, 2021; 1(2).

194.

Wahl

, Cossy-Gantner

, Germann

, et al. Artificial Intelligence (AI) and global health: How can AI contribute to health in resource-poor settings? BMJ Glob Health, 2018; 3(4):e000798.

195.

Meskó

, Topol

. The imperative for regulatory oversight of large language models (or generative AI) in healthcare. NPJ Digit Med, 2023; 6(1):120.

196.

Jindal

, Lungren

, Shah

. Ensuring useful adoption of generative artificial intelligence in healthcare. J Am Med Inform Assoc, 2024; 31(6):1441–1444.

197.

American Society for Clinical Oncology. Principles for the Responsible Use of Artificial Intelligence in Oncology. 2024. Available from: https://society.asco.org/sites/new-www.asco.org/files/ASCO-AI-Principles-2024.pdf

198.

Shah

, Halamka

, Saria

, et al. A nationwide network of health AI assurance laboratories. JAMA, 2024; 331(3):245–249.

199.

Vandersluis

, Savulescu

. The selective deployment of AI in healthcare: An ethical algorithm for algorithms. Bioethics, 2024; 38(5):391–400.

200.

Drazen

, Haug

. Trials of AI interventions must be preregistered. Nejm AI, 2024; 1(4):AIe2400146.

201.

Longhurst

, Singh

, Chopra

, et al. A call for artificial intelligence implementation science centers to evaluate clinical effectiveness. Nejm AI, 2024; 1(8):AIp2400223.

202.

Sertkaya

, Wong

, Jessup

, et al. Key cost drivers of pharmaceutical clinical trials in the United States. Clin Trials, 2016; 13(2):117–126.

203.

Switzer

. Disabled rights: American disability policy and the fight for equality. Georgetown University Press; 2003.

204.

Hoffman

, Stovall

. Survivorship perspectives and advocacy. J Clin Oncol, 2006; 24(32):5154–5159.

205.

Bowen

, Murshid

. Trauma-informed social policy: A conceptual framework for policy analysis and advocacy. Am J Public Health, 2016; 106(2):223–229.

206.

Gaffney

, McCormick

. The affordable care act: Implications for health-care equity. Lancet, 2017; 389(10077):1442–1452.

The Most Disruptive Near-Term Use of AI in Cancer Care: Patient Empowerment Through Software Agents

Abstract

Introduction: Engaged Patients and Caregivers Are Critical to Achieve the Transition to Personalized Medicine

A Significant Opportunity for AI in Medicine is to Address the Challenges That Patients and Caregivers Face

Treatment volatility and complexity

Testing access and test interpretation

Care fragmentation and time pressures

Reimbursement and paying for care

The Solution Concept: Applying AI Through Software Agents to Enable Patients and Caregivers to Be Copilots

Enable 24/7 support for patients through AI-supplemented care navigation services

Empower patients in evidence-based treatment decision-making through the use of AI and health care interoperability technologies

Streamline reimbursement, care coordination, and trial access by leveraging agent technologies

Call to Action: A Roadmap for Using Data and AI to Enable Democratized, Patient-Led Care

Achieving ubiquitous distribution of AI-enabled consumer health services through cooperation between health care, technology startups, and technology platforms

Ensuring robust and rapid adoption of AI-enabled software tools that touch patients directly through sensible regulation

Conclusion

Footnotes

Authors’ Contributions

Author Disclosure Statement

Funding Information

References