Utilizing Circadian Heart Rate Variability Features and Machine Learning for Estimating Left Ventricular Ejection Fraction Levels in Hypertensive Patients: A Composite Multiscale Entropy Analysis

Figures

• 

  Figure 1. QRS wave detection.

• 

  Figure 2. RMSE corresponding to feature sets for 20:00–21:00 time interval utilizing LR model. Red triangle marks iteration at which feature addition ceases.

• 

  Figure 3. Model performance fluctuates throughout the 24 h period and differs across various regression models with the best feature subset. The heatmaps highlight the best-performing regression models for each hour of the day. Time points are arranged in ascending order, progressing clockwise. Model performance is visualized based on RMSE (A) and MAE (B) (red is better, blue is worse). The numbered circles (I–X) correspond to the ten regression models evaluated in the study.

• 

  Figure 4. Bland–Altman plots for four specific hours using LR (a), Treebag (b), SVMPoly (c), and LR (d), respectively. Differences were computed as predicted minus actual LVEF values. The light blue dotted lines represent the 95% confidence intervals. The legend indicates different types of work (red for night work, black for day work).

• 

  Figure 5. Feature importance value corresponding to the best feature sets for the 20:00–21:00 time interval based on the LR model. To better visualize the importance of different CMSE-based HRV feature categories to the model, the lollipop chart is organized in two stages: first, sorted by the total scores of individual features, and then grouped and sorted within each category. Features are divided into four distinct color-coded groups: Scale 1, Scale 5, Slope 1–5, and Area 1–5. The y-axis lists the features derived from various CMSE parameters, arranged within each category from the most to least important. The x-axis displays the importance scores. Each feature’s importance is represented by a circle, with the size reflecting the score and the exact value labeled within each circle for clarity.

• 

  Figure 6. SHAP analysis. (A) The importance ranking of the top 20 features according to the mean (|SHAP value|); (B) the importance ranking of the top 20 features using the optimal regression model. The suffixes of different features represent various CMSE parameters: “sc1” denotes Scale factor 1, “sc5” denotes Scale factor 5, “a15” represents Area factor 1–5, and “s15” stands for Slope factor 1–5.