MedADL: high-throughput information extraction of functional status from electronic health records to advance frailty assessment in older adults.

BACKGROUND

Functional status is essential for assessing frailty and planning care in older adults but is often under-documented in the structured fields of electronic health records. Manual chart review can capture functional status information but is labor-intensive and time-consuming. In this study, we developed and validated a scalable natural language processing (NLP) model to extract functional status information from unstructured electronic health record (EHR) notes.

METHODS

This retrospective cohort study included 110 older adults seen at a geriatric osteoporosis clinic. A large language model-augmented symbolic NLP pipeline (MedADL) was developed to extract activities of daily living (ADL) and instrumental activities of daily living (IADL) impairments from 13 182 clinical notes and derive the Cumulative Functional Burden Index (CFBI), validated against grip strength, walking speed, balance score, and frailty status using ordinal logistic regression with 10-fold cross-validation.

RESULTS

The NLP model demonstrated moderate to strong performance across 16 ADL and IADL categories, with an average precision of 0.868, sensitivity of 0.828, specificity of 0.937, and F1 score of 0.816. CFBI was significantly associated with frailty status (odds ratio [OR] = 0.799, reciprocal OR = 1.25, p = .049, AUC (Area Under the Curve) = 0.66), indicating that each additional functional deficit was associated with approximately 25% higher odds of being classified in a less robust frailty category. CFBI was correlated with walking speed (r = 0.43, p < .001), balance (r = -0.42, p < .001), grip strength (r = -0.24, p = .019), and fall history (r = 0.21, p = .040).

CONCLUSION

MedADL demonstrated moderate-to-strong performance across 16 bADL and iADL categories, and the derived CFBI was significantly associated with frailty status and correlated with 4 objective physical performance measures. Future studies will prospectively evaluate long-term outcomes and integrate the model into the EHR to enable real-time clinical decision support.