JMIR Publications

ABSTRACT

Background:

The Functional Independence Measure (FIM) is a standard functional assessment tool in rehabilitation medicine; however, in routine clinical practice, formal assessments are often conducted only at limited intervals, making it difficult to capture patients’ functional recovery trajectories during hospitalization at a high temporal resolution. In contrast, free-text clinical notes routinely documented by physical therapists, occupational therapists, and speech-language pathologists contain rich observational information on patients’ functional status, but this information has not been systematically utilized for quantitative assessment.

Objective:

The aim of this study was to estimate FIM scores at multiple time points from Japanese free-text rehabilitation notes using a large language model (LLM), to evaluate estimation performance by disease and therapy type, and to analyze biases in the distribution of functional information inherent in clinical notes.

Methods:

We retrospectively analyzed free-text notes written by physical therapists, occupational therapists, and speech-language pathologists for patients hospitalized with cerebral infarction, hip fracture, or vertebral fracture at Saiseikai Moriyama Municipal Hospital between 2019 and 2024. Using zero-shot in-context learning, an LLM was employed to estimate scores for all 18 FIM items at each time point. The estimated scores were aligned with ground-truth FIM assessments recorded on the same day and evaluated using mean absolute error (MAE), root mean squared error (RMSE), and weighted kappa. In addition, stratified analyses based on admission FIM scores and therapy-specific FIM item coverage analyses were conducted.

Results:

Overall, the FIM scores estimated by the LLM demonstrated the feasibility of estimating FIM scores from rehabilitation notes. Estimations based on occupational therapy notes showed consistently lower MAE and higher weighted kappa across all disease groups. In contrast, speech-language pathology notes contained limited information on motor items, resulting in relatively larger estimation errors. Patients with lower admission FIM scores tended to exhibit larger estimation errors, suggesting that differences in observational context across therapy records influenced estimation performance.

Conclusions:

Overall, the FIM scores estimated by the LLM demonstrated the feasibility of estimating FIM scores from rehabilitation notes. Estimations based on occupational therapy notes showed consistently lower MAE and higher weighted kappa across all disease groups. In contrast, speech-language pathology notes contained limited information on motor items, resulting in relatively larger estimation errors. Patients with lower admission FIM scores tended to exhibit larger estimation errors, suggesting that differences in observational context across therapy records influenced estimation performance. This study demonstrates that FIM scores can be estimated from free-text rehabilitation notes using the LLM. The proposed approach has the potential to complement sparsely observed FIM assessments without increasing bedside evaluation burden, enabling higher-frequency monitoring of functional recovery trajectories. Time-series functional assessments estimated by LLMs may further support early prediction of discharge outcomes and the development of personalized rehabilitation planning.