(M1130-05-32) Fatigue Forecasting in Long COVID Using Wearable Heart Rate Variability Monitoring

Purpose: Fatigue is the most common post-acute sequelae of COVID-19 (PASC; also known as Long COVID), affecting millions globally with no approved pharmacological interventions. Current clinical trials for fatigue therapeutics rely on subjective patient-reported outcomes (PROs) that suffer from recall bias, placebo effects, and poor sensitivity to treatment response. This study aims to develop and validate objective digital biomarkers for fatigue using heart rate variability (HRV) metrics captured continuously via wearable sensors, establishing quantifiable clinical trial endpoints for therapeutic development. Fatigue manifests as autonomic nervous system dysfunction, with sympathetic overdrive and parasympathetic withdrawal [3-5]. By establishing real-time, objective fatigue signatures in PASC patients, we create a framework for: (1) stratifying patients for clinical trial enrollment based on fatigue phenotypes, (2) monitoring therapeutic response with greater sensitivity than PROs, and (3) enabling adaptive trial designs with continuous endpoint assessment. This digital biomarker approach could accelerate drug development timelines by providing early efficacy signals and dose-response relationships. The methodology has implications for developing therapeutics across multiple fatigue-associated conditions including multiple sclerosis (MS) and myalgic encephalomyelitis/chronic fatigue syndrome (ME/CFS), potentially transforming how we evaluate fatigue reduction agents and develop personalized treatment strategies.

Methods: We collected physiological data from 27 PASC patients with persistent fatigue using Empatica E4 wristbands [1] measuring heart rate, electrodermal activity, movement, and skin temperature. Ten participants completed mentally fatiguing tasks consisting of six blocks while wearing the device. Participants reported fatigue levels (1-100 visual analogue scale) [2] before and after each block. Raw photoplethysmography signals underwent preprocessing including low-pass denoising filters and motion artifact removal using Padasip adaptive filtering [6] with accelerometer reference signals. Data were segmented into 10, 30, and 60-second windows to increase training data points. R-R intervals were extracted using peak detection algorithms, with ectopic beats filtered out to obtain normal-to-normal intervals [7]. HRV features were extracted using NeuroKit2 Python package [8]. Feature importance was determined through wrapper and filter-based methods, with permutation importance analysis on validation sets. A Recursive Neural Network with Long Short-Term Memory nodes (RNN-LSTM) was employed for fatigue prediction using leave-one-subject-out cross-validation (Figure 1). Additionally, participants wore the E4 wristband continuously in real-world settings while reporting fatigue levels six times daily to validate laboratory findings. The metrics that are evaluated are mean squared error (MSE) for accuracy and R² value for predictability calculations in comparison to randomness in the data. Instead of absolute fatigue ratings, the change from baseline is calculated to account for individual biases of self-reported fatigue.

Results: The preprocessing pipeline successfully extracted clean HRV signals from wearable sensor data across all participants (Figure 2). Feature selection identified three primary HRV metrics for fatigue severity prediction: inverse of the average length of the acceleration/deceleration segments (HRV_IALS), percentage of inflection points of the RR intervals series (HRV_PIP), and the spectral power of high frequencies (HRV_HF) . The RNN-LSTM model achieved normalized test mean squared error (MSE) of 0.0304 and R² of 0.4874, demonstrating moderate predictive capability for self-reported fatigue levels. Leave-one-subject-out cross-validation showed consistent performance across participants, with the model successfully predicting changes from baseline fatigue levels for individual participants using sequential time-series data. The model captured temporal trends in fatigue progression throughout the mentally demanding tasks, with predictions closely tracking self-reported values. Real-world continuous monitoring data collection was successfully completed, with participants maintaining compliance for six daily fatigue reports while wearing the device. The identified HRV features showed consistent patterns between laboratory and real-world settings, supporting the ecological validity of our approach.

Conclusion: This study demonstrates the feasibility of using wearable sensor-derived HRV metrics to objectively detect and predict fatigue levels in Long COVID patients experiencing chronic fatigue. The identification of specific HRV features (HRV_IALS, HRV_PIP, HRV_HF) as reliable predictors of fatigue severity provides a physiological basis for real-time fatigue monitoring with lightweight cloud orchestrated prediction pipeline. With R² of 0.4874, our machine learning approach shows promise for developing personalized fatigue management systems that could deliver timely interventions. The successful implementation in both controlled laboratory and real-world settings validates the practical applicability of this technology. This methodology establishes a framework for objective fatigue assessment that could revolutionize management strategies for PASC patients and potentially extend to other conditions with chronic fatigue, including MS and ME/CFS. Future work will refine predictive models, perform patient stratification into distinct autonomic phenotypes for precision medicine approaches, potentially matching patients to mechanism-based therapies (sympatholytics, parasympathomimetics, or combination treatments), and develop real-time intervention protocols based on detected fatigue patterns.

References: 1. McCarthy C, Pradhan N, Redpath C, Adler A. Validation of the Empatica E4 wristband. 2016 IEEE EMBS international student conference (ISC); 2016: IEEE; 2016. p. 1-4.
2. Carlsson AM. Assessment of chronic pain. I. Aspects of the reliability and validity of the visual analogue scale. Pain 1983; 16(1): 87-101.
3. Lee KFA, Gan W-S, Christopoulos G. Biomarker-Informed Machine Learning Model of Cognitive Fatigue from a Heart Rate Response Perspective. Sensors 2021; 21(11): 3843.
4. Escorihuela RM, Capdevila L, Castro JR, et al. Reduced heart rate variability predicts fatigue severity in individuals with chronic fatigue syndrome/myalgic encephalomyelitis. Journal of translational medicine 2020; 18(1): 1-12.
5. Fagundes CP, Murray DM, Hwang BS, et al. Sympathetic and parasympathetic activity in cancer-related fatigue: more evidence for a physiological substrate in cancer survivors. Psychoneuroendocrinology 2011; 36(8): 1137-47.
6. Cejnek M, Vrba J. Padasip: an open-source Python toolbox for adaptive filtering. Journal of Computational Science 2022; 65: 101887.
7. Campanella S, Altaleb A, Belli A, Pierleoni P, Palma L. A method for stress detection using empatica E4 bracelet and machine-learning techniques. Sensors 2023; 23(7): 3565.
8. Makowski D, Pham T, Lau ZJ, et al. NeuroKit2: A Python toolbox for neurophysiological signal processing. Behavior research methods 2021: 1-8.

Acknowledgements: This study was funded by the Rutgers University Research Council Collaborative Multidisciplinary Award. No conflicts of interest. Due process with HIPAA compliance and IRB approved protocols maintained for data collection and analysis.

Figure 1. Data Processing steps, from raw signal from Empatica E4 wristband to preprocessed signal for RNN-LSTM

Figure2: Preliminary results of RNN-LSTM predicted fatigue levels vs self-reported target fatigue levels for each test subject. The length of sequences varies due truncation of blocks with irregular R-R intervals indicating corrupted data.