Explainable AI for Tracking and Predicting Human Behavioral Changes Across the Lifespan

Abstract

Background: Everyday behaviours such as physical activity, sedentary time and sleep change systematically across the human lifespan, yet most behavioural prediction models are trained on pooled populations and explained, if at all, with a single global attribution profile.
Problem: Pooled models can obscure age-dependent behavioural mechanisms, and post-hoc explanations are rarely tested for stability.
Objective: To develop and evaluate a Lifespan-Aware Explainable Behavioral Prediction Framework (LA-EBPF) that predicts short-term behavioural change and explains its predictions separately for lifespan groups.
Framework: LA-EBPF integrates behavioural-change feature engineering (day-to-day differences, trends, variability, transitions), five literature-informed lifespan strata, partially pooled shared-plus-group gradient-boosted models, lifespan-conditional Shapley attribution, a Lifespan Attribution Divergence (LAD) index and an Explanation Stability Index (ESI).
Data and simulation: The variable schema, wear protocol and age structure follow the public NHANES 2011–2014 wrist-accelerometer Physical Activity Monitor files; the numerical evaluation uses a controlled, calibrated synthetic cohort (N = 11,000) with known, age-varying behavioural drivers. The task is to predict a person-relative downturn in day-7 activity from six preceding days.
Simulation findings: Across ten repeated participant-level splits, LA-EBPF reached ROC-AUC 0.770 ± 0.013, F1 0.525 ± 0.017 and MCC 0.337 ± 0.023, versus 0.732 ± 0.013, 0.495 ± 0.020 and 0.298 ± 0.027 for the strongest baseline (gradient-boosted trees on raw day-level features). Removing behavioural-change features produced the largest loss (AUC 0.725). Group attribution profiles diverged well beyond a permutation null (mean LAD 0.171 vs 0.0018; p = 0.005), and 13 of 15 simulated group-specific drivers were recovered among the top-eight attributions. Stability-aware screening reduced the feature set by about 28% without loss of accuracy but did not raise rank-level ESI. Novelty and significance: LA-EBPF contributes an integrated, auditable procedure for lifespan-stratified behavioural prediction with explanation-stability reporting. All numerical results are simulation-based and require empirical replication on real longitudinal data.