Key result
A machine-learning prognostic model incorporating Holter-derived heart rate variability indices did not significantly improve discrimination (AUC 0.77 vs 0.71, P=0.263) compared to a clinical-only model, but showed higher threshold-based classification performance.
Why the study?
It was uncertain whether 24-hour ambulatory ECG-derived heart rate variability indices provide complementary prognostic information in machine-learning models for heart failure beyond conventional clinical variables.
Does the addition of 24-hour Holter-derived heart rate variability indices to clinical variables improve machine-learning prognostic models for predicting cardiovascular death or heart failure hospitalization in patients with chronic heart failure?
Observational (n=353)
No
Does the addition of 24-hour Holter-derived heart rate variability indices to clinical variables improve machine-learning prognostic models for predicting cardiovascular death or heart failure hospitalization in patients with chronic heart failure?
Absolute Event Rate: 0.77% vs 0.71%
p-value: p=0.263
Incorporating 24-hour Holter-derived heart rate variability indices into machine-learning prognostic models improves risk calibration and classification performance for predicting adverse outcomes in patients with chronic heart failure.
To determine whether heart rate variability (HRV) indices derived from 24-hour ambulatory electrocardiography provide complementary prognostic information in machine-learning prognostic models using conventional clinical variables in patients with heart failure, particularly with regard to calibration and threshold-based performance. We analyzed 353 patients with heart failure. The primary endpoint was cardiovascular death or hospitalization for worsening heart failure. Patients were randomly divided into training and test cohorts in a 3:2 ratio. Two gradient boosting models were developed: one using clinical variables and HRV indices, and the other incorporating clinical variables only. To ensure a fair comparison and limit model complexity, the number of selected predictors was restricted to five in both models. The primary endpoint occurred in 49 patients (14%) during a median follow-up of 740 days (interquartile range 566–870 days). In the independent test cohort ( n = 141), discrimination (AUC) was similar between the two models (0.77 versus 0.71; DeLong P = 0.263). However, the HRV model showed higher threshold-based classification performance (higher F1 score and accuracy) and a calibration slope closer to 1.0 (slope 1.01 versus 0.73). In machine-learning-based prognostic modeling for heart failure, Holter-derived HRV indices may be associated with clinically relevant classification performance and a calibration slope closer to 1.0, although the improvement in discrimination as assessed by AUC was not statistically significant. These findings suggest that heart rate variability may provide complementary prognostic information not fully captured by standard clinical variables alone.
No takes yet. Share an insight, caveat, or question.
Matsuwaki et al. (2026) conducted an observational in Chronic heart failure (n=353). Machine-learning prognostic model incorporating clinical variables and HRV indices vs. Machine-learning prognostic model incorporating clinical variables only was evaluated on Discrimination (AUC) for cardiovascular death or hospitalization for worsening heart failure (p=0.263). A machine-learning prognostic model incorporating Holter-derived heart rate variability indices did not significantly improve discrimination (AUC 0.77 vs 0.71, P=0.263) compared to a clinical-only model, but showed higher threshold-based classification performance.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: