MetaHARP: Meta-Learning-Driven Rapid Adaptation for Individual-Heterogeneity-Aware Behavioral Risk Prediction

Behavioral risk prediction increasingly relies on longitudinal signals from wearable devices, smartphones, online activities, and multimodal health records. However, individual behavioral patterns and risk-triggering mechanisms are highly heterogeneous, making population-level predictors difficult to adapt to new individuals with limited observations. This paper proposes MetaHARP, a meta-learning-driven rapid adaptation framework for individual-heterogeneity-aware behavioral risk prediction. MetaHARP treats each individual as a meta-task, encodes support observations into a heterogeneity prototype, retrieves adaptation priors from behaviorally similar individuals, and generates lightweight low-rank adapters for personalized risk prediction. An individual-conditioned calibration module is further introduced to improve the reliability of predicted risk probabilities. Experiments on four public behavioral and health-related datasets show that MetaHARP consistently outperforms recent time-series and foundation-model baselines in predictive performance, few-shot adaptability, probability calibration, and adaptation efficiency. These results demonstrate that heterogeneity-aware meta-adaptation provides an effective solution for personalized behavioral risk prediction.