Predicting acute decompensated heart failure using circadian markers from heart rate time series

Hospital admissions for acute decompensated heart failure (ADHF) are linked to high readmission rates, emphasizing the need for early intervention. Dysregulation of the circadian rhythm that regulates key physiological processes, such as heart rate (HR), blood pressure and sleep-wake cycles, may precede weight gain and clinical symptoms of worsening heart failure (HF) by weeks, providing a window for timely intervention. This study aims to develop a predictive algorithm for early and accurate ADHF detection.

Sixty-five patients discharged after ADHF hospitalization monitored HR with a wrist-worn device for 6 months after reaching stable HF. Circadian parameters (mesor, amplitude and acrophase) were extracted via cosinor analysis and used to train a long short-term memory neural network. The algorithm analysed 21-day periods before an HF event, defined as unplanned outpatient visits for congestion episode, increased diuretics, ADHF hospitalization or sudden cardiac death. Circadian changes appeared in the 21 days preceding HF events, with elevated mesor (70.6 vs. 73.6 b.p.m.; P < 0.001), reduced amplitude (8.3 vs. 4.9 b.p.m.; P = 0.046) and acrophase shifts (11.3 vs. 12.2 h; P = 0.706). The classification algorithm showed 74% sensitivity, 73% specificity and a 74% AUC (P < 0.001). Amplitude was the strongest predictor, contributing 62% to the algorithm's feature importance.

Circadian metrics from a wrist-worn device showed progressive alterations over the 3 weeks preceding ADHF, offering potential early detection of HF decompensation with moderate prediction performance. Future research should refine these metrics and results in larger, diverse populations, using various sensor types and explore early interventions.