Large Language Model–Driven Feature Weighting for Cross-Device Harmonization of Wearable Data: A Generalizable Framework for Digital Health Models

초록

Wearable-based lifelog data offer a promising approach for digital mental health monitoring, but cohort heterogeneity, including device-specific measurement differences and variability in ecological momentary assessment (EMA) protocols, often limits model generalizability. To address this issue, we propose a large language model (LLM)-driven feature-weighting framework for cross-device harmonization in mood prediction. Leveraging few-shot learning, the framework infers cohort-specific rules from a small number of representative samples and converts them into executable weighting functions without requiring large-scale retraining. Four cohorts with distinct characteristics were used to generate cohort-specific weights, and an independent external cohort was used for validation. Model performance was evaluated by comparing classifiers before and after applying the generated weights, and traditional machine learning-based weighting methods were included as baselines to assess the validity of the LLM-based approach. Performance was measured using the area under the receiver operating characteristic curve (AUROC) and F1-score across five random seeds. LLM-based weighting improved external validation F1-score more consistently than machine learning-based weighting, particularly when GPT-OSS was used. SHapley Additive exPlanations (SHAP) and correlation analyses indicated that the generated weights were actively used by the predictive models and were associated with relevant physiological and behavioral features. These findings suggest that LLM-driven rule-based feature weighting may improve the external generalizability of wearable-based digital health models across heterogeneous cohorts. © 1975-2011 IEEE.

키워드