Invited Commentary: Transforming Occupational Therapy Research With Artificial Intelligence

Abstract

Artificial intelligence (AI) in occupational therapy is growing in scope and scale. AI use in occupational therapy practice and education has been explored recently; however, it is imperative that the field must reflect on the use of AI in occupational therapy research. In this invited commentary, we highlight recent research in occupational therapy that effectively harnesses the power of AI and algorithms. We also discuss the lessons that have been learned and the considerations we must take to ensure we produce high-quality evidence to support occupational therapy practice. Finally, we provide recommendations for occupational therapy researchers and the field to ensure best practices and excellence in research.

Plain Language Summary

Artificial intelligence, commonly shortened to “AI”, involves the use of computers to complete tasks that usually require human thinking. AI has been used more and more in healthcare, including the practice of occupational therapy. Research is important to develop new treatments, tools, and other outputs that support occupational therapy practice. We do not currently have any recommendations for using AI in occupational therapy research. In this article, we discuss some of the current research in occupational therapy that uses AI. We also discuss some of the challenges and lessons learned in that research. Importantly, there are many things to consider before a researcher uses AI in their research. Lastly, we provide suggestions and recommendations for researchers in occupational therapy that want to use AI.

Introduction

Research in occupational therapy supports the advancement of occupational therapy clinical practice and builds the evidence base for occupational therapy service providers. There is a rapidly growing field of research on the clinical implications of using artificial intelligence (AI) in occupational therapy clinical practice, but less has been written on AI for occupational therapy research. AI is simply defined by the Oxford dictionary as the science of analyzing and extracting information from large sets of data. The rapid advancement of generative AI tools can greatly enhance the critical thinking and clinical reasoning required for skilled occupational therapy services. As detailed in their 2025 paper, Drs. Karen Jacobs and Alyson Stover call on occupational therapy practitioners to “embrace AI as a partner” (p. e7) and consider how to thoughtfully integrate AI into their clinical practice workflows (Stover & Jacobs, 2025); this recommendation is in line with a recent call to utilize collaborative AI in rehabilitation, in which AI and clinicians work together with AI distilling large amounts of complex data that clinicians can then use to augment, but not replace, their own clinical decision-making (Liew et al., 2025).

With regard to AI for occupational therapy practice, a recent survey and interviews of Canadian occupational therapy practitioners indicated that practitioners simultaneously recognize the benefit and potential/actual use of AI in clinical practice, and realize the limitations, concerns, and fears surrounding use and development (Matharu et al., 2025). Study participants highlighted the benefits of AI in clinical practice including increasing efficiency of documentation, assisting with assessments, alerting providers to changes in health status, and personalization of interventions. They also raised concerns around safety, privacy, and the perception of the use of AI in health care by colleagues and clients alike. Similarly, others have highlighted concerns in precision medicine around patient privacy, bias in AI models (e.g., overfitting models, models based on skewed representations of certain populations), trustworthiness of AI, replication challenges, and avoiding the dehumanization of medicine (Abujaber & Nashwan, 2024; Briganti & Le Moine, 2020; Liew et al., 2025).

Regarding AI and occupational therapy research, similar concerns have been raised in clinical research design and execution. Park and colleagues outline the use of AI as it pertains to the stages of health care clinical trials (2020). Stage 0 is idea conception and initial experiments, which involves the careful development of AI-based algorithms alongside substantial assessment of user needs. In Stages 1 and 2, which are safety and initial efficacy, respectively, the algorithm model performance. Clinical consequences of those algorithms are evaluated alongside clinical efficacy. Stages 3 and 4 involve exploring algorithm performance and outcomes in the short- and long-term alongside clinical effectiveness and implementation. The authors also include recommendations for using AI to assist with study designs at each phase. Complementing this work, Askin and colleagues make recommendations for how to use AI in research such as for optimizing trial design, subject recruitment, safety oversight, and handling missing data (Askin et al., 2023). Taken together, the stage has been readily set for researchers to move the field forward using AI.

In this invited commentary, we discuss current approaches in occupational therapy that involve the use and/or development of AI, including the use, management, and analysis of large data sets. Occupational therapy researchers are working in all stages outlined by Park et al. (2020), and we highlight a few cases that demonstrate responsible and ethical forward progress regarding AI in occupational therapy research. Furthermore, we see researchers attempting the seemingly impossible task of answering the question raised by Kaelin and colleagues: “How should complex OT concepts and values be translated into machine-readable information without losing important information and while ensuring AI technology is meaningful for OT and OT clients?” (Kaelin et al., 2024, p. 3) Taking all of this work into account, we conclude by making recommendations for the future of occupational therapy research in AI, including setting the stage for truly collaborative AI (Liew et al., 2025).

Current Occupational Therapy Research Using AI

The boom in rehabilitation technology over the past two decades has facilitated the growth and application of algorithms and AI in real-world participant monitoring. One of the most exciting examples is the uptick in technology created for use in the home, capturing data in a space that has not been well-integrated into health care research. The availability and lower cost of mobile devices, smart home sensors, and wearable technologies have opened avenues for the collection of data and intervention outcomes that reflect real-world performance to occupational therapy researchers. Previously, we relied on participant self-reports when capturing home-based data. Although participant report provides rich subjective information that is vital not only to treatment planning but also to accurate assessments and algorithms to predict outcomes, objective, quantitative data based on activity patterns in the home and factors such as falls, walking, upper body activities, and community participation provide an additional level of understanding that can greatly enhance both research and clinical decisions (Proffitt et al., 2024). In addition, in-home sensors can provide an unobtrusive way to monitor participants in their home spaces and provide health metrics to the participants themselves, their care partners, and their health care providers in real-time. The integrated algorithms in these systems can provide alerts about changes in function and status, oftentimes before clinical assessments and the participant themselves notice changes (Demiris et al., 2025; Dodge et al., 2015; Wu et al., 2021). These alerts can be critical in participant populations with conditions such as Amyotrophic Lateral Sclerosis (ALS) or Parkinson’s disease (PD), where disease progression can be rapid, but health care or research visits are only scheduled every 3 to 6 months. Research has begun to explore the development of predictive algorithms, based on in-home sensor data, to detect when significant changes occur for individuals with neurodegenerative or cardiovascular diseases. Early detection of functional declines and safety concerns may improve personalized care (aka precision rehabilitation) at an earlier stage when intervention can be more effective (Janes et al., 2025; Marchal et al., 2024).

The use of AI in occupational therapy research has also focused on advancing precision medicine, and more specifically, precision rehabilitation, which aims to provide the right treatment at the right time for each person (Collins & Varmus, 2015; French et al., 2022). These efforts assume that a given therapy might be suitable for some people but not all, and that the timing and dosage may differ among individuals based on their unique personal characteristics. This is due to the inherent heterogeneities of disease progression, patient comorbidities, age effects, and social determinants of health, to name but a few factors, which render a one-size-fits-all approach to rehabilitation ineffective. However, it currently remains unclear who will benefit from what rehabilitation treatment; developing algorithms that can help identify and predict responders to specific treatments is a critical area of occupational therapy research where AI has been helpful. In addition, many patient-centered outcomes, such as measures of cognition and activities of daily living (ADLs) are inherently variable, and their clinical meaningfulness, as well as research utility, differs across individuals. For example, a 1-point change in cognition may not have an impact on a healthy young adult but may represent a highly meaningful sign of recovery in a stroke participant. Even no change in cognition may be meaningful if the person was expected to decline without therapy. Still, in current practice, discerning true responses in many patient-centered outcomes remains difficult. This challenge also extends to clinical trials where conventional analyses focus on average treatment effects (ATEs), which do not account for the proportion of participants who genuinely benefit and are susceptible to outliers, particularly in smaller trials.

An AI/machine learning (ML)-driven approach is to consider the individual-level treatment response (ITR), which estimates the probability that an individual will respond to a given treatment. Several AI/ML algorithms exist for this purpose. Counterfactual prediction analysis with permutation tests, for example, estimates the probability of improvement or decline had a participant not received the treatment, providing a personalized measure of net change. Excitingly, the ITR method has been applied in clinical trial research already (Pang et al., 2024; Wu et al., 2024; Zhao et al., 2013) and has shown a high potential to strengthen occupational therapy clinical trials. One application is called virtual controls (Wu et al., 2025). The idea of virtual controls (also known as twin-controls, synthetic controls, or digital twins) is to create replicas of patient cohorts that can be used as a surrogate to predict the effects of treatments on a personalized level. The U.S. Food and Drug Administration (FDA) and European Medicines Agency are also increasingly emphasizing virtual control methods as part of their initiatives to modernize trial development. Utilizing natural history cohorts containing extensive records on disease progression or disability trajectory, such as the National Alzheimer’s Coordinating Center Uniform Data Set (NACCUDS) and Health and Retirement Study (HRS) data set, researchers can create another arm in their trials for replicability. This method could potentially lead to precisely matched control groups and reliable trial evaluations, while addressing key methodological issues in the field. For example, Phase I–II trials have limited sample sizes, and their lack of replicability is often one of the reasons for failed Phase III studies. With a synthetic-control group design, researchers can identify the characteristics of treatment responders to further develop larger trials with a higher likelihood of success. Digital twin methods are also used in current precision rehabilitation algorithms that attempt to predict likelihood of response to certain treatments or treatment schedules given specific individual characteristics (Liew et al., 2025). Notably, using virtual controls, or digital twins, requires large, unbiased data sets that can be used to generate reliable virtual controls (see below regarding large data sets).

Finally, AI is starting to be explored in many other ways throughout occupational therapy research. Recent reviews demonstrate that AI is being used in occupational therapy research primarily in the areas of robotics (Bulan et al., 2025), motor assessments, and telerehabilitation (Kaelin et al., 2024), as well as generally within human–computer interactions (Kansizoglou et al., 2025). There are also burgeoning efforts to use AI to begin to generate predictive models that are informed by occupational therapy researchers. For instance, an algorithm was recently trained to predict early autism diagnoses based on both the child’s and the mother’s risk factors in electronic health records (EHRs) data (Li et al., 2025); another AI algorithm that assesses the “brain age” of individuals was found to predict poststroke sensorimotor and cognitive outcomes (Marin-Pardo et al., 2025). These types of predictive algorithms, when validated with real-world data, can be extremely useful in identifying people who may benefit from more careful screening, which could enable earlier treatment, and those who may benefit from specific focuses for more efficient rehabilitation treatment, respectively.

Implementing AI Today: Factors to Consider

Although we advocate strongly for the implementation and utilization of AI in occupational therapy research as well as in clinical practice, it is also essential that we continue to consider the practical risks and limitations. Although the potential for improved health care outcomes with AI is remarkable, the implementation of AI into our research must be done with full consideration of the limitations and even potential harm that AI also holds.

Inequity and Algorithm Bias

AI models, and in fact all large data set models, cannot accurately extrapolate beyond the data used to train the model. While this core tenet of statistics should be ingrained in the training of every researcher, AI is a tool that is not always transparent about the data being used to create the logic being used for decision-making. If the training data for an AI model is not diverse, the resulting model can perpetuate or worsen existing health care disparities and further bias in health care (Cross et al., 2024). The resulting misdiagnoses or suboptimal treatment strategies will disproportionately affect minority, underserved and/or underrepresented populations. Unfortunately, this is a well-documented phenomenon (Norori et al., 2021) that has been demonstrated across race (Chowkwanyun & Reed, 2020), skin color (Wen et al., 2022), gender (Maserejian et al., 2009; Obermeyer et al., 2019), and age (Choi et al., 2025; Muralidharan et al., 2024). Beyond the biases created by the data, or lack of data, the biases of the controlling interests must also be considered. All AI and ML algorithms were designed to some extent by humans. Decisions to include certain data sources or exclude particular geographic regions are made by humans who have their own set of assumptions and biases. Transparency around decisions made relevant to coding parameters, data inclusions and exclusion, and reporting metrics help to make these biases known, but do not remove the potential for harm. There is growing attention to ethical considerations when using AI within occupational therapy specifically, also around these issues of bias and transparency (Kaelin et al., 2024).

Data Security and Privacy

As discussed previously, large data sets are required for accurate AI processing. The vast amount of sensitive patient/participant data extracted from sources such as EHRs, research databases, and genomic data creates significant challenges to ensuring data security, regulatory compliance (e.g., Health Insurance Portability and Accountability Act [HIPAA]), and prevention of malicious re-identification of “anonymized” data. In addition, AI tools are often available to multiple end users, thus further complicating data control and leading to concerns related to data privacy. Tools to protect large-scale data privacy are currently being developed to address these issues.

Accountability

If data from a research lab are deidentified and shared with an AI agent, and a bad actor used the data to reidentify the sample and acquire protected health data about a population of older, marginalized adults with a neurological condition, who is responsible? The researcher who provided the data? The owner of the central repository who granted individual access? The local institutional review board (IRB) who approved the study? As data sets grow, control of the data and accountability for its use, and misuse, is essential. Similarly, the “black box” nature of many AI models makes it difficult to understand how a diagnosis or conclusion was reached. If an AI system is used by a researcher or clinician to design a personalized treatment plan for a participant or client, and the AI tool makes an error that causes harm, who is accountable or liable for the error? These are critical questions for which there are currently no clear answers, but which should be carefully considered for each new application. These ethical issues should be paramount in discussions of the use of AI in occupational therapy research.

Translation Failures

Machine learning models have been around for decades. However, algorithms that perform well in controlled lab settings with ideal controls and clean data sets often fail in complex real-world settings where data are not always complete and patients do not always follow standardized protocols. The shift from clean experimental data to messy real-world data is highlighting new issues that are emerging due to this. A comparison has been made between the promises of the EHR in the early 2000s and the promises of AI today (Rose & Chen, 2024). Initially, EHR adoption was driven by intentions toward reducing health care costs and increasing efficiency. Unfortunately, we see interfaces and workflows that are complex, frustrating, and decrease overall efficiency for health care workers. The promises of quality were usurped by a rush toward availability (Rose & Chen, 2024). It is critical that any use of AI is supervised and validated by a knowledgeable human who can use AI as a tool that requires oversight, rather than blindly trusting the AI to complete the task correctly.

Looking Forward to AI in Occupational Therapy Research

Occupational therapy research has already begun to benefit tremendously from the use of AI, and this is only expected to grow in the upcoming years. As the field is rapidly evolving, it is time for researchers, clinicians, and educators to consider where AI might benefit their current work and to develop ways to stay abreast of AI as new advances emerge.

AI methods are a critical means for advancing occupational therapy research across a diverse range of fields including health outcomes research using wearable sensors (Hicks et al., 2019), rehabilitation-related meta-analyses (Lang et al., 2015; Lohse et al., 2014), health services research (Ottenbacher et al., 2019), and stroke neurorehabilitation research (Liew et al., 2022). Accordingly, experts have identified an emergent need to modify and expand education and training to better incorporate data science procedures into rehabilitation research, particularly for occupational therapy practitioners (Ottenbacher et al., 2019). However, most occupational therapists do not have training in or exposure to the basics of computer programming or engineering. Thus, a common question among occupational therapy practitioners is: “where should researchers who want to learn these skills start?” Fortunately, free online courses and programs exist to help teach these skills. In addition to the many general free courses, such as those found on Coursera or YouTube, rehabilitation-specific programs, such as the National Institutes of Health (NIH)-funded Reproducible Rehabilitation Research Education Program (ReproRehab), are addressing this gap by providing hands-on assistance to researchers and professionals to teach them how to use computer programming and data science methods to enhance their research (https://www.reprorehab.usc.edu/). ReproRehab also provides public resources for the broader rehabilitation research community in the form of open-access, curated web database of resources for rehabilitation scientists (https://reprorehabdb.usc.edu/), YouTube tutorials, hands-on exercises on GitHub, and more, most of which are designed for people who have no prior background in computer programming or engineering. By gaining a foundation on how computers process information, and how AI uses data to generate new information, occupational therapy researchers will be well-positioned to better understand new advances as they emerge. Understanding the basics of how AI works will make talking and collaborating with statisticians, data scientists, and data managers, who have different mindsets about mathematics, modeling, and technical skills, more efficient and provide a common vocabulary to facilitate interdisciplinary communication.

Furthermore, a key component of using AI is data management. Terms such as “data scrubbing” or “data cleaning” refer to the process of organizing data sets so that they are consistent and complete. This includes removing or filling in any missing values, correcting any inconsistent labels or values (e.g., days posttreatment vs. months posttreatment; age coded as “Age” in one data set and “AGE” in another), and removing any information that is irrelevant to the overall data set (e.g., visualizing your data, identifying key columns, and removing redundant or irrelevant ones). Doing this allows a computer to accurately read and process a data set, which is critical for using AI (see ReproRehab YouTube for talks on practical data management tips). This preparatory work often is manual and tedious, although proper planning and maintenance of data management practices at the outset of starting a study can save enormous amounts of time compared with trying to make sense of messy data at the end of the study. This work can also be thought of as getting data “AI-ready.”

Furthermore, using a common data standard, or common data model, can help to ensure data are AI-ready similarly across research labs and individual researchers so that data sets can become interoperable. A new NIH P50-funded Data Science and Analytics for Precision Rehabilitation (DAPR) Center aims to help rehabilitation researchers prepare, manage, and store data using common data models (https://dapr.usc.edu/). The DAPR Center will provide hands-on assistance to researchers wishing to harmonize their data, along with recommendations of data standards for storing rehabilitation-specific terms that may not exist in other common data models, such as the NINDS CDE (Saver et al., 2012) or OMOP (Hripcsak et al., 2015).

Finally, there are many key advantages to getting data AI-ready in a standardized manner. First, most AI-based approaches require large, diverse data sets to train and test algorithms that are accurate, reliable, and generalizable. However, there is often a lack of large, detailed, AI-ready data sets as rehabilitation research studies are usually collected in smaller samples (e.g., 50–100 subjects, versus the 1,000 subjects needed by AI algorithms). The creation of AI-ready data sets means that data sets from multiple studies on the same topic can be pooled together, which allows for larger, harmonized data sets to support better-powered analyses. One example of this driving principle at work is the ENIGMA Stroke Recovery working group, which is led by an occupational therapy researcher (Liew) and which brings together high-resolution brain imaging and behavioral data from over 2,000 stroke patients worldwide, through a consortium of over 100 stroke researchers working together (Liew et al., 2022). In ENIGMA, individual stroke research groups collect smaller samples for their own research studies and then contribute this data to ENIGMA, where it is harmonized into one large, AI-ready data set centrally by the core ENIGMA Stroke Recovery team. This large data approach has enabled well-powered analyses that have detected more subtle patterns in brain repair and recovery over time (Domin et al., 2023; Ferris et al., 2023; Liew et al., 2021, 2022, 2023; Zavaliangos-Petropulu, Lo, et al., 2022; Zavaliangos-Petropulu, Tubi, et al., 2022). Extending this approach to other areas of occupational therapy research where large data sets are hard to collect, but where data from smaller studies can be combined for larger analyses, could improve the statistical power and generalizability of occupational therapy research.

Finally, occupational therapy researchers should be knowledgeable about the processes they are using and what they expect the AI to output. That is, researchers should know the “ground truth” they are expecting from the AI model and carefully test and validate that the model is providing them with the expected responses. This may sometimes mean breaking down the task into smaller subtasks for AI that can be easily verified. Researchers can also use concepts such as virtual controls or digital twins, as mentioned previously, to confirm that the AI is producing the expected output. Although AI can be a powerful tool, knowledgeable clinical oversight is always required to ensure responsible and effective use.

Conclusion

AI is here to stay. As highlighted earlier in the paper, occupational therapy researchers are responsibly developing, using, and implementing AI across all stages of research. Furthermore, new tools and approaches have already been developed that are being implemented in clinical occupational therapy research and practice. We hope that the advice and recommendations in this paper are strongly considered by researchers, practitioners, institutions, and systems at all levels of influence, and we look forward to the future of AI and occupational therapy research. (And no, we did not write this paper using AI!)