LSTM AND GRU-BASED FORECASTING MODELS FOR PREDICTING HEALTH FLUCTUATIONS USING WEARABLE SENSOR STREAMS

Abstract

This quantitative, cross-sectional, case-study–based research addresses the problem that short-horizon health fluctuations in wearable streams are difficult to operationalize and forecast, limiting actionable monitoring. The purpose was to define fluctuation outcomes, identify wearable indicators linked to perceived variability, and compare LSTM versus GRU forecasting pipelines under matched preprocessing and evaluation. The bounded case sample comprised 120 monitored participant cases retained after quality screening (mean wear-time 12.8 hours/day, SD 2.1; overall missingness 9.6%, SD 6.4; 84.2% high adherence at or above 10 hours/day). The primary subjective variable, Health Fluctuation Severity, was a 6-item 5-point Likert composite with acceptable reliability (Cronbach’s alpha .86; mean 3.18, SD 0.74), and the primary objective variable was an Objective Fluctuation Index (mean 0.61, SD 0.19). The analysis plan combined descriptive statistics, reliability checks, Pearson correlations, and regression with data-quality controls, alongside forecasting evaluation using MAE and RMSE and a paired t-test on participant-level errors. Headline findings showed convergence between subjective and objective measures (r .52, p < .001) and significant associations with resting heart-rate shift (r .41), sleep disruption (r .46), activity instability (r .29), and HRV proxy (r −.38) (all p ≤ .001). GRU outperformed LSTM on the test set (MAE 0.072 vs 0.081; RMSE 0.094 vs 0.106; t(119) 4.62, p < .001) and increased explanatory power when combined with wearable features (R squared 0.41; delta R squared 0.07, p = .004). The results imply that baseline-centered indices plus GRU forecasting can support validated fluctuation monitoring in real-world programs when adherence and missingness are managed.