DOI: 10.1161/circ.148.suppl_1.15326 ISSN: 0009-7322

Abstract 15326: A Wearable Accelerometry and AI Framework for Phenotyping Step-Count Trajectories in Older Adults Undergoing Cardiac Rehabilitation

Souptik Barua, Kelly Tellez, Yuchen Meng, Camila Johanek, Stephanie Pena, Samrachana Adhikari, Antoinette Schoenthaler, John A Dodson
  • Physiology (medical)
  • Cardiology and Cardiovascular Medicine

Background: Wearable accelerometers, which record daily physical activity (PA) patterns over time, are a common component of mobile health-enabled cardiac rehabilitation (mHealth-CR). Artificial intelligence (AI) algorithms may be able to identify personalized insights of PA patterns to optimize mHealth-CR adherence.

Methods: We used data from the RESILIENT trial, an ongoing randomized clinical trial of mHealth-CR in patients aged ≥65 years with ischemic heart disease. All participants in the intervention arm received Fitbit devices for activity tracking. We analyzed Fitbit step-count data collected over 3 months in these individuals using an AI framework based on functional principal component analysis (FPCA) and k-means clustering to identify distinct step count-based PA phenotypes. We then also computed adherence to using the Fitbit over time and clustered resultant adherence trajectories to identify distinct adherence patterns within the participant cohort.

Results: We analyzed data from the first 75 intervention arm participants in RESILIENT (mean age 72 years, 19% female, 32% non-White). Using our AI framework, we identified four PA phenotypes based on their step count trends over 3 months: Unchanged (n=32), Steady Increase (n=14), Increase then Decrease (n=13), and Decrease and Increase (n=14) ( Figure 1A ). Two participants’ step count trajectories were not classifiable within this framework. We also examined changes in adherence to Fitbit use over 3 months based on a minimum step count cutoff for nonadherence (<128 steps). Applying our AI framework, we identified four adherence phenotypes: Constantly adherent (n=39) , Mostly adherent (n=13) , Moderately adherent (n=14) , and Mildly adherent (n=9) (( Figure 1B )).

Conclusions: Wearable accelerometry combined with AI can provide novel insights into PA patterns during CR, which may help inform personalized PA regimens that can maximize adherence to mHealth-CR.

More from our Archive