Detection of congestive heart failure from RR intervals during long-term electrocardiographic recordings

Figure 1. Illustration of the classification pipeline. [For more details about dynamical detrended fluctuation analysis (DDFA) density and DDFA α(HR, s), see Figure 5.]

Figure 2. Scaling exponents of detrended fluctuation analysis (DFA) for each of the studied subgroups and the receiving operating characteristics area under curve (AUC) values between the different subgroups. a: First-order DFA (DFA-1) short-scale scaling exponent α1. b: DFA-1 long-scale scaling exponent α2. c: Second-order DFA (DFA-2) short-scale scaling exponent α1. d: DFA-2 long-scale scaling exponent α2. Square brackets indicate 95% confidence intervals for the AUC values estimated by DeLong’s algorithm.36 Statistically different AUC values from 0.5 are indicated by asterisks. AF = atrial fibrillation; CHF = congestive heart failure.

Figure 3. Confusion matrices for the XGBoost classifier with congestive heart failure (CHF), atrial fibrillation (AF), and healthy controls. The classification features include the short- and long-scale scaling exponents of both the first- and second-order detrended fluctuation analysis.

Figure 4. Examples of scaling exponent landscapes (in color) α(t, s) (a, c, e), and α(HR, s) (b, d, f) computed with dynamical detrended fluctuation analysis for RR intervals collected from a healthy subject (a, b), congestive heart failure patient (c, d), and atrial fibrillation patient (e, f).

Figure 5. Aggregated plots of the scaling exponents α(HR, s) (in color) as functions of the scale s and heart rate (HR) resulting from dynamical detrended fluctuation analysis. a: All the healthy controls. b: All the congestive heart failure (CHF) subjects. c: All the atrial fibrillation (AF) subjects. The HR range is limited to 50–180 for visual purposes. d–f: Corresponding aggregated densities of the scaling exponent.

Figure 6. Confusion matrices for the XGBoost classifier with congestive heart failure (CHF), atrial fibrillation (AF), and healthy controls. The classification features are extracted from dynamic detrended fluctuation analysis with principal component analysis dimensionality reduction.

Figure 7. Aggregated plots of the scaling exponents α(HR, s) (in color) as functions of the scale s and heart rate (HR) resulting from dynamical detrended fluctuation analysis. The results are grouped based on the severity of the condition. a: Atrial fibrillation (AF) with >80% of the measurement duration in sinus rhythm (SR) (N = 26). b: AF with 1%–80% in SR (N = 24). c: AF with <1% in SR (N = 34). d: Congestive heart failure (CHF) with New York Heart Association (NYHA) class I or II (N = 96). e: CHF with NYHA class III or IV (N = 51). f: Healthy controls (N = 274).