Unsupervised machine learning to investigate trajectory patterns of COVID-19 symptoms and physical activity measured via the MyHeart Counts App and smart devices

The study addresses how COVID-19 symptom trajectories (long versus short duration) manifest in non-hospitalized healthcare workers and whether these trajectories are associated with real-world physical activity patterns measured passively by smartphones and wearables. Contextually, Long COVID is prevalent but variably estimated, and frontline HCWs had high infection rates. Prior work has leveraged smartphones and wearables to detect COVID-19 and monitor physiology (e.g., HRV), but less is known about how physical activity relates to symptom persistence in mildly symptomatic or non-hospitalized cases. The authors hypothesized that unsupervised trajectory modeling would reveal distinct long/short symptom patterns and that physical activity would also form distinct trajectory clusters, potentially associated with symptom persistence.

Background literature highlights: (1) Long COVID presents persistent symptoms (fatigue, dyspnea, neuropsychological issues) with prevalence estimates ranging 5–50%; ONS reported 13.7% with symptoms beyond 12 weeks. (2) Hospitalized cohorts show persistent symptoms up to 12 months, with fatigue and dyspnea most frequent. (3) HCWs had high seroprevalence early in the pandemic. (4) Wearables and smartphones have been used to detect COVID-19 onset, distinguish COVID-19 from influenza, and monitor physiology (e.g., HRV) and activity; combining symptoms with wearable data improves detection. (5) Evidence that physical inactivity is associated with worse COVID-19 outcomes and interest in leveraging wearable-derived metrics for cardio-respiratory diseases is growing. Gaps remain regarding the utility of HRV and activity metrics in mildly symptomatic/non-hospitalized individuals and their relation to symptom duration.

Design and participants: 204 HCWs were recruited (July 2020–July 2021) into the Sheffield Teaching Hospitals NHS Foundation Trust observational study (STH-ObS, 18/YH/0441). Inclusion required age >18 and iPhone 6 or later; Apple Watch Series 4 was offered. No participants were hospitalized for COVID-19. COVID-19 status was based on PCR and/or serology (spike and nucleocapsid). Demographics, symptoms, testing, and vaccination were recorded at research clinics. Ethical conduct adhered to the Declaration of Helsinki and ICH-GCP. Data sources: (a) Clinician-recorded symptoms at up to five clinic visits across >1 year; 17 symptoms assessed (fever, lost taste, lost smell, fatigue, confusion, nausea/vomit, headache, wheeze, sore throat, abdominal pain, joint pain, runny nose, muscle ache, cough, shortness of breath, diarrhoea, chest pain). (b) MyHeart Counts app self-reports (10 overlapping symptoms; loss of appetite excluded as not in clinic list). (c) Apple HealthKit activity metrics from iPhone/Apple Watch: basalEnergyBurned, activeEnergyBurned, flightsClimbed, distanceWalkingRunning, heartRate, walkingHeartRateAverage, heartRateVariabilitySDNN, stepCount. Cohorts for analysis: 121 participants with symptomatic COVID-19 and at least two symptom timepoints (clinical or app) and a known onset date (116 symptom onset; 5 PCR date) were used for symptom trajectory analysis. Activity data sufficient for longitudinal analysis were available for 34 participants; 21 had both symptom trajectories and all 8 activity measures. Symptom trajectory modeling: For each participant, a generalized linear model (probit link; Matlab 2017a glmfit) estimated the probability of symptom presence over time using the ratio P = (number of positive symptoms)/(number assessed), with Nq=17 for clinic visits and Nq=10 for app self-reports. To align trajectories for unsupervised clustering, estimated probabilities were computed at five common timepoints across one year (0, 90, 180, 270, 360 days). Latent Class Growth Analysis (LCGA; Mplus) identified trajectory classes using BLRT p<0.05, high entropy and average posterior probabilities, assessment of linear/quadratic growth, and convergence checks. The same approach (with Nq=1) estimated probabilities for each specific symptom. Physical activity trajectory clustering: Time series for each of the 8 activity measures (mean ~487 timepoints; range 42–596) were analyzed using a shape-based approach (R 4.2.0; kmlShape v0.9.5). Frechét distance was used to merge trajectories into 30–33 “senator” trajectories per activity; these were resampled to 100 timepoints for shape similarity. K-means-based clustering (k=2–5 tested) yielded stable two-cluster solutions (high vs low activity). Differences between mean representative trajectories (senators) were assessed via Welch two-sample t-tests. Association analyses: Associations between symptom trajectory class (long vs short) and activity cluster (high vs low) were tested using Fisher’s Exact tests (contingency tables). Univariable two-sample t-tests assessed differences in mean activity levels between symptom groups for windows after onset (3 days, 1 week, 2 weeks, 1 month, 3 months). Models adjusted for age, gender, and comorbidities where specified. Sensitivity analyses included excluding certain baseline periods, down-sampling participants to test clustering robustness, varying time intervals for symptom probability estimation, and pre/post COVID-19 comparisons (paired Wilcoxon) for available activity data. Reinfection analysis: Among 38 reinfected participants (24 short, 14 long), odds of recurrence of 17 specific symptoms were compared between long vs short trajectory groups.

- Symptom trajectories: LCGA identified two classes (short vs long) for any symptom across 121 participants (77 short, 44 long). Prevalence of long trajectories for any symptom: 36.36% [95% CI: 26.42, 48.82]. Two-class solutions (posterior probability >0.95) were supported for nine specific symptoms with uneven distributions, including: lost taste (109/12 short/long), lost smell (95/26), fatigue (91/30), headache (109/12), joint pain (112/9), muscle ache (102/19), cough (101/20), shortness of breath (100/21), chest pain (97/24). The most prevalent long-term individual symptoms were fatigue 24.79% [16.73, 35.39] and loss of smell ~21.5%. - Baseline symptom burden predicted persistence: Higher number of symptoms at baseline had higher odds of long trajectories across several symptoms (significant in multiple models). - Reinfection: In 38 reinfected participants, recurrence of specific symptoms had higher odds among those with long trajectories for any symptom: nausea/vomit OR 21.0 [1.035, 426.0], joint pain OR 12.8 [1.31, 125.1], muscle ache OR 6.071 [1.412, 26.03]. - Physical activity trajectories: For 34 participants with wearables, two trajectory clusters (high vs low) were identified for each activity. Significant differences between high vs low clusters were observed for most activities (p ≤ 2.858e-07), with walkingHeartRateAverage not differing (p=0.3593). High-activity participants had roughly double distanceWalkingRunning (10,955 m vs 4,964.5 m) and higher activeEnergyBurned (949,650 vs 538,190 calories). FlightsClimbed differences were largest, especially in the first 30 days. - Symptom–activity association (n=21 with both): Only distanceWalkingRunning trajectory clusters were significantly associated with symptom trajectories (OR=0.12; Fisher’s p=0.03), indicating high distanceWalkingRunning aligned with short symptom trajectories. Individuals in the long-symptom group had lower mean distanceWalkingRunning, with significant differences emerging from 1 week post-onset and persisting at later windows. StepCount and flightsClimbed were not significantly associated with symptom trajectory groups.

Unsupervised modeling distinguished robust long vs short COVID-19 symptom trajectories in non-hospitalized HCWs and identified parallel high vs low physical activity trajectories from wearable data. The key observation is a specific association between higher distanceWalkingRunning and shorter symptom duration, suggesting that greater real-world ambulatory activity may correlate with or reflect more rapid symptom resolution. Other activity metrics (e.g., stepCount, flightsClimbed) did not show consistent associations, potentially due to measurement context (e.g., distanceWalkingRunning recorded as workouts with higher sampling frequency) or the nature of the metric (distance and effort combined). The prevalence and pattern of long-term symptoms align with prior reports, particularly for fatigue and anosmia. Reinfection findings suggest individuals with long trajectories may be more prone to recurrence of certain symptoms. These results support the feasibility and value of integrating consumer wearable data with clinical and app-based symptom tracking to study recovery trajectories and potentially guide monitoring of post-acute sequelae.

This study demonstrates a methodology to identify longitudinal patterns of COVID-19 symptoms and physical activity using unsupervised learning, showing a high prevalence of long-term symptoms (36%) in non-hospitalized HCWs and a specific association between higher distanceWalkingRunning and shorter symptom duration. The approach highlights the potential of routinely collected wearable data to complement clinical follow-up for COVID-19 and other cardio-respiratory diseases. Future work should include larger, more diverse cohorts, improved continuous data capture (including during working hours), validation of associations across devices and platforms, exploration of causal mechanisms, and integration of additional physiological markers (e.g., HRV, respiratory rate) to refine predictive models of recovery.

- Limited testing availability early in the pandemic led to reliance on retrospective symptom recall for some participants, introducing potential recall bias. - Possible unmeasured co-infections cannot be fully excluded, though circulation of other respiratory viruses was low and serology/RT-PCR for SARS-CoV-2 was robust. - Wearability constraints: HCWs could not wear wrist devices during work hours, reducing completeness of heart rate–based metrics; phone-based measures (steps, distance) may be less affected but still incomplete. - Small sample sizes for wearable analyses (34 with activity data; 21 with both symptom and activity trajectories) limit power and generalizability; missing data and loss to follow-up are common in remote monitoring. - Demographic skew (higher proportion female HCWs) and single health system context may limit generalizability. - Inconsistencies in data capture frequency and asynchronous timepoints required modeling assumptions and imputation; activity clusters rely on shape-based methods that may behave differently with larger datasets. - Insufficient sample to compare high/low activity trajectories pre- vs post-infection in a paired manner.