An IoT-Ready Context-Aware Patient State Framework with LLM-Driven Recommendation for Shoulder Rehabilitation

초록

In Internet-of-Things (IoT)-based rehabilitation, patient data from wearable sensors (IMU, EMG, heart rate), clinical assessments, and surveys differ in format and granularity, complicating unified patient-state construction. Existing AI-based systems often omit fatigue level or rehabilitation stage, or they place large language models (LLMs) in the clinical decision-making role, exposing patients to hallucination risk. We propose an IoT-ready framework comprising three components: an input-source-independent context abstraction pipeline, a digital twin that simulates clinical-score trajectories, and an LLM Agent that interprets the outputs of a deterministic algorithm and a digital twin as (subject, predicate, object) Triplets to produce a retrieval-augmented clinical report. The algorithm-not the LLM-selects the 13-exercise sequence from Shoulder Pain and Disability Index (SPADI) item-level responses. Explicit per-layer schemas let IoT-sensor branches be added without changing the downstream interface. Here we evaluate the framework on the clinical-score path (three patient-reported outcome measures and six range-of-motion measures); the multi-modal IoT branches maintain deployment-target functionality. On 48 IRB-approved shoulder rehabilitation patients (144 longitudinal records; augmented to 7200 only to train the trajectory generator), real-only leave-one-subject-out evaluation gave VAS RMSE 0.658 (0-10) and SPADI RMSE 6.499 (0-100). Ablation across four surface forms revealed a fidelity-accuracy trade-off-Narrative highest on template fidelity (BERTScore F1 0.250), raw JSON highest on judge-rated accuracy-and the framework adopts the balanced-midpoint Triplet form. A claim-level audit of the Triplet-form reports left 22-31% of atomic claims unsupported (Claude Sonnet and GPT-4o judges) regardless of guideline retrieval. A raw-LLM control never reproduced the algorithm-defined sequence exactly (0 of 48), supporting deterministic, auditable sequence selection and the need for clinician review before clinical use.

키워드