Prediction of Overall Sleep Score Outcomes in Patients with Post-acute Sequelae of COVID-19 (PASC) Using Smartwatch Data: An Artificial Neural Network (ANN) Approach

Research Objectives: To predict the overall sleep score of patients with PASC using data collected from Garmin smartwatches harnessing an innovative modeling approach.

Design: Retrospective study.

Setting: Outpatient physical rehabilitation hospital.

Participants: Eleven patients with PASC, including fatigue, post-exertional malaise, brain fog, weakness, and limited ability to complete instrumental activities of daily living and work/life roles were provided with Garmin Venu SQ2 smartwatches for therapeutic purposes.

Interventions: Patients completed a post-COVID outpatient program including physical, occupational, and speech therapy (average length=13 weeks, range=7-24). Therapy consisted of graded exercise, diaphragmatic breathing, pacing, sleep, stress and continence management, and cognitive-communication treatment. Smartwatch data was integrated across therapies to guide exercise intensity, stress, and sleep strategies.

Main Outcome Measures: Clinical [patient age, anxiety (GAD-7), depression (PHQ-9), continence] and smartwatch [steps, heart rate zones (50%‒70%) and (70%‒100%)] data were used to train ANN (architecture: Hidden-Layer-Size=80, Training-Ratio=70%, Test-Ratio=15%, and Validation-Ratio=15%) and predict patients’ smartwatch overall sleep quality score (SQS).

Results: Age [Importance-Score (IS)=3.01] was the most important ANN factor predicting SQS with a strong negative correlation. Higher depression scores (IS=2.09) negatively impacted SQS. Anxiety (IS=1.99) showed an overall positive but inverted U-shaped correlation with SQS. Moderate anxiety scores correlated with better SQS, while higher or lower scores did not. Continence (IS=1.83) and time spent within 70%-100% heart rate zones (IS=0.25) demonstrated U-shaped relationships with SQS such that moderate scores in these factors were associated with better SQS. In contrast, time spent within 50%-70% heart rate zones (IS=1.11) negatively impacted SQS. Steps (IS=0.38) exhibited an overall positive but inverted U-shaped relationship with SQS. Moderate physical activity correlated with better sleep, while higher or lower levels did not.

Conclusions: The ANN model demonstrated moderate accuracy in predicting overall SQS, with feature importance analysis highlighting key variables impacting SQS. Results highlight the potential of wearable technology and machine learning to monitor and predict sleep quality in individuals with PASC. Further research with larger datasets and additional features could enhance predictive accuracy and clinical utility. Additionally, research is warranted to explore whether treatments addressing psychological health, continence, and physical activity improve sleep quality in individuals with PASC.

Author(s) Disclosures: This project is being supported, in whole or in part, by federal award number SLFRP1965 (PI: Burnfield), awarded to Douglas County, and by federal award number SLFRP1615 (PI: Burnfield) awarded to the Nebraska Department of Health and Human Services, by the U.S. Department of the Treasury. The co-authors’ work on this abstract was supported, in part, by these grants.