Classification of Individuals With COVID-19 and Post-COVID-19 Condition and Healthy Controls Using Heart Rate Variability: Machine Learning Study With a Near-Real-Time Monitoring Component.

BACKGROUND: Heart rate variability (HRV) is a validated biomarker of autonomic and inflammatory regulation and has been associated with both acute COVID-19 and post-COVID-19 condition. Although reverse transcription polymerase chain reaction remains the diagnostic gold standard for acute infection, there is a lack of accessible, noninvasive physiological tools to support ongoing monitoring and stage differentiation of COVID-19 and its sequelae. The growing availability of wearable devices capable of real-time HRV data collection opens up opportunities for early detection and health status classification using machine learning.

OBJECTIVE: This study aimed to identify HRV patterns capable of distinguishing individuals with active COVID-19 and post-COVID-19 condition and healthy controls using data collected from wearable devices and processed using machine learning models. A secondary objective was to assess the feasibility of a near-real-time health monitoring system based on these patterns using wearable-derived HRV data.

METHODS: HRV indexes (SD of the normal-to-normal intervals [SDNN], root mean square of successive differences [RMSSD], low-frequency relative power [LF%], and high-frequency relative power [HF%]) were collected from 61 participants (n=21, 34% with active COVID-19; n=20, 33% with post-COVID-19 condition; and n=20, 33% healthy controls) using 2 standardized datasets. Classification models were developed using supervised machine learning algorithms (decision tree, support vector machines, k-nearest neighbor, and neural networks) and evaluated through cross-validation. A contextual clinical variable indicating recent SARS-CoV-2 infection was incorporated into 1 model configuration to assess its impact on classification performance. In addition, a prototype system for near-real-time monitoring was implemented and tested in a separate group of 4 participants.

RESULTS: Participants with active COVID-19 showed significantly lower HRV indexes (SDNN, RMSSD, LF%, and HF%) compared to both participants with post-COVID-19 condition and healthy controls (P<.001), whereas differences between the post-COVID-19 condition and healthy groups were not statistically significant. Decision tree models trained solely on HRV features achieved 76.4% accuracy with high discriminative performance for active COVID-19 (F1-score=88%; area under the curve=0.85) but limited detection of post-COVID-19 condition (F1-score=56%). When a contextual clinical variable indicating recent SARS-CoV-2 infection was included, overall accuracy increased to 87%, and the F1-score for post-COVID-19 condition rose to 92%, with improved area under the curve metrics across classes. A prototype system tested on 4 independent participants correctly classified their status, demonstrating feasibility for near-real-time application.

CONCLUSIONS: HRV patterns collected from wearable devices and analyzed via machine learning successfully distinguished individuals with active COVID-19 from healthy individuals with high accuracy using physiological data alone. When a clinical contextual variable indicating recent infection was added, the model also achieved strong performance in identifying post-COVID-19 condition cases. A prototype system demonstrated feasibility for near-real-time application, reinforcing the potential of HRV for individualized health monitoring.