Uncertainty-aware federated temporal learning with explainable LLM-based coaching for privacy-preserving wearable health systems

Abstract

Introduction The growing use of wearable sensors enables continuous health and activity monitoring; however, challenges such as noisy data, privacy concerns, and device heterogeneity limit the effectiveness of centralized learning systems. This paper presents a privacy-first, AI-enhanced lifestyle coaching framework that integrates multi-stage signal cleaning, federated learning (FL), and explainable large language models (LLMs) with structured prompting for intelligent, decentralized health guidance. Unlike prior FL-based human activity recognition (HAR) systems, this work introduces (i) a quantitatively validated multi-stage denoising pipeline, (ii) local latent-space semantic imputation under federated constraints, and (iii) an explainable LLM-based coaching layer whose reasoning is explicitly grounded in aggregated federated activity representations rather than raw sensor data. Methods The proposed system employs a robust preprocessing pipeline comprising Hampel filtering, adaptive wavelet–Butterworth denoising, Variational Autoencoder (VAE)-based semantic imputation, and Kalman smoothing. The cleaned signals are used to train highly efficient Deep LSTM models on-device, and local updates are aggregated using a FedProx-based framework to mitigate client drift and preserve data privacy. To ensure rigorous evaluation and eliminate temporal data leakage, strict chronological data splitting is enforced alongside a 50% overlap stride constraint. Results The multi-stage signal cleaning pipeline successfully achieved a 7–10 dB improvement in signal-to-noise ratio (SNR). Under the strict evaluation conditions, the global federated model established a highly realistic, leak-free baseline, achieving an average global accuracy of 68.08% and a Macro-F1 score of 0.60 on highly imbalanced, strictly unseen future time-series data. Furthermore, the LLM-based coaching module achieved a faithfulness score of 0.87. Discussion The high faithfulness of the coaching module proves that high-level, uncertainty-aware summaries are sufficient to generate personalized, transparent, and context-aware lifestyle recommendations without ever exposing raw sensor streams. Overall, the integration of uncertainty-aware signal preprocessing, federated temporal modeling, and explainable LLM-driven coaching establishes a mathematically sound, scalable, and secure architecture for next-generation AI-driven health systems.