Sage Journals: Discover world-class research

Abstract

Comprehensive cancer rehabilitation programs that incorporate evidence-based physical activity (PA) and exercise are currently recommended as a standard component of cancer care. However, reach and access to cancer rehabilitation is fragmented due to patient-, healthcare provider-, and organizational-level barriers. Artificial intelligence (AI), including both generative AI (e.g. chatbots that use large language models) and predictive AI techniques (e.g. forecasting future outcomes), holds potential to scale cancer rehabilitation at a relatively low cost, while filling critical gaps in care. The purpose of this narrative review is to introduce the concept of AI-supported cancer rehabilitation and synthesize emerging evidence focused on PA and structured exercise interventions. We found that existing research on the role of AI to support cancer rehabilitation is in its early stages. To-date, AI has been used to support cancer rehabilitation to: 1) screen and identify patients in need of rehabilitation; 2) predict exercise training responses and outcomes; 3) enhance patient engagement and behavior change (e.g., through feedback, coaching, or conversational agents); and 4) support precision exercise prescription. Early AI-supported interventions have demonstrated modest improvements in PA levels, although evidence remains limited. We outline priority research questions and summarize key challenges relating to the ethics, equity, and implementation of AI-tools to support cancer rehabilitation. By leveraging multidisciplinary collaboration and patient-engagement, ethically and effectively designed AI-supported cancer rehabilitation tools have the potential to overcome barriers to cancer rehabilitation access and delivery, while remaining trustworthy and meaningful to end-users.

Plain Language Summary

AI and Exercise: New Ways to Support Cancer Rehabilitation: Artificial intelligence (AI) is changing many areas of cancer care, and cancer rehabilitation is no exception. People living with and beyond cancer often face challenges, from debilitating treatment side-effects to reduced physical function. Exercise and physical activity are proven ways to lessen cancer treatment side-effects and improve overall patient quality of life and health. Yet, many patients do not access evidence-based programs. This is where AI can step in. Our paper reviews the emerging evidence on how AI can support exercise and physical activity in cancer rehabilitation settings. AI tools can analyze large amounts of health data, predict future risks, and may help to tailor exercise programs to individual needs. For example, wearable devices (like a smartwatch or activity tracker) and smartphone apps can monitor physical activity levels, track symptoms, and provide real-time feedback. AI can also help clinicians design programs that adapt as patients’ needs change, potentially making rehabilitation more flexible and responsive. We found that early studies show promise, particularly with improving physical activity levels. At the same time, important challenges remain. These include ensuring patient privacy, addressing ethical concerns, and making sure AI tools are accessible to diverse populations. More research is needed to confirm the effectiveness of AI-supported cancer rehabilitation, understand risks, and guide safe implementation in clinical practice. Altogether, AI has the potential to make cancer rehabilitation more personalized, accessible, and effective. By combining technology with evidence-based exercise programs, we can better support patients in staying active during and after cancer treatment.

Keywords

oncology machine learning generative artificial intelligence physical activity

1. Introduction

Increasingly, cancer rehabilitation that prioritizes physical activity (PA) and structured exercise is recognized as an essential means to support patients and to manage the adverse effects of cancer and its treatment.^1-3 Relative to age-matched individuals without a history of cancer, those diagnosed with cancer have lower physical fitness (e.g., aerobic capacity and strength) and reduced physical function (e.g., ability to perform daily activities), contributing to significantly diminished quality of life (QoL).⁴ Mounting evidence highlights that PA and exercise-based support can help to directly address many common adverse effects of a cancer diagnosis and cancer treatment, such as fatigue, declines in physical function and reduced QoL.^5-7

Current cancer-specific exercise guidelines from leading oncology and exercise science organizations advocate for integrating exercise as a standard component of cancer care,^1-3 including the 2022 recommendations from the American Society of Clinical Oncology.² Despite published evidence-based recommendations, comprehensive cancer rehabilitation programs, with high quality PA and exercise support, are currently the exception rather than the rule, globally.^8,9 Even when such services are available, they are frequently underutilized.^10-12 Traditional cancer rehabilitation models, of which much of the current evidence is based, typically rely on “person-to-person” delivery that involves qualified exercise professionals working directly with patient participants either in-person or through live-remote formats. However, access to these services is hindered by several barriers at the patient-, provider- and organizational-level,¹³ from workforce shortages to inadequate funding and infrastructure. As a result, designing and delivering scalable, inclusive cancer rehabilitation programs continues to be a significant challenge.

Emerging digital health solutions, including artificial intelligence (AI), offer innovative strategies to expand access to cancer rehabilitation, enhance patient engagement, and personalize care.¹⁴ In this context, “patient engagement” refers to increasing active participation in PA and exercise behaviors, while “personalization” describes tailoring support based on individual needs, preferences, and predicted responses. The ground-breaking results of the CHALLENGE Trial recently showed a 28% reduced risk of cancer recurrence and 37% lower risk of death, with 8-year overall survival of 90.3% among people with colon cancer who received structured exercise support after adjuvant chemotherapy compared to 83.2% among those who received health education alone.¹⁵ Consequently, the potential use of AI to mimic a similar exercise behavior change intervention that involves extensive patient counselling for wide scale implementation is a new priority for the field. From predictive modeling to personalized exercise or PA planning and real-time monitoring, AI may be able to enhance cancer rehabilitation delivery. However, research on the utility and application of AI-supported cancer rehabilitation is urgently needed. Further, because AI-supported tools often depend on large, sensitive data sets and digital platforms, they raise critical questions about data privacy, transparency, and equitable access, particularly given that existing AI studies in oncology frequently draw on non-representative, homogeneous populations.¹⁴

The purpose of this review is to introduce the concept of AI-supported cancer rehabilitation and present emerging research, focusing specifically on PA and structured exercise applications. In this narrative review, we begin by briefly introducing what AI is, providing examples of how it is used in oncology, then summarize existing AI applications relevant to PA and exercise in oncology, describing how these tools have been used and tested to date. Lastly, we discuss future research directions and key considerations for AI research in cancer rehabilitation, including ethical, privacy, equity, and implementation issues.

1.1. What Is Artificial Intelligence and How Is It Used in Cancer Care?

AI refers to a system or machine simulating human intelligence and is designed to mimic human thinking and behaviour, including predicting, learning, and problem solving¹⁶ (Table 1). Under the umbrella of AI is machine learning (ML), a term used for methods to train AI. ML can train applications of predictive AI, such as through supervised or unsupervised learning, where algorithms classify outputs to predict future outcomes. Another rapidly developing subset of AI is generative AI (GenAI). GenAI refers to AI tools that can generate new types of data, including text, images, music, programming code, among other outputs.¹⁷ Since the release of ChatGPT to the public in 2022, GenAI has become increasingly mainstay, with multiple applications in daily life. Many GenAI tools, such as conversational agents (often known as chatbots), depend on large language models (LLM), which allow pre-trained knowledge to be used to accomplish tasks with limited input data.¹⁸

Table 1.

Common AI and ML Approaches and Terms With Healthcare Applications

Term	Main features	Example applications in healthcare
Predictive AI	Uses data to forecast outcomes or risks; often implemented via supervised or unsupervised learning	Predicting disease progression, forecasting hospital readmissions
Generative AI	Creates new data (e.g., text, images, or recommendations) that resemble training data; can include chatbots	Patient-facing chatbots, synthetic data generation, clinical documentation support
Supervised Learning	Learns from labeled input-output data; can use neural or non-neural methods	Predicting patient outcomes, supporting diagnoses
Unsupervised Learning	Identifies hidden patterns, clusters, or associations in unlabeled data; can use neural or non-neural methods	Identifying symptom–activity associations, detecting anomalies in records
Reinforcement Learning (RL)	Learns optimal actions via trial-and-error feedback (rewards/penalties); combines elements of supervised and unsupervised approaches	Personalizing treatment strategies, optimizing resource allocation
Neural Networks/Deep Learning	Neural network methods (CNNs, RNNs, ANNs, transformers); sometimes referred to as “deep” when many layers are used but definition varies	Medical image interpretation, signal analysis, natural language tasks
Natural Language Processing (NLP)	AI methods for understanding and generating human language; may use supervised, unsupervised, or reinforcement learning; neural or non-neural methods	Extracting information from clinical notes, summarizing medical records
Large Language Models (LLMs)	A recent advance of NLP; transformer-based deep learning models trained on massive text datasets; generate human-like language by predicting next words in sequence	Conversational agents (chatbots), summarizing medical records, real-time translation, clinical documentation support

CNN: convolutional neural networks; RNN: recurrent neural networks; ANN: artificial neural networks.

Current academic discourse on the potential for AI to “revolutionize healthcare” is growing, with promises of AI tools to increase accuracy, reduce costs, save time, all while minimizing human errors.¹⁹ In oncology, AI tools show promise on a number of fronts, including cancer screening and diagnosis,^20,21 drug discovery and design,²² predicting patient responses to treatment,^23,24 among other applications.²⁵ A recent systematic review of AI in supportive and palliative cancer care included 29 studies that employed AI tools to either forecast future clinical outcomes by analyzing patient data or to classify current states through text-based screening, such as identifying changes in symptoms or patient well-being.²⁶

AI-powered chatbots, or conversational agents, also hold potential as useful tools in oncology care. Chatbots use dialogue systems to enable natural language conversations with users by means of speech, text, or both.²⁷ The conversational capacity with chatbots can vary from constrained conversation tools that do not use AI, where users select predefined conversational lines, to unconstrained conversation tools that leverage AI and allow for free-text input.²⁸ In oncology settings, AI-powered chatbots have been tested in areas like treatment adherence and symptom management, showing high patient satisfaction and, in several cases, improved outcomes relative to standard care.²⁹ In an equivalence study, Chen et al randomly selected 200 patients and compared how three different LLM-chatbots answered patient questions against the answers provided by six verified oncologists.³⁰ The authors found that the best chatbot consistently outperformed physicians on ratings of response quality, empathy, and readability (all p<0.001).³⁰ In a collaborative pilot study, OpenAI GPT-4 combined with retrieval augmented generation was tested on 104 frequently asked questions relating to breast cancer to deliver personalized answers.³¹ Overall, 85.6% of responses were deemed comprehensible, 87.5% correct, and 89.4% free of undue harm, and 69.2% complete.³¹ The authors concluded that the chatbot can provide largely accurate information, but noted limitations regarding response completeness and potential harm underscoring the need for clinician oversight.³¹ A major advantage of chatbots is their 24/7 availability, which may help to better prepare patients for conversations with their healthcare providers or answer lingering questions after healthcare appointments. Current evidence is laying the much-needed groundwork for future research to develop chatbots that can complement or enhance cancer care, serving as another clinical tool to better support patients.

Current evidence illustrates the extent of AI’s potential in oncology, from diagnostics to supportive and palliative care, and illustrate its growing relevance in person-centered interventions. As AI tools continue to evolve, their adaptability, scalability, and capacity for personalization position them as promising tools for cancer rehabilitation. The next section introduces how AI is currently being applied to PA and exercise interventions in oncology and reviews the current state of research on AI-supported cancer rehabilitation.

2. Current Landscape: AI Applications in Cancer Rehabilitation

2.1. Search Strategy

To identify literature describing applications of AI within PA or exercise-based cancer rehabilitation, a search of MEDLINE (Ovid) was conducted from database inception to August 1, 2025, The search strategy used Boolean operators to combine free-text terms across three core concepts: 1) artificial intelligence (e.g., “artificial intelligence,” “machine learning,” “deep learning”); 2) cancer (e.g., “cancer*,” “neoplasm*,” “oncolog*”); and 3) exercise (e.g., “cancer rehabilitation,” “exercise,” “exercise therap*,” “physical activit*”). All terms were searched in the title and abstract fields only; controlled vocabulary terms were not used.

Key articles known to the research team were used to seed the search, and backward and forward citation tracking was performed to identify additional publications. Studies were eligible if they: 1) were published in English; 2) described the use of AI to support PA or exercise-based cancer rehabilitation (e.g., patient screening, PA engagement, prediction of training responses, precision exercise prescription); and 3) focused on individuals living with or beyond cancer. Study protocols describing planned AI-supported applications were also included. Exclusion criteria were: 1) conference abstracts or presentations; 2) studies addressing other supportive care domains (e.g., psychology, nutrition) without a physical rehabilitation component; and 3) digital health interventions not explicitly supported by AI (e.g., rules-based systems).

2.2. Screening and Early Identification

In oncology, Ma et al constructed a predictive AI model to predict and proactively screen for low PA levels to form the basis for formulating individualized PA rehabilitation programs for people with lung cancer.³² The authors reported that their predictive model could effectively predict PA levels, providing healthcare providers with quantitative references to support early identification of people with low PA. The model relied on 15 input features that were identified through multivariate logistic regression, such as demographic, disease-specific, physical function (e.g., six-minute walk test), and patient-reported outcomes. A practical online prediction model calculation tool was then created to enhance the clinical applicability. Clinical staff can collect relevant patient data during hospitalization, access the website, and enter the 15 variables to generate a real-time prediction of a patient’s PA level.³² However, currently the prediction tool is limited by its hardware facilities and can only be used within a local area network. Of note, any one of these measures on their own could potentially be flags to healthcare providers to refer to or prescribe PA interventions. However, in practice, referrals to cancer rehabilitation services and conversations about PA between patients and their care teams are underutilized.³³ Thus, flags generated by predictive AI may help to prompt healthcare providers to ensure that the most appropriate cancer rehabilitation service or advice about PA is offered consistently.

2.2. Predicting Exercise Training Responses and Outcomes

Predictive AI has been used to identify impaired cardiorespiratory fitness and to predict poor cardiorespiratory response to aerobic exercise training in women with metastatic breast cancer.³⁴ With multidimensional data, including echocardiography, hematologic and patient-reported outcomes, Novo et al categorized participants (n=64) with metastatic breast cancer into two phenotypes, one of which had a blunted cardiorespiratory training response to aerobic exercise. Participants with this phenotype were more like to have greater than three lines of previous anticancer therapy for metastatic disease, and reduced resting left ventricular systolic and diastolic function, cardiac output reserve, hematocrit, lymphocyte count, patient-reported outcomes, and cardiorespiratory fitness (p < .05) at baseline. Participants in this phenotype who took part in aerobic exercise training had a blunted cardiorespiratory training response compared to people who did not fall into the phenotype (-1.94 ± 3.80 mL O₂·kg^-1·min^-1 compared to 0.70 ± 2.22 mL O₂·kg^-1·min^-1).³⁴ Thus, ML-driven phenotypic clustering may be able to improve cardiovascular risk stratification and inform exercise interventions for people with cancer. These applications highlight growing potential of predictive AI to support precision exercise interventions by identifying who is most likely to benefit and how to optimize outcomes across diverse populations.³⁵

2.3. Behaviour Change and Engagement Tools

Hassoon et al tested two AI-driven interventions compared to a control condition to increase PA among 42 people with cancer who were considered overweight or obese and physically inactive in a three-arm randomized controlled pilot study.³⁶ Participants were randomized to either: 1) on-demand AI coaching delivered via an interactive digital voice assist through an Amazon Echo smart speaker, called “MyCoach” (n=14), 2) autonomous data-driven text messaging via mobile phone, called “SmartText” (n=14), or 3) educational materials on PA, which acted as the control condition (n=14). Both MyCoach and SmartText used AI-based recommendation systems to personalize coaching, but had key differences. SmartText used a supervised learning, goal-based model, whereas MyCoach used an unsupervised, reinforcement learning goal-based model with reward conditions. In the MyCoach arm, the participant initiated all interactions through the smart speaker, such that intensity and frequency were determined by user engagement rather than clinician input. MyCoach generated automated rewards based on learned participant behavior and general response patterns, and incorporated real-time data from a wearable sensor (Fitbit Charge HR2) to enable reward feedback. While not explicitly referred to as a chatbot, MyCoach uses voice-based interaction, reinforcement learning, and bidirectional communication, and thus functioned as a voice-enabled chatbot tool. SmartText, by contrast, used a supervised learning model to deliver daily messages based on participants’ preferences, wearable sensor data, and progress over time. After four weeks, participants in the MyCoach group increased their daily steps by an average of 3,618 steps/day, suggesting that automated, voice-assisted AI coaching may be a scalable and engaging approach to promote PA in this population. Participants in the SmartText group also increased the average number of steps by a mean of 1,619 steps/day.³⁶

Nurse AMIE (Addressing Malignancies in Individuals Everyday) is a technology-enabled, rule-based supportive care tablet platform that provides symptom assessment and guideline-concordant self-care for rural patients with advanced cancer.^37,38 Nurse AMIE asks about symptom level (e.g. pain, fatigue, sleep, distress) on a scale of 0 to 10. Afterwards, it asks participants to record their previous day step totals on the tablet. A decision algorithm determines which self-care intervention to offer the participant, and then Nurse AMIE suggests a resource or intervention, including relevant PA and exercise resources (including pedometer step goals and exercise videos). In a feasibility trial (n=68), participants interacted with the tablet on 48% of possible days. Walking was rated as the most helpful activity (83% of days offered), whereas psychological strategies were least helpful overall (49% helpful, 51% not helpful). No significant changes in symptom outcomes were found, but small nonsignificant improvements were found for fatigue (d=0.24). and non-significant but potentially clinically meaningful, moderate reductions were found for sleep (d=0.65) and distress (d=0.74).³⁸ The current intervention-logic for Nurse AMIE is not AI-driven, however, there are plans for a new trial to embed AI at the infrastructure level by pairing Nurse AMIE with the Amazon Echo Show (an AI-enabled voice platform).³⁹ While clinical decision-making within Nurse AMIE remains rule-based, the Amazon Echo Show will use AI for speech recognition and Natural Language Processing to facilitate conversational interaction. The Amazon Echo Show acts like an ML-powered chatbot for participant interaction and interpreting voice commands.³⁹ By using the Amazon Echo Show, participants can use their voice to communicate with Nurse AMIE, which may be more user-friendly and accessible, compared to other technologies.

2.4. Precision Exercise Prescription Systems

Moreno-Gutierrez et al developed an AI-enabled, knowledge-based digital health system called ATOPE+ that integrates rule-based expert systems, ML, and wearable data analytics.⁴⁰ ATOPE+ was designed to remotely assess participant’s workload-recovery ratio and prescribe individualized exercise dose for people with cancer. This is accomplished by combining expert knowledge (i.e. physical therapists with therapeutic exercise experience) with information from wearable devices and other health data sources. Patients use a smartphone app to gather their data, interact with the experts, and receive the personalized exercise prescriptions. The app also connects to external devices (e.g., Bluetooth ECG) to record heart rate variability and collects patient-reported measures, such as recovery, distress, sleep quality, and fatigue. Patients also use a wearable activity tracker to collect their overall and training-specific PA levels. In a small usability investigation with providers (n=8, physiotherapists) using ATOPE+, the app scored 91.6 ± 7.8 points (average ± standard deviation) and the web dashboard 85.6 ± 20.9 (deemed, “excellent”) on the System Usability Scale (range 0-100). In qualitative interviews, the providers noted benefits including how the app was intuitive and how it may foster patient autonomy and exercise adherence; however, they also described the need to train patients with fewer technological skills, add prompts for patients to remind them of the protocol, and offered suggestions to increase engagement (such as gamification elements).⁴⁰ The system was prototyped in 16 participants (n=11 women with breast cancer), and the app is now being tested in a clinical trial for the prevention of cardiotoxicity for women with breast cancer (ClinicalTrials.gov registration number NCT06518200).

Gao et al also published a protocol paper for a retrospective ML modeling cohort study.⁴¹ The authors plan to use predictive AI to develop exercise prescriptions that follows the 2019 ACSM guidelines, tracked using individual characteristics, such as demographics, cancer type, and lifestyle factors, along with wearable devices to monitor daily activity and weekly adherence. ML models will analyze the collected data and identify which combinations of baseline characteristics are associated with successful completion of exercise prescriptions. By grouping patients with similar characteristics and outcomes, the authors aim to build a system that makes it easier for clinicians to match individuals with the exercise plans most likely to work for them.⁴¹

3. Challenges and Future Directions

Current evidence emphasizes the various contexts in which AI can be used to enhance PA and exercise interventions for people with cancer. However, the evidence base is still emerging, and applications of AI to support cancer rehabilitation remain mostly in the planning and development phases. Our review shows that AI-supported cancer rehabilitation applications are largely using predictive AI to: 1) screen and identify patients in need of rehabilitation; 2) predict exercise training responses and outcomes; 3) enhance patient engagement and behavior change (e.g., through feedback, coaching, or conversational agents); and 4) support precision exercise prescription. Notably, several limitations to predictive AI exist within these contexts. When it comes to predicting exercise responses, for example, there is a high degree of variability, influenced by factors such as treatment type, symptom burden, and physical function, which complicates generating accurate predictions. Many predictive models only rely on surrogate markers (e.g., step count, heart rate variability) that may not fully capture clinical relevance or patient priorities. Further, in studies such as Novo et al,³⁴ that explored ML-driven phenotypic clustering, the authors reported that the high number of statistical comparisons performed may increase the risk of Type I error, leading to false associations. As predictive tools continue to evolve, ongoing validation and contextual adaptation are needed to ensure their safe and effective use in cancer rehabilitation.

While evidence specific to cancer rehabilitation is in its infancy, AI is increasingly used to support PA in other populations, helping to set the stage for future research in oncology. In a recent narrative review, Canzone et al reported on 15 studies that had incorporated AI to promote PA among children, healthy adults, older adults, and people with disabilities.⁴² Across the studies, AI was embedded into interventions to deliver messages and notifications, generate predictions about future outcomes, analyze gestures and postures, collect data, or serve as a virtual training guide. Most studies used observational or pilot intervention designs and incorporated AI primarily for automated messaging, PA monitoring via wearable or smartphone sensors, and basic behavioral prediction. However, few studies evaluated clinical populations, including people living with and beyond cancer.⁴²

Notably, while AI-powered chatbots hold promises to increase cancer rehabilitation access and flexibility, to-date, voice or text-based chatbots powered by GenAI have not been fully leveraged to offer PA or exercise-based support to people with cancer. Chatbots are increasingly recognized as a novel AI-tool to improve the accessibility and effectiveness of PA or other lifestyle behaviour change interventions.^28,43 GenAI chatbots overcome critical limitations of rules-based systems that restrict users to predefined input and often lead to repetitive program content, which may decrease user satisfaction.⁴⁴ Another advantage to GenAI chatbots is the low degree of digital literacy required, beyond typing or voice interaction. In a systematic review and meta-analysis, Singh et al reported that AI-supported and non-AI chatbot interventions significantly increased PA levels, fruit and vegetable consumption, and sleep duration and quality across diverse populations and age groups.⁴⁵ Across 19 trials, the authors reported significant effects on PA (SMD=0.28), steps (+735/day), focusing on goal setting, personalized feedback, reminders, and motivational support for standalone chatbot interventions or as part of multicomponent intervention with wearables.⁴⁵ A subgroup analysis illustrated that AI-supported chatbots were more effective than non-AI chatbots for fruit and vegetable consumption. While we noted that Hassoon et al used chatbot interventions in people with cancer, both MyCoach and SmartText were described as using traditional ML and reinforcement learning, not LLMs or generative text.³⁶ Exploring how GenAI chatbots can be designed and evaluated within cancer rehabilitation is an important direction for future research.

Significant work remains to develop GenAI chatbots that fully meet end-user needs and produce meaningful changes in relevant clinical outcomes. Because the public is familiar with GenAI chatbots, like ChatGPT, there are now expectations for this level of seamless, intelligent conversations for all conversational agents. Publicly available chatbots however, should be used cautiously, given notable risks from hallucinations, generating responses based on inaccurate information, to possible personal data violation.⁴⁶ Moreover, many existing chatbots are trained on English language data, posing a significant barrier to access for non-English-speaking patients. In a critical evaluation of OpenAI’s GPT-4 model, Dergaa et al concluded that AI could act as a “baseline” for safe exercise recommendations; but at this stage more research is needed to determine the degree of complexity or adaptability required to optimize physical fitness and health recommendations.⁴⁷ However, since that evaluation, the additional iterations of GPT have been released, highlighting the speed at which these tools advance and the need for evaluation to be on-going to decern how updates to AI tools perform. Future chatbot enhancements may also include approaches such as retrieval-augmented generation, which allows the chatbot to base its conversations on pre-specified documents. Using retrieval-augmented generation can help prevent the chatbot from generating false information and offers more up-to-date and quality-controlled information for the LLM to draw from when answering questions.

GenAI-supported relational agents may also offer unique advantages to cancer rehabilitation delivery. RAs are typically powered by a LLM, such as GPT, but are presented as animated digital characters that use verbal and non-verbal communication to build rapport and support behavior change.⁴⁸ Through strategies like empathy, storytelling, and humor, relational agents aim to mimic face-to-face human interaction.⁴⁹ Early evidence from health promotion programs for adults without health conditions suggests relational agents may enhance engagement, including changes in PA levels, but require further refinement for broader application in cancer rehabilitation.⁴⁹ Advancing GenAI-enabled cancer rehabilitation will ultimately require tools that consider individual patient preferences and priorities, while integrating diverse data sources, such as diagnostic information, lifestyle factors, and treatment plans and responses.^14,50

4. Current Research Questions

Harnessing AI for cancer rehabilitation is an innovative, and likely inevitable solution, to address well-identified barriers to widespread implementation and adoption of cancer rehabilitation within healthcare systems. Healthcare staffing, infrastructure, and funding are currently insufficient to meet the scope of the current need for rehabilitation services.⁵¹ Leveraging AI technology may allow rehabilitation professionals to more effectively use their time to address patient goals and needs, and triage for services, offering “person-to-person” options and less intensive, more “hands-off” solutions where appropriate. Consequently, AI holds potential to bridge existing healthcare gaps, where shortages are limiting access to vital care.

Well-designed, large trials with appropriate, person-centered endpoints and long-term follow-up are needed to address several emerging research questions regarding the use of AI-supported cancer rehabilitation. In developing this paper, our author team generated some key future research questions for the field (Table 2). Multidisciplinary collaboration is an essential component of cancer rehabilitation and will need to be harnessed to design AI tools to support future rehabilitation efforts. Hilbers et al reported that after completing a scoping review on patient attitudes towards AI in cancer care, a consistent theme is that patients want AI as a decision-support tool, not as an independent decision maker, and they do not want AI to de-personalize their care.⁵² In response, AI developers need to co-design interventions, including those pertaining to cancer rehabilitation, with people with lived experience to ensure their values and expectations are incorporated to build trust and acceptability.⁵² Further, while technology, including AI, may improve outcomes, reduce health disparities, and enhance healthcare access, in some instances, disparities may be exacerbated.⁵³ People do not have equal access to technology due to differences in geographic locations, health literacy, and socioeconomic status. Advancing equity in cancer rehabilitation requires designing and implementing AI-supported tools in partnership with underserved communities, ensuring their perspectives, priorities, and lived experiences meaningfully shape the research and development process.

Table 2.

Key Research Questions for Using AI in Cancer Rehabilitation

Research question	Rationale
What are patient and healthcare provider perceptions of AI-delivered or AI-supported cancer rehabilitation, and what factors influence their acceptability and trust?	User trust, perceived usefulness, and ethical concerns will strongly influence uptake and engagement with AI tools.
Can AI tools assist with early identification and referral of patients who would benefit from cancer rehabilitation across the cancer continuum?	AI could help automate identification of eligible patients and reduce missed opportunities for timely intervention.
What ethical, privacy, and data governance considerations must be addressed in developing, testing, and deploying AI tools in cancer rehabilitation?	Transparent, responsible AI development is essential, particularly when handling sensitive health data and influencing care decisions.
What design or co-design strategies are needed to ensure AI-tools for cancer rehabilitation are usable, acceptable, and scalable across diverse settings?	Involving patients, caregivers, healthcare providers, and qualified exercise professionals in tool design can support relevance, usability, and long-term adoption.
Which AI methods (e.g., models, architectures) are most effective for prescribing exercise and predicting patient responses using real-time clinical and patient-reported data?	AI may be able to enable individualized, adaptive exercise prescriptions, but the most appropriate and effective approaches remain unknown in cancer rehabilitation contexts.
How much human oversight is necessary to ensure AI-supported cancer rehabilitation is safe, feasible, effective at different stages of the cancer continuum?	Clarifying the role of healthcare providers and qualified exercise professionals is critical to balancing automation with cancer rehabilitation effectiveness and patient safety and trust.
What training and support do healthcare providers and qualified exercise professionals need to implement AI tools in clinical and community settings?	Successful implementation depends on provider readiness and confidence, yet education and workflow integration needs are not well understood.
How can AI be developed and implemented to improve access to cancer rehabilitation and exercise programs for underserved or minority populations without reinforcing existing health disparities?	Equitable AI design is essential to avoid perpetuating bias and to ensure access for underserved communities.
How can AI adapt exercise programs in real time in response to patients’ symptoms, treatment side effects, and changes in health status?	AI-enabled responsiveness could enhance exercise safety and effectiveness, especially for patients with fluctuating symptoms or rapid changes in health status.
What are the priority endpoints for trial delivering AI-supported cancer rehabilitation and exercise programs?	Robust evaluation is needed to determine if AI-supported cancer rehabilitation can improve patient-reported outcomes, physical function, and health outcomes.

5. Important Considerations for AI-Related Research

AI advancements are demonstrating success within oncology care.²⁵ However, large scale implementation into real-world clinical settings has not occurred due to several challenges. A fulsome review of existing problems and potential solutions to AI in healthcare settings is outside the scope of this review. However, comprehensive summaries of the main challenges to the adoption of AI use in healthcare have been well covered by other authors, including ethical and legal issues, security, and societal acceptance, while also summarizing emerging solutions.^54-57 We briefly outline some key concerns that impact cancer rehabilitation clinical practice and research (summarized in Table 3).

Table 3.

Challenges of Applying AI in Clinical Practice

Challenge	Key issues	Implications for cancer rehabilitation
Data Ethics & Security	Informed consent, data ownership, privacy risks, vulnerability to cyberattacks	Patients must understand how their physiological, patient-reported, and clinical data is used; secure systems are essential
Bias in AI Tools	Bias introduced from data, training, and AI applications themselves	Can perpetuate disparities in accessing and utilizing cancer rehabilitation and lead to unsafe or unfair outcomes
Explainability & Transparency	AI decisions can be opaque (“black box”) or difficult to interpret	Lack of transparency may reduce trust among patients, healthcare providers and qualified exercise professionals, limiting adoption
Need for Human Oversight	Healthcare providers must supervise AI decisions; lack of trust or training limits adoption	Oversight builds patient trust; healthcare providers and qualified exercise professionals need AI literacy
Integration into Clinical Workflows	AI tools must fit into existing clinical workflows to avoid burdening healthcare providers	Poor integration into existing cancer rehabilitation models risks low uptake or disruption in care delivery
Regulatory & Legal Uncertainty	Lack of unified global standards; unclear liability in medical errors involving AI	Legal gaps may hinder implementation or shift risk burdens to healthcare providers or qualified exercise professionals

5.1. Informed Consent, Data Ownership, and Security

Ethical AI use requires transparent data practices, including informed consent, patient data ownership, and robust privacy protections.^55,58 Patients must be clearly informed how their data may be used by AI systems, especially when clinical decisions are involved. Health data systems are also vulnerable to cyberattacks, underscoring the need for strong cybersecurity measures.⁵⁶

5.2. Bias and Fairness

Bias is a key concern with the use of AI in healthcare settings, where AI outputs may influence patient care and health outcomes. Bias can be introduced from training data, the training process, and the AI applications themselves. AI tools trained on historical or non-representative data may perpetuate health inequities.⁵⁷ Bias can be unintentionally embedded, or even deliberately introduced through “poisoning attacks”, where data is maliciously introduced into datasets to intentionally introduce bias. Mitigating bias requires representative data, input from diverse end-users including patients, and cybersecurity measures.

5.3. Human Oversight and Trust

AI should support, not replace, qualified healthcare providers, including rehabilitation providers such as physiotherapists, occupational therapists, and exercise physiologists. An AI tool can present accurate and false information with equal confidence. While provider involvement remains important to ensure safe, ethical, and patient-centered care, the degree of oversight required may depend on the specific application and safeguards in place. For example, systems that draw on curated databases or use retrieval-augmented generation with guardrails may allow patients to interact directly with AI tools, whereas applications such as AI scribes or clinical documentation typically require provider review. Such oversight helps foster patient trust and acceptability of AI-tools in healthcare settings, while ensuring nuanced and ethical clinical decision-making.⁵⁶ To serve this role effectively, providers require relevant training to equip them with AI literacy and the skills to critically assess all AI outputs.

5.4. Regulation and Legal Accountability

Clear, consistent legal frameworks are essential to guide ethical AI integration in healthcare. Current gaps include questions about liability. Who is responsible when AI errors occur: the developer, the healthcare provider, or the system itself?^54,59

6. Conclusion

The future of AI-supported cancer rehabilitation to deliver PA and exercise-based interventions is promising but research on this topic is in its early stages. Early applications of AI have included leveraging predictive AI to identify people with cancer in need of rehabilitation and personalizing PA and exercise prescriptions, along with exploring how chatbot-like AI tools promote behavior change and increase PA levels. The opportunities for AI to support cancer rehabilitation are exciting but warrant cautious optimism. Building evidence in this field requires an approach that is ethical and person-centered.

Footnotes

ORCID iDs

Kelcey A. Bland

Ignacio Catalá-Vilaplana

John-Jose Nunez

Lauren C. Capozzi

Kristin L. Campbell

Ethical Considerations

Ethics approval was not required.

Author Contributions

KAB and ICV drafted the manuscript. JJN, LCC, and KLC critically reviewed and revised the manuscript.

Funding

The authors disclosed receipt of the following financial support for the research, authorship, and/or publication of this article.

Declaration of Conflicting Interests

The authors declared the following potential conflicts of interest with respect to the research, authorship, and/or publication of this article: JJN has received in-kind support from Amazon Web Services Canada, and indirect research support from BC Cancer Foundation and Pfizer Canada.

Data Availability Statement

Data sharing is not applicable to this article as no new datasets were generated or analyzed.

References

Campbell

Winters-Stone

Wiskemann

, et al. Exercise Guidelines for Cancer Survivors: Consensus Statement from International Multidisciplinary Roundtable. Med Sci Sports Exerc. 2019;51(11):2375-2390. doi:10.1249/mss.0000000000002116.

Ligibel

Bohlke

May

, et al. Exercise, Diet, and Weight Management During Cancer Treatment: ASCO Guideline. J Clin Oncol. 2022;40(22):2491-2507. doi:10.1200/JCO.22.00687.

Rock

Thomson

Sullivan

, et al. American Cancer Society nutrition and physical activity guideline for cancer survivors. CA Cancer J Clin. 2022;72(3):230-262. doi:10.3322/caac.21719.

Ness

Wall

Oakes

Robison

Gurney

. Physical performance limitations and participation restrictions among cancer survivors: a population-based study. Ann Epidemiol. 2006;16(3):197-205.

Alfano

Cheville

Mustian

. Developing High-Quality Cancer Rehabilitation Programs: A Timely Need. Am Soc Clin Oncol Educ Book Am Soc Clin Oncol Annu Meet. 2016;35:241-249. doi:10.1200/EDBK_156164.

Cheville

Smith

Barksdale

, et al. 29 - Cancer Rehabilitation. In: Braddom’s Physical Medicine and Rehabilitation. Elsevier; 2021: Philadelphia, PA, 568-593.e7. doi:10.1016/b978-0-323-62539-5.00029-1.

Gallegos-Kearin

Knowlton

Goldstein

, et al. Outcome Trends of Adult Cancer Patients Receiving Inpatient Rehabilitation: A 13-Year Review. Am J Phys Med Rehabil. 2018;97(7):514-522. doi:10.1097/PHM.0000000000000911.

Canestraro

Nakhle

Stack

, et al. Oncology Rehabilitation Provision and Practice Patterns across Canada. Physiother Can Physiother Can. 2013;65(1):94-102. doi:10.3138/ptc.2011-53.

Silver

Stout

Pratt-Chapman

Haylock

Sharma

. The state of cancer rehabilitation in the United States. J Cancer Rehabil. 2018;1:1-8.

10.

McEwen

Rodriguez

Martino

, et al. “I didn’t actually know there was such a thing as rehab”: survivor, family, and clinician perceptions of rehabilitation following treatment for head and neck cancer. Support Care Cancer. 2016;24:1449-1453.

11.

Pergolotti

Deal

Lavery

Reeve

Muss

. The prevalence of potentially modifiable functional deficits and the subsequent use of occupational and physical therapy by older adults with cancer. J Geriatr Oncol. 2015;6(3):194-201.

12.

Stubblefield

. The underutilization of rehabilitation to treat physical impairments in breast cancer survivors. Pm&r. 2017;9(9):S317-S323.

13.

Hendren

Chin

Fisher

, et al. Patients’ barriers to receipt of cancer care, and factors associated with needing more assistance from a patient navigator. J Natl Med Assoc. 2011;103(8):701-710.

14.

Hao

. Harnessing artificial intelligence for cancer rehabilitation: A call to action. Support Care Cancer. 2025;33(11):961. doi:10.1007/s00520-025-10047-1.

15.

Courneya

Vardy

O’Callaghan

, et al. Structured Exercise after Adjuvant Chemotherapy for Colon Cancer. N Engl J Med. 2025;393(1):13-25. doi:10.1056/NEJMoa2502760.

16.

Liu

Cao

, et al. Artificial intelligence: A powerful paradigm for scientific research. Innov Camb. 2021;2(4):100179. doi:10.1016/j.xinn.2021.100179.

17.

Storey

Yue

Zhao

Lukyanenko

. Generative Artificial Intelligence: Evolving Technology, Growing Societal Impact, and Opportunities for Information Systems Research. Inf Syst Front. 2025;27:2081-2102. doi:10.1007/s10796-025-10581-7. Published online February 25, 2025.

18.

Chen

Parsa

Swanson

, et al. Large language models in oncology: a review. BMJ Oncol. 2025;4(1):e000759.

19.

Alowais

Alghamdi

Alsuhebany

, et al. Revolutionizing healthcare: the role of artificial intelligence in clinical practice. BMC Med Educ. 2023;23(1):689. doi:10.1186/s12909-023-04698-z.

20.

Sangers

Reeder

van der Vet

, et al. Validation of a Market-Approved Artificial Intelligence Mobile Health App for Skin Cancer Screening: A Prospective Multicenter Diagnostic Accuracy Study. Dermatol Basel Switz. 2022;238(4):649-656. doi:10.1159/000520474.

21.

Wang

Zhao

Marostica

, et al. A pathology foundation model for cancer diagnosis and prognosis prediction. Nature. 2024;634(8035):970-978. doi:10.1038/s41586-024-07894-z.

22.

MHN

Nguyen

TPT

, et al. An in-depth review of AI-powered advancements in cancer drug discovery. Biochim Biophys Acta BBA - Mol Basis Dis. 2025;1871(3):167680. doi:10.1016/j.bbadis.2025.167680.

23.

Sinha

Vegesna

Mukherjee

, et al. PERCEPTION predicts patient response and resistance to treatment using single-cell transcriptomics of their tumors. Nat Cancer. 2024;5(6):938-952. doi:10.1038/s43018-024-00756-7.

24.

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;5(8):1158-1175. doi:10.1038/s43018-024-00772-7.

25.

Kolla

Parikh

. Uses and limitations of artificial intelligence for oncology. Cancer. 2024;130(12):2101-2107. doi:10.1002/cncr.35307.

26.

Reddy

Nafees

Raman

. Recent advances in artificial intelligence applications for supportive and palliative care in cancer patients. Curr Opin Support Palliat Care. 2023;17(2):125-134. doi:10.1097/spc.0000000000000645.

27.

Laranjo

Dunn

Tong

, et al. Conversational agents in healthcare: a systematic review. J Am Med Inform Assoc. 2018;25(9):1248-1258.

28.

Zhang

Lange

Fukuoka

. Artificial intelligence chatbot behavior change model for designing artificial intelligence chatbots to promote physical activity and a healthy diet. J Med Internet Res. 2020;22(9):e22845.

29.

Sanders

Chow

JCL

. Chatbot for Health Care and Oncology Applications Using Artificial Intelligence and Machine Learning: Systematic Review. JMIR Cancer. 2021;7(4):e27850. doi:10.2196/27850.

30.

Chen

Parsa

Hope

, et al. Physician and Artificial Intelligence Chatbot Responses to Cancer Questions From Social Media. JAMA Oncol. 2024;10(7):956-960. doi:10.1001/jamaoncol.2024.0836.

31.

Boie

Glastetter

Lux

, et al. Evaluating a Chatbot as a Companion for Patients With Breast Cancer: Collaborative Pilot Study. JMIR Cancer. 2025;11:e68426. doi:10.2196/68426.

32.

Wang

, et al. Development of a Machine Learning Algorithm-Based Predictive Model for Physical Activity Levels in Lung Cancer Survivors: A Cross-Sectional Study. J Clin Nurs. 2026;35:719-728. doi:10.1111/jocn.70033. Published online July 8, 2025.

33.

Cheville

Kornblith

Basford

. An Examination of the Causes for the Underutilization of Rehabilitation Services Among People with Advanced Cancer. Am J Phys Med Rehabil. 2011;90(5):S27-S37. doi:10.1097/PHM.0b013e31820be3be

34.

Novo

Thomas

Khouri

, et al. Machine Learning–Driven Phenogrouping and Cardiorespiratory Fitness Response in Metastatic Breast Cancer. JCO Clin Cancer Inform. 2024;8(8):e2400031. doi:10.1200/cci.24.00031

35.

Buffart

Kenkhuis

Newton

May

Galvão

Courneya

. Accelerating precision exercise medicine in cancer patients using pooled individual patient data: POLARIS experience. JNCI Cancer Spectr. 2025;9:pkaf078. doi:10.1093/jncics/pkaf078. Published online August 7, 2025.

36.

Hassoon

Schrack

Naiman

, et al. Increasing physical activity amongst overweight and obese cancer survivors using an alexa-based intelligent agent for patient coaching: protocol for the physical activity by technology help (PATH) trial. JMIR Res Protoc. 2018;7(2):e9096.

37.

Schmitz

Baker

Werts-Pelter

, et al. Nurse AMIE randomized controlled trial to address symptom management among rural advanced Cancer patients: Addressing malignancies in everyday life. Contemp Clin Trials. 2025;155:108005. doi:10.1016/j.cct.2025.108005.

38.

Schmitz

Schleicher

Doerksen

, et al. Testing the acceptability and feasibility of a tablet-based supportive cancer platform for patients with metastatic breast cancer. J Cancer Surviv. 2021;15(3):410-413. doi:10.1007/s11764-021-01021-x.

39.

Gordon

Qiu

Doerksen

, et al. Addressing metastatic individuals everyday: Rationale and design of the nurse AMIE for Amazon Echo Show trial among metastatic breast cancer patients. Contemp Clin Trials Commun. 2023;32:101058. doi:10.1016/j.conctc.2023.101058.

40.

Moreno-Gutierrez

Postigo-Martin

Damas

, et al. ATOPE+: An mHealth System to Support Personalized Therapeutic Exercise Interventions in Patients With Cancer. IEEE Access. 2021;9:16878-16898. doi:10.1109/access.2021.3049398.

41.

Gao

Ren

, et al. Development of an interpretable machine learning-based intelligent system of exercise prescription for cardio-oncology preventive care: A study protocol. Front Cardiovasc Med. 2023;9:1091885.

42.

Canzone

Belmonte

Patti

, et al. The multiple uses of artificial intelligence in exercise programs: a narrative review. Front Public Health. 2025;13:1510801. doi:10.3389/fpubh.2025.1510801.

43.

Zhang

Fang

Fukuoka

. A systematic review of artificial intelligence chatbots for promoting physical activity, healthy diet, and weight loss. Int J Behav Nutr Phys Act. 2021;18(1):160. doi:10.1186/s12966-021-01224-6.

44.

Luo

Aguilera

Lyles

Figueroa

. Promoting Physical Activity Through Conversational Agents: Mixed Methods Systematic Review. J Med Internet Res. 2021;23(9):e25486. doi:10.2196/25486.

45.

Singh

Olds

Brinsley

, et al. Systematic review and meta-analysis of the effectiveness of chatbots on lifestyle behaviours. Npj Digit Med. 2023;6(1):118. doi:10.1038/s41746-023-00856-1.

46.

Tangsrivimol

Darzidehkalani

Virk

HUH

, et al. Benefits, limits, and risks of ChatGPT in medicine. Front Artif Intell. 2025;8:1518049. doi:10.3389/frai.2025.1518049.

47.

Dergaa

Ben Saad

El Omri

, et al. Using artificial intelligence for exercise prescription in personalised health promotion: A critical evaluation of OpenAI’s GPT-4 model. Biol Sport. 2024;41(2):221-241. doi:10.5114/biolsport.2024.133661.

48.

Bickmore

Caruso

Clough-Gorr

Heeren

. It’s just like you talk to a friend’ relational agents for older adults. Interact Comput. 2005;17(6):711-735. doi:10.1016/j.intcom.2005.09.002.

49.

Sillice

Morokoff

Ferszt

, et al. Using Relational Agents to Promote Exercise and Sun Protection: Assessment of Participants’ Experiences With Two Interventions. J Med Internet Res. 2018;20(2):e7640. doi:10.2196/jmir.7640.

50.

Al-Akayleh

Al-Remawi

Ali

AASA

. AI-Driven Physical Rehabilitation Strategies in Post-Cancer Care. In: 2024 2nd International Conference on Cyber Resilience (ICCR); Dubai, United Arab Emirates; 2024:1-6. IEEE. doi:10.1109/iccr61006.2024.10532883.

51.

Kennedy

Bayes

Newton

, et al. Implementation barriers to integrating exercise as medicine in oncology: an ecological scoping review. J Cancer Surviv. 2022;16(4):865-881. doi:10.1007/s11764-021-01080-0.

52.

Hilbers

Nekain

Bates

Nunez

. Patient Attitudes Toward Artificial Intelligence in Cancer Care: Scoping Review. JMIR Cancer. 2025;11:e74010. doi:10.2196/74010.

53.

Van Kessel

Wong

BLH

Clemens

Brand

. Digital health literacy as a super determinant of health: More than simply the sum of its parts. Internet Interv. 2022;27:100500. doi:10.1016/j.invent.2022.100500.

54.

World Health Organization . Ethics and Governance of Artificial Intelligence for Health: WHO Guidance. Geneva, Switzerland: World Health Organization; 2021.

55.

Geis

Brady

, et al. Ethics of Artificial Intelligence in Radiology: Summary of the Joint European and North American Multisociety Statement. J Am Coll Radiol. 2019;16(11):1516-1521. doi:10.1016/j.jacr.2019.07.028.

56.

Jiang

, et al. Opportunities and challenges of artificial intelligence in the medical field: current application, emerging problems, and problem-solving strategies. J Int Med Res. 2021;49(3):3000605211000157. doi:10.1177/03000605211000157.

57.

Rajkomar

Hardt

Howell

Corrado

Chin

. Ensuring Fairness in Machine Learning to Advance Health Equity. Ann Intern Med. 2018;169(12):866-872. doi:10.7326/M18-1990.

58.

Mittelstadt

Floridi

. The Ethics of Big Data: Current and Foreseeable Issues in Biomedical Contexts. Sci Eng Ethics. 2016;22(2):303-341. doi:10.1007/s11948-015-9652-2.

59.

De Micco

Grassi

Tomassini

Di Palma

Ricchezze

Scendoni

. Robotics and AI into healthcare from the perspective of European regulation: who is responsible for medical malpractice? Front Med Lausanne. 2024;11:1428504. doi:10.3389/fmed.2024.1428504.

Artificial Intelligence Meets Cancer Rehabilitation: Emerging Evidence for Exercise and Physical Activity Interventions

Abstract

Plain Language Summary

Keywords

1. Introduction

1.1. What Is Artificial Intelligence and How Is It Used in Cancer Care?

2. Current Landscape: AI Applications in Cancer Rehabilitation

2.1. Search Strategy

2.2. Screening and Early Identification

2.2. Predicting Exercise Training Responses and Outcomes

2.3. Behaviour Change and Engagement Tools

2.4. Precision Exercise Prescription Systems

3. Challenges and Future Directions

4. Current Research Questions

5. Important Considerations for AI-Related Research

5.1. Informed Consent, Data Ownership, and Security

5.2. Bias and Fairness

5.3. Human Oversight and Trust

5.4. Regulation and Legal Accountability

6. Conclusion

Footnotes

ORCID iDs

Ethical Considerations

Author Contributions

Funding

Declaration of Conflicting Interests

Data Availability Statement

References