Machine learning for detecting physical function and quality of life deterioration in patients with heart failure using circadian rhythm

Abstract Background Heart failure (HF) has a significant impact on patients’ physical functional capacity and quality of life (QoL). Poor physical functional capacity and reduced QoL are associated with increased risk of (re)hospitalization and mortality in patients with HF. Monitoring physiological parameters like heart rate (HR) offer the potential for early detection of changes in physical function and QoL, allowing timely, personalized interventions. Purpose This study aims to validate the predictive accuracy of longitudinal changes in HR–based circadian parameters (mesor, amplitude, acrophase), measured by a wrist-worn device, in differentiating between deterioration and improvement of physical functional capacity and QoL in patients with HF using machine learning. Methods Twenty-nine patients wore a wrist-worn device after being admitted for acute decompensated HF. Outcome measurements were determined at baseline (T0), defined as stable HF with optimal medical therapy, and again 18 weeks later (T1). QoL was measured by Kansas City Cardiomyopathy Questionnaire (KCCQ-12) and Minnesota Living with Heart Failure Questionnaire (MLHFQ), physical functional capacity by the Short Physical Performance Battery (SPPB), and hand-grip strength using a hand dynamometer. Patients’ status was classified as "deterioration" when at least four out of seven previously defined clinically relevant decline thresholds were exceeded*. Patients not meeting the "deterioration" classification were defined as "stable/improvement". At T0 and T1, circadian parameters were extracted from the continuous wrist-worn device data, and weekly averages and changes (Δ) were calculated. These changes were labelled accordingly and stratified into 70% training and 30% testing sets. A bagged tree ensemble (496 trees, 6 splits) optimized via Bayesian tuning with five-fold cross-validation was retrained on all training data and tested on the hold-out set. Results Seventeen patients were classified as deteriorated and twelve patients had at least stable physical functional capacity and QoL. The algorithm achieved an accuracy of 75%, sensitivity of 80%, and specificity of 67% with an AUC of 0.80 (p .05) on identifying changes in physical capacity and QoL (Fig.1). Acrophase contributed the most with 38%, followed by mesor and amplitude with a contribution of 35% and 27%, respectively. Circadian parameters were similar in both groups (Table 1). Conclusion Machine learning analysis of circadian metrics can predict shifts in physical functional capacity and QoL with a relatively high accuracy when measured with a wrist-worn device. These findings present a promising avenue for delivering timely patient-tailored lifestyle interventions to patients with heart failure when continuous monitoring is integrated into standard care. *ΔKCCQ ≤–10 points1; ΔMLHFQ ≥5.0 points2; ΔSPPB ≤–1 point; Δchair stand ≥2.6 s3; Δgait speed ≤–0.08 m/s4; Δgrip strength ≤–2.5 kg5