Benchmarking information extraction of physical activity from electronic health record with large language models: an...

We aimed to develop a data model and a natural language processing (NLP) pipeline for representing physical activity (PA) in Electronic Health Records (EHRs), and to evaluate transformer- and Large Language Model (LLM)-based classifiers for sentence-level PA attribute classification.

We analyzed PA documentation across three patient cohorts (cancer, COVID, and Alzheimer's disease) using structured and unstructured EHR data. A conceptual schema was developed to represent PA and its linguistic attributes. Five BERT models and three modern LLMs (Llama3-8B, MedAlpaca-13B, and PMC-Llama-13B) were evaluated for classifying PA attributes (binary status, negation, exclusion, and an eleven-class Category) on pre-extracted PA-related sentences.

Clinical notes were a richer source of PA information than structured ICD or SDoH data. On binary tasks, the best BERT model reached F1 0.619 (Exclusion); with Supervised Fine-Tuning (SFT), Llama3-8B reached F1 0.689 (Exclusion). On the 11-class Category task, performance was modest (best macro-F1 0.262, ROC-AUC 0.803, by Llama3-8B).

In-Context Learning (ICL) was highly variable: while Llama3-8B-ICL achieved the best ROC-AUC on Category, the domain-specific MedAlpaca-13B and PMC-LLaMA-13B essentially failed. These results, together with sparsely represented PA elements (Amount, Frequency, Assessment) and the absence of a downstream evaluation, position this work as an initial proof of feasibility, with supervised domain adaptation still required for reliable clinical PA extraction.

We contribute a PA data model, annotation schema, and a working NLP pipeline with a BERT/LLM benchmark for sentence-level PA attribute classification. The pipeline supports future end-to-end PA extraction and downstream applications such as phenotyping, risk prediction, and cohort identification.