Establishing a novel model to assess exercise capacity in chronic heart failure based on stress echocardiography

Abstract

Aims: The aim of this study was to develop a simple, fast and efficient clinical diagnostic model, composed of exercise stress echocardiography (ESE) indicators, of the exercise capacity of patients with chronic heart failure (CHF) by comparing the effectiveness of different classifiers.

Methods and results: Eighty patients with CHF (aged 60 ± 11 years; 78% male) were prospectively enrolled in this study. All patients underwent both cardiopulmonary exercise test (CPET) and ESE and were divided into two groups according to the VE/VCO₂ slope: 30 patients with VE/VCO₂ slope ventilation classification (VC)1 (i.e., VE/VCO₂ slope < 30) and 50 patients with VC2 (i.e., VE/VCO₂ slope ≥ 30). The analytical features of all patients in the four phases (rest, warm-up, peak and recovery phases) of ESE included the following parameters: left ventricular (LV) systolic function, LV systolic function reserve, LV diastolic function, LV diastolic function reserve and right ventricular function. Logistic regression (LR), extreme gradient boosting trees (XGBT), classification regression tree (CART) and random forest (RF) classifiers were implemented in a K-fold cross-validation model to distinguish VC1 from VC2 (LVEF in VC1 vs. VC2: 44 ± 8% vs. 43 ± 11%, P = 0.617). Among the four models, the LR model had the largest area under the curve (AUC) (0.82; 95% confidence interval [CI]: 0.73 to 0.92). In the multiple-variable LR model, the differences between the peak-exercise-phase and resting-phase values of E (ΔE), s'_peak and sex were strong independent predictors of a VE/VCO₂ slope ≥ 30 (P value: ΔE = 0.002, s'_peak = 0.005, sex = 0.020). E/e'_peak, ΔLVEF, ΔLV global longitudinal strain and Δstroke volume were not predictors of VC in the multivariate LR model (P > 0.05 for the above).

Conclusions: Compared with the LR, XGBT, CART and RF models, the LR model performed best at predicting the VE/VCO₂ slope category of CHF patients. A score chart was created to predict VE/VCO₂ slopes ≥ 30. ΔE, s'_peak and sex are independent predictors of exercise capacity in CHF patients.

Keywords: Cardiopulmonary exercise test; Chronic heart failure; Exercise capacity; Exercise stress echocardiography; Machine learning; VE/VCO2 slope.

Figure 1

Changes in various parameters—E (A), E/e′ (B), e′ (C), s′ (D), LVEF (E) and LVGLS (F)—of heart failure patients with different ventilation classifications at four phases of exercise stress echocardiography. E, early diastolic peak flow velocity of the mitral valve; E/e′, ratio of transmitral E velocity to tissue Doppler mitral annulus e′ velocity; e′, early diastolic velocity of the mitral annulus; LVEF, left ventricular ejection fraction; LVGLS, left ventricular global longitudinal strain; s′, systolic velocity of the mitral annulus; VC, ventilation classification.

Figure 2

Diagnostic effectiveness and feature importance of each model. (A) Receiver operating characteristic curves of the XGBT, RF, CART and LR models. (B) Bar chart of the SHAP values of the LR model. (C) Ribbon graphs of the importance of each feature in the XGBT, RF and CART algorithms. CART, classification regression tree; CO, cardiac output; LR, logistic regression; RF, random forest; s′_peak, systolic velocity of the mitral annulus in the peak phase; SHAP, Shapley additive explanations; XGBT, extreme gradient boosting trees; ΔE, difference of E values between peak and rest; Δe′, difference of e′ values between peak and rest.

Figure 3

Graphical score chart for the probability of a VE/VCO₂ slope ≥ 30 in patients with chronic heart failure based on the LR model. (Dark red, probability > 90%; light red, 75–90%; orange, 50–75%; yellow, 30–50%; green, <30%.) ΔE, difference in E values between peak and rest; s′_peak, systolic velocity of the mitral annulus in the peak phase.