SARS-CoV-2 infection presents a wide spectrum of symptoms, from asymptomatic to severe, hospital-requiring cases, with some patients experiencing long-lasting effects (Long COVID). The prevalence of Long COVID varies widely across studies (5-50%). Healthcare workers (HCWs) were at high risk during the pandemic, with studies reporting high rates of infection and long-term symptoms, primarily fatigue. Smartphones and wearable devices offer opportunities to collect real-world longitudinal data on symptoms and physical activity. Previous research has shown the utility of these technologies in COVID-19 prediction and symptom monitoring. However, the association between physical activity patterns and symptom trajectories in mildly symptomatic COVID-19 patients remains less understood. This study aimed to identify symptom and physical activity trajectory patterns in a cohort of HCWs using unsupervised machine learning and to investigate the association between them.

Existing literature demonstrates a correlation between COVID-19 symptom severity and physical activity levels. Studies using smartphone apps and wearable devices have shown their effectiveness in predicting COVID-19 and distinguishing it from other respiratory illnesses. While research highlights the potential of heart rate variability (HRV) in predicting outcomes in hospitalized patients, less is known about its role in mildly symptomatic individuals. Several studies have focused on the diagnostic use of physical activity parameters from smartwatches; however, the use of wearable technology to monitor and understand the relationship between physical activity and symptoms in individuals with mild COVID-19 symptoms is still in its early stages.

This study recruited 204 HCWs, of whom 121 had confirmed COVID-19 and at least two research clinic visits. 140 participants were enrolled in the MyHeart Counts Study App, with 51 provided Apple Watch Series 4 devices. Data included self-reported symptoms from clinic visits and the app, and various physical activity parameters from the Apple Watch. Latent class growth analysis (LCGA) was used for unsupervised clustering of symptom trajectories, identifying two classes: long and short symptom duration. A similar unsupervised clustering method, using Fretchet-based distance clustering, was applied to eight physical activity features (basalEnergyBurned, activeEnergyBurned, flightsClimbed, distanceWalkingRunning, heartRate, walkingHeartRateAverage, heartRateVariabilitySDNN, stepCount) to identify distinct activity trajectories. Finally, the study investigated the association between COVID-19 symptom trajectories and physical activity trajectory patterns using Fisher's Exact test and a two-sample t-test. Sensitivity analyses were conducted to assess the robustness of the findings by excluding participants with late symptom onset, down-sampling participants, and altering the time intervals for symptom probability calculations.

The study identified two distinct trajectory patterns for COVID-19 symptoms: long and short duration. The prevalence of any long-term symptom was 36%, with fatigue (24.8%) and loss of smell (21.5%) being the most prevalent. Analysis of eight physical activity features also revealed two distinct clusters: high and low activity. Among these, only 'distance moved walking or running' showed a significant association with COVID-19 symptom trajectories. Specifically, individuals with shorter symptom durations exhibited significantly higher levels of this activity. This association was robust even when considering different time windows from COVID-19 onset (starting from one week post-onset). The study found no significant association between other demographic factors (age, gender, ethnicity, comorbidity) and long vs short symptom trajectories, except that having a higher number of symptoms at baseline was associated with longer symptom trajectories. An analysis of re-infections showed that nausea, joint pain, and muscle ache had a higher mean odds of recurrence in the long-trajectory group.

This study's findings support the high prevalence of long-term COVID-19 symptoms in non-hospitalized HCWs and demonstrate a method to identify physical activity patterns and their association with these symptoms. The weak but significant association between higher 'distanceWalkingRunning' and short COVID-19 trajectories suggests a potential link between physical activity and symptom duration, although the reasons for the association with only this specific activity are unclear. This could reflect a difference in data collection frequency, combined effort and distance, or other factors. The study's findings contribute to the growing body of evidence highlighting the importance of long-term symptom monitoring and underscores the potential of wearable technologies in understanding the relationship between physical activity and post-COVID-19 outcomes. Further research with larger sample sizes and more diverse populations is necessary to validate these findings and to explore the underlying mechanisms.

This study demonstrated the utility of unsupervised machine learning techniques in identifying distinct trajectories of COVID-19 symptoms and physical activity in a cohort of HCWs. A significant association was found between higher levels of 'distanceWalkingRunning' and shorter symptom durations, suggesting a potential link between physical activity and symptom recovery. The study highlights the value of wearable technologies in longitudinal health monitoring and calls for further research to explore these associations.

The study's limitations include the relatively small sample size, particularly for the physical activity analysis. The reliance on self-reported symptoms and the potential for recall bias should also be considered. The study was conducted during the early stages of the pandemic, potentially affecting the availability of frequent COVID-19 testing and the ability to accurately determine symptom onset. Finally, significant data loss occurred due to issues with device usage, app setup, data syncing, and data sharing, resulting in a substantial reduction in the number of participants with complete data for all analyses.