Seeing Past the Event Horizon: A Framework for Integrating Artificial Intelligence and Machine Learning Into Physical Therapy

Artificial intelligence (AI) has emerged as a prominent force within various realms of health care, promising to revolutionize clinical decision-making, reduce health care expenditures, and enhance patient outcomes.1 The exciting potential of AI, and particularly machine learning (ML) for clinicians such as physical therapists, lies in its capacity to undertake intricate tasks and identify subtle patterns within diverse datasets.2 Despite apprehensions surrounding the possibility of AI/ML automation displacing health care providers, physical therapists ultimately stand to gain from its integration by leveraging domain-specific augmentation and task automation. However, the swift advancement of this technology underscores the pressing need for the physical therapist field to delineate optimal strategies for both developing, integrating, and adopting AI/ML. Considering these opportunities and challenges, this paper seeks to provide a comprehensive framework for the development, integration, and adoption of AI/ML within physical therapist practice. By addressing the complexities of AI/ML implementation while highlighting the potential benefits for both physical therapists and patients, this paper aims to catalyze meaningful discussions and actions to ensure the responsible and equitable integration of AI/ML technologies into physical therapist care delivery. AI, broadly defined, refers to the simulation of human-like intelligence by machines, enabling them to perform tasks that typically require human intelligence, such as learning, reasoning, and problem-solving. Most recently, generative AI, such as large language models, have gained notoriety for their ability to generate expert-level responses to queries such as generating summaries of patient data.3 Within the realm of AI, ML is a subset of techniques that focus on the development of algorithms allowing computers to improve their performance on a task through experience, without being explicitly programmed. ML encompasses various approaches, including supervised learning, where models are trained on labeled data to make predictions or classifications; unsupervised learning, which involves finding hidden patterns or structures in unlabeled data; and reinforcement learning, where algorithms learn to make decisions by interacting with an environment and receiving feedback. ML models can be trained on large datasets of patient outcomes and clinical variables to predict optimal treatment strategies or detect early signs of musculoskeletal disorders. AI/ML has potential to improve how clinical research is conducted, reduce the cost of developing evidence-based interventions, and potentially increase the efficacy and generalizability of clinical trials.3 Bringing efficacious interventions and technologies to clinical practice takes an immense amount of time and resources.4 AI/ML can improve pre-trial planning, patient selection, and data analytics by managing large and heterogeneous sources of data, identifying unseen patterns, and predicting endpoint outcomes. Importantly for physical therapists, AI/ML can be used for multiple applications salient to physical therapy including diagnosis prediction, personalization of care, and predicting patient events and outcomes.1 For example, Rommers et al5 used ML to determine injury risk in youth athletes with 85% precision. Additionally, ML has been used to predict intervention adherence with an area under the curve of 0.90 and poststroke upper-extremity function, assessed using the Fugl-Meyer Assessment, after physical therapy (root of mean squared error of 0.36; R2 = 0.91).6 Together, optimizing and potentially increasing the successful development of efficacious interventions and improving clinical decision making and potentially improve patient outcomes by delivering more personalized and data-driven care. Excitement surrounding the potential impact of AI/ML applications in clinical practice often overlooks significant practical challenges in their translation. One such challenge lies in the requirement for extensive datasets for training and validation. The acquisition of large and diverse datasets raises questions about their source, accessibility, and reliability in reflecting real-world clinical scenarios without inherent biases. Addressing these concerns necessitates collaboration between health care institutions, research organizations, and technology developers to establish robust data-sharing networks. Initiatives promoting data standardization and interoperability can facilitate the pooling of data from multiple sources while ensuring compliance with privacy regulations and ethical guidelines. A formidable hurdle to AI/ML algorithm development is access to high-quality data is crucial to the development of reliable AI/ML algorithms. Rigorous data curation processes, including data cleaning, normalization, and validation, are essential to minimize biases and ensure the generalizability of the models across diverse patient populations and clinical settings.7 Additionally, transparency in data collection methods and validation procedures is paramount to instill trust in the reliability and integrity of the datasets used for AI/ML algorithm development. Furthermore, prospective, multisite study designs are imperative for validating AI/ML algorithms in real-world clinical settings. This involves separating training and validation datasets to mitigate the risk of overfitting and ensuring that the algorithms perform consistently across different geographies, centers, or patient populations. Ultimately, collaborative efforts among researchers, clinicians, and data scientists are essential for designing robust study protocols and implementing standardized reporting practices to enhance the reproducibility and transparency of AI/ML research in health care. Recently, researchers have created standardized reporting tools for AI/ML algorithm development studies (TRIPOD2,8 TRIPOD-AI,9 DECIDE-AI10) and clinical trials (CONSORT-AI11). Additionally, the Food and Drug Administration has begun to develop a regulatory framework for AI/ML algorithms as medical devices.12 Reporting of algorithm methodologies and clear regulatory pathways for AI/ML products may address concerns over algorithm transparency and explainability. Once AI/ML algorithms are ready for widespread deployment, additional logistic barriers of data curation and management, in heterogeneous clinical environments, may present an even more challenging hurdle for real-world use. The performance of AI/ML algorithms is largely beholden upon the quality of input data. As such, clinics will have to ensure the proper infrastructure, resources, and workflows are in place prior to deploying AI/ML algorithms into clinical practice. This will require extensive partnership between clinicians, administrators, and AI/ML developers to ensure algorithms are able to perform in their intended manner. The clinical translation of AI/ML into physical therapy requires careful consideration over the level of autonomy provided to AI/ML algorithms in clinical decision making. Assuming sound execution of algorithm development, the decision on where the tools will sit on the spectrum of assistive to autonomous should account for the algorithm’s performance, complexity of the model, and acceptability from patients, payers, and providers.13 Bitterman et al14 proposed a graded categorization of AI/ML automation for medical applications which consists of 2 levels of assistive (ie, Level 1: Data presentation; Level 2: Clinical decision-support) and 3 levels of autonomous (ie, Level 3: Conditional automation; Level 4: High automation; Level 5: Full automation). This model accounts for the agent responsible for decision making, domain or task-specificity, and party ultimately liable for decisions. For example, an AI/ML algorithm developed for diagnosis and therapy recommendations in subacromial pain syndrome is presented in Figure 1. Within the assistive algorithm space, the clinician is ultimately responsible for decision making and assumes liability. In contrast, as autonomy increases, with levels such as conditional automation, high automation, and full automation, the responsibility for decision-making shifts toward the algorithm itself, and liability may shift accordingly to AI/ML developers or manufacturers. This transition raises important ethical and legal considerations regarding accountability, transparency, and patient safety. As AI/ML algorithms approach the point of clinical translation, physical therapists, AI/ML developers, and regulators will need to consider how the level of autonomy will not only influence clinical care, but also how the level of autonomy will influence clinical adoption, patient trust, and the (un)foreseen consequences of the algorithm’s integration into clinical workflows. Framework for decision support autonomy of AI/ML algorithms for physical therapist practice. The example provided is for a diagnostic tool for SPS. AI = artificial intelligence; AI/ML = artificial intelligence/machine learning; SPS = subacromial pain syndrome. AI/ML algorithm development model based on the SALIENT framework by Vergt et al.12 The example provided is for the development of a diagnostic tool for SPS. aPreparation phase. AI/ML = artificial intelligence/machine learning; HCI = human computer interface; SPS = subacromial pain syndrome. Furthermore, the level of autonomy conferred to AI/ML algorithms may influence patient trust and acceptance. Patients may exhibit varying levels of trust in AI/ML-enabled decision-making processes based on their perceptions of algorithmic reliability, transparency, and alignment with their preferences and values. As AI/ML autonomy increases, patients may express concerns regarding the loss of human oversight and the potential for algorithmic errors or biases to influence care decisions. Establishing patient trust in AI/ML-enabled care requires clear communication, transparency regarding algorithm capabilities and limitations, and opportunities for patient involvement in decision-making processes. Moreover, robust safeguards and accountability mechanisms are essential to mitigate risks and ensure patient safety and confidence in AI/ML-enabled care delivery. Bringing clinical decision and diagnostic support AI/ML algorithms to physical therapist will require alignment on a research framework for development. Van der Vegt et al15 proposed an end-to-end process for AI/ML algorithm development for clinical use, titled “The staged clinical AI implementation (SALIENT) framework.” SALIENT consists of 5 stages of implementation and 5 elements. In stage 1, the clinical problem or application is defined in which Element A is nested. Stages 2 to 5 pertain to increased levels of study phases and development. Element B comprises the development of the AI/ML model, the data pipeline, human-computer interface (HCI), and clinical workflow design and is nested across all 5 stages. Element C takes components from Element B into functioning systems over stages 3 to 5 in which the data pipeline, AI/ML model, and HCI are updated iteratively over each progressive stage. Element D, nested in stages 4 and 5, involves the integration of outputs from Element C into clinical settings and workflows. Finally, Element E entails the ongoing monitoring and evaluation of model performance across stages 3 to 5 (Fig. 2). While the SALIENT framework offers a structured approach, its complexity may not provide clear guidance for practitioners. Instead, let’s consider a more straightforward example to elucidate the development process. Take, for instance, the development of an AI/ML algorithm for the diagnosis and therapy recommendations in subacromial pain syndrome. The process typically involves defining the clinical problem or application, developing the AI/ML model, designing the data pipeline and HCI, integrating the system into clinical settings, and monitoring and evaluating the model’s performance iteratively (Fig. 2). Regardless of the approach to developing AI/ML algorithms for physical therapy, without accounting for key adoption drivers the ultimate realization of these tools will falter. Adler-Milstein et al16 recently presented a model of clinical adoption focused on 4 domains including the reasons, means, method, and desire to use AI/ML. While the SALIENT framework provides a means for AI/ML algorithm development, physical therapists’ adoption will require additional consideration during the developmental process to ensure these domains are addressed. At the forefront of physical therapist adoption is the ability of AI/ML to address a critical need or inefficiency within a desired clinical setting. Improved clinical decision making and patient outcomes will likely be the primary driver of physical therapist adoption. For instance, AI/ML could be used for identifying patients at risk of non-compliance with prescribed exercise regimens, allowing for early intervention and improved treatment adherence. Additionally, ML algorithms can analyze large datasets of patient outcomes to identify trends and patterns, informing personalized treatment plans tailored to individual patient needs. However, AI/ML algorithms targeting clinical inefficiencies which result in the optimization of how and what physical therapists do in the clinic, in conjunction with sufficient reimbursement, will likely be primary drivers of administrators to maximize financial margins. Ultimately, AI/ML developers need to consider these factors when creating algorithms for clinical use with concurrent alignment with physical therapists and administrators on standards for sufficient “reasons for use.” Physical therapists and AI/ML developers must account for data and computing infrastructure, as well as the human resources required to implement algorithms for clinical use. The input data (eg, electronic medical records, patient reported outcomes, wearables), level of training required to educate physical therapists, level of (pre)processing, and computing requirements will likely be heterogeneous across AI/ML algorithms and may not concurrently align with the capacity within health care systems considering adopting this technology. For example, the implementation of AI/ML-powered gait analysis systems in physical therapy clinics may require investments in specialized hardware and training for physical therapists to interpret and apply the insights generated by these systems effectively. Furthermore, health care systems may need to invest in additional resources to scale up the underlying infrastructure supporting and maintaining the AI/ML algorithm. This is not only a financial and resource barrier to adoption but may indirectly create health care delivery disparities based on socioeconomic disparities or geographic constraints. Ultimately, physical therapists need to advocate and incentivize AI/ML algorithms which optimize the balance between data and human infrastructure requirements and performance. Adoption of AI/ML tools will also be influenced by their usability and workflow integration to maximize uptake, while minimizing or optimizing the impact on clinical workflows. Setting, task, or domain of use aside, the tool must not burden physical therapists or interdisciplinary teams and should prioritize interpretability, generalizability, and user experience to minimize workflow disruption and ensure seamless integration into clinical practice. For example, generative AI documentation tools that automatically generate progress notes based on patient interactions can streamline administrative tasks, allowing physical therapists to focus more time and attention on direct patient care. However, if the generative AI system was only applicable to a specific patient population it may not achieve sufficient adoption due to suboptimal integration into a clinical workflow. Finally, ensuring professional fulfillment and trust in the algorithm decisions is fundamental in the adoption of AI/ML in physical therapy. The importance of developing AI/ML tools to support complex and difficult tasks, at the same time reducing or replacing the mundane, is important for ensuring physical therapists do not view the AI/ML tools as a replacement for themselves within the clinic. This balance would ensure the human-centric and rewarding contribution of physical therapy is maintained while mitigating the undesired aspects of physical therapists which contribute to burnout. Additionally, developing trust in AI/ML tools in physical therapy, through clinical evidence and education, may prove to bolster their desire to integrate the tools into clinical practice. Ultimately, striking a delicate equilibrium between preserving the human-centric and complex contributions of physical therapists and fostering trust through clinical evidence and education is pivotal for widespread adoption of AI/ML tools. As health care embarks on integrating AI/ML tools into clinical practice, it is imperative for practitioners, developers, and regulators to collaborate in shaping the future landscape. This paper provides a framework for the development and implementation of AI/ML algorithms. However, to realize the full potential of these tools while retaining the human-centric aspects of physical therapy, clinicians must actively engage in addressing key adoption drivers. This calls for a collective commitment from the physical therapy community, including the American Physical Therapy Association, to actively participate in the development, validation, and ethical deployment of AI/ML tools and governance, thereby ensuring a harmonious integration that enhances patient care and the professional satisfaction of physical therapy. The American Physical Therapy Association provides resources for physical therapists interested in developments in the digital health space at: https://www.apta.org/your-practice/practice-models-and-settings/digital-health-in-practice. Three motions were passed at the 2024 House of Delegates on inappropriate use of AI by payers and on digital health technologies, digital therapeutics, and AI in physical therapist practice; the final language is available on the APTA website. The American Physical Therapy Association supports the ethical development and integration of artificial intelligence that reduces administrative burden and enhances physical therapist practice, education, and research for the best interest of patients, clients, the profession, and society. The American Physical Therapy Association opposes the inappropriate use of artificial intelligence by payers to deny or restrict access to, or reduce payment for, physical therapist services or other medically necessary services. The American Physical Therapy Association believes digital health technologies, digital therapeutics, and artificial intelligence have the potential to augment physical therapist practice by expanding access, enhancing care delivery models, promoting safety, reducing administrative burden, and improving outcomes when all of the following criteria are met: Physical therapists, in collaboration with the patient or client, shall determine the optimal modes to access physical therapist services, whether this is in person, remotely, or some combination. A physical therapist is responsible for all aspects of patient and client management. Physical therapist services shall be conducted in a manner that allows for patient and client engagement and supports the therapeutic alliance (ie, fosters participation in the patient and client-provider relationship). The physical therapist and the physical therapist assistant adhere to scope of practice, including requirements that are based on the jurisdiction in which the patient or client is located when receiving physical therapist services. Physical therapist services shall be provided consistent with appropriate direction and supervision requirements of assistive personnel and other support personnel, as well as all relevant Association policies, positions, and binding ethical documents. Physical therapy is represented, provided, and promoted only when it meets term and title protections as defined in state law and in accordance with federal health, communications, and trade authorities. There are no funders to report. All authors completed the ICMJE Form for Disclosure of Potential Conflicts of Interest. N. Morelli is an employee of Medtronic PLC.