대형 언어 모델을 활용한 맞춤형 디지털 물리·작업 치료 처방 시스템

📝 Abstract

Digital health interventions increasingly deliver home exercise programs via sensor-equipped devices such as smartphones, enabling remote monitoring of adherence and performance. However, current software is usually authored before clinical encounters as libraries of modules for broad impairment categories. At the point of care, clinicians can only choose from these modules and adjust a few parameters (for example, duration or repetitions). As a result, individual limitations, goals, and environmental constraints are often not reflected, limiting personalization and benefit. We propose a paradigm in which large language models (LLMs) act as constrained translators that convert clinicians’ exercise prescriptions into intervention software. Clinicians remain the decision makers: they design exercises during the encounter, tailored to each patient’s impairments, goals, and environment, and the LLM generates matching software. We conducted a prospective single-arm feasibility study with 20 licensed physical and occupational therapists who created 40 individualized upper extremity programs for a standardized patient; 100% of prescriptions were translated into executable software, compared with 55% under a representative template-based digital health intervention (p < 0.01). LLM-generated software correctly delivered 99.7% of instructions and monitored performance with 88.4% accuracy (95% confidence interval, 0.843-0.915). Overall, 90% of therapists judged the system safe for patient interaction and 75% expressed willingness to adopt it in practice. To our knowledge, this is the first prospective evaluation of clinician-directed intervention software generation with an LLM in health care, demonstrating feasibility and motivating larger trials in real patient populations.

💡 Analysis

Digital health interventions increasingly deliver home exercise programs via sensor-equipped devices such as smartphones, enabling remote monitoring of adherence and performance. However, current software is usually authored before clinical encounters as libraries of modules for broad impairment categories. At the point of care, clinicians can only choose from these modules and adjust a few parameters (for example, duration or repetitions). As a result, individual limitations, goals, and environmental constraints are often not reflected, limiting personalization and benefit. We propose a paradigm in which large language models (LLMs) act as constrained translators that convert clinicians’ exercise prescriptions into intervention software. Clinicians remain the decision makers: they design exercises during the encounter, tailored to each patient’s impairments, goals, and environment, and the LLM generates matching software. We conducted a prospective single-arm feasibility study with 20 licensed physical and occupational therapists who created 40 individualized upper extremity programs for a standardized patient; 100% of prescriptions were translated into executable software, compared with 55% under a representative template-based digital health intervention (p < 0.01). LLM-generated software correctly delivered 99.7% of instructions and monitored performance with 88.4% accuracy (95% confidence interval, 0.843-0.915). Overall, 90% of therapists judged the system safe for patient interaction and 75% expressed willingness to adopt it in practice. To our knowledge, this is the first prospective evaluation of clinician-directed intervention software generation with an LLM in health care, demonstrating feasibility and motivating larger trials in real patient populations.

📄 Content

Clinician-Directed Large Language Model Software Generation for Therapeutic Interventions in Physical Rehabilitation Edward Kim1*†, Yuri Cho1†, Jos´e Eduardo E. Lima2, Julie Muccini2, Jenelle Jindal2, Alison Scheid3, Erik Nelson1, Seong Hyun Park1, Yuchen Zeng1, Alton Sturgis1, Caesar Li1, Jackie Dai1, Sun Min Kim1, Yash Prakash1, Liwen Sun1, Isabella Hu1, Hongxuan Wu1, Daniel He1, Wiktor Rajca1, Cathra Halabi3, Maarten Lansberg2, Bjoern Hartmann1, Sanjit A. Seshia1 1*Electrical Engineering and Computer Sciences, University of California, Berkeley, CA, USA. 2Neurology, Stanford University, CA, USA. 3Neurology, University of California, San Francisco, CA, USA. *Corresponding author(s). E-mail(s): ek65@eecs.berkeley.edu; †These authors contributed equally to this work. Abstract Digital health interventions are increasingly used in physical and occupational therapy to deliver home exercise programs via sensor-equipped devices such as smartphones, enabling remote monitoring of adherence and performance. In the current digital intervention paradigm, however, exercise software is typically pro- grammed before clinical encounters as libraries of pre-defined modules aimed at broad impairment categories. At the point of care, clinicians can only select from this library and adjust a narrow set of parameters such as duration and repetitions. Patient-specific needs that emerge during encounters—such as dis- tinct movement limitations, personal goals, or home and work constraints—are rarely reflected in the software, limiting personalization and contributing to lower adherence and reduced therapeutic benefit. In this study, we propose a digital intervention paradigm in which large language models (LLMs) act as constrained translators that convert clinicians’ exercise prescriptions into inter- vention software. Clinicians preserve their role as clinical decision-makers: they 1 arXiv:2511.18274v2 [cs.HC] 6 Dec 2025 design exercises during the encounter, tailored to each patient’s impairments, goals, and environment, and the LLM generates software to match these prescrip- tions. We conducted a prospective single-arm feasibility study with 20 licensed physical and occupational therapists and a standardized patient. Clinicians cre- ated 40 individualized upper extremity exercise programs (398 total instructions), which were automatically translated into executable software. A standardized patient was used because the safety of LLM-generated intervention software had not previously been evaluated. We benchmarked each therapist-designed pre- scription against a representative template-based DHI to estimate how many could be delivered under the status quo paradigm. The proposed paradigm yielded a 45% increase in the proportion of personalized prescriptions that could be implemented as software compared with the status quo (100% vs 55%; p <0.01), with unanimous agreement among therapists that it was easy to use. The LLM-generated software correctly delivered 99.7% (397/398) of instructions as prescribed and monitored performance with 88.4% (95% CI, 0.843–0.915) accuracy. Overall, 90% (18/20) of therapists judged the system safe for patient interaction, and 75% expressed willingness to adopt it in their clinical practice. To our knowledge, this is the first prospective evaluation of clinician-directed intervention software generation with an LLM in healthcare, demonstrating fea- sibility and motivating larger trials to assess clinical effectiveness and safety in real-patient populations. Keywords: large language model, LLM software generation, digital health intervention, physical rehabilitation 1 Introduction Digital health interventions (DHIs) [1, 2] in physical and occupational therapy increas- ingly take the form of clinician-prescribed intervention software deployed to a patient’s smartphone or connected wearable device [3, 4]. Such software delivers step-by-step exercise instructions and can objectively capture adherence and performance using on-device sensors (e.g., camera-based motion tracking on phones or inertial sensing in wearables). As a result, it can quantify whether exercises were completed and how they were performed—such as range of motion, repetition count, tempo, and posture—and return concise summaries to clinicians for timely adjustment of therapy (e.g., mod- ifying difficulty or introducing new exercises). This capability represents a practical advance over standard paper exercise worksheets, which cannot capture how patients perform prescribed exercises at home between visits. Despite these gains, the prevailing digital prescription paradigm with monitoring remains rigid in its capacity to personalize therapy. In current DHI platforms, prior to clinical encounters, therapeutic exercises are encoded as a library of parametrized intervention software, targeting broad populations with common impairments. Con- sequently, patient-specific needs identifiable only during encounters (for example, distinct defic

This content is AI-processed based on ArXiv data.