Cross-Database Generalization of a Single-Feature Threshold Model for Congestive Heart Failure Detection from Short-Term Heart Rate Variability

Objective To evaluate the transferability of interpretable single-feature threshold models across the five evaluated databases for congestive heart failure (CHF) detection from short-term heart rate variability (HRV), and to compare their performance with optimized multi-feature classifiers. Approach Twenty HRV features from 60-second inter-beat interval windows in five databases were evaluated using leave-one-database-pair-out nested cross-validation with six outer-test sets and their corresponding outer-development sets. Within each outer-development set, directions and thresholds were fitted on inner-train sets, and features were selected according to mean inner-validation F1 score. The selected model was refitted on the complete outer-development set. The most frequently selected feature was further evaluated across the six outer-test sets with separately fitted directions and thresholds. Eight multi-feature classifiers were optimized and evaluated using the same data splits. Performance was assessed using record-balanced window-level metrics and database-stratified recording-level bootstrap confidence intervals. Main results The short-term detrended fluctuation analysis scaling exponent (f14) was selected most frequently. The f14 threshold model achieved the highest mean test AUC (0.881 ± 0.038), F1 score (0.718 ± 0.079), and sensitivity among the evaluated single-feature and multi-feature models. Among the eight multi-feature classifiers, Random Forest achieved the highest mean test AUC (0.858 ± 0.067). Significance An interpretable single-feature threshold model showed strong CHF-Normal Sinus Rhythm (NSR) discrimination in the five evaluated databases. f14 was the strongest candidate in this evaluation, although its thresholds varied across outer-development sets and its clinical specificity requires further validation.