Abstract
BACKGROUND: Post-COVID fatigue (pCF) represents a significant burden for many individuals following SARS-CoV-2 infection. The unpredictable nature of fatigue fluctuations impairs daily functioning and quality of life, creating challenges for effective symptom management. OBJECTIVE: This study investigated the feasibility of developing predictive models to forecast next-day fatigue levels in individuals with pCF, utilizing objective physiological and behavioral features derived from wearable device data. METHODS: We analyzed data from 68 participants with pCF who wore an Axivity AX6 device on their non-dominant wrist and a VitalPatch electrocardiogram (ECG) sensor on their chest for up to 21 days while completing fatigue questionnaires every other day. HRV features were extracted from the VitalPatch single-lead ECG signal using the NeuroKit Python package, while activity and sleep features were derived from the Axivity wrist-worn device using the GGIR package. Using a 5-fold cross-validation approach, we trained and evaluated the performances of two machine learning models to predict next-day fatigue levels using Visual Analogue Scale (VAS) fatigue scores: Random Forest and XGBoost. RESULTS: Using five-fold cross-validation, XGBoost outperformed Random Forest in predicting next-day fatigue levels (mean R² = 0.79 ± 0.04 vs. 0.69 ± 0.02; MAE = 3.18 ± 0.63 vs. 6.14 ± 0.96). Predicted and observed fatigue scores were strongly correlated for both models (XGBoost: r = 0.89 ± 0.02; Random Forest: r = 0.86 ± 0.01). Key predictors included heart rate variability features-sample entropy, low-frequency power, and approximate entropy-along with demographic (age, sex) and activity-related (moderate and vigorous duration) factors. These findings underscore the importance of integrating physiological, demographic, and activity data for accurate fatigue prediction. CONCLUSIONS: This study demonstrates the feasibility of combining heart rate variability with activity and sleep features to predict next-day fatigue levels in individuals with pCF. Integrating physiological and behavioral data show promising predictive accuracy and provides insights that could inform future personalized fatigue management strategies.