Prediction of the left ventricular ejection fraction by machine learning algorithms based on heart rate variability parameters in patients with ischemic heart disease

Abstract

Background The left ventricular ejection fraction (LVEF) plays a pivotal role in the diagnosis and prediction of ischemic heart disease (IHD). Current techniques to measure LVEF have led to some complications and are relatively expensive despite the high accuracy. Heart rate variability (HRV) is an alternative for the assessment of cardiac function and its related parameters are easily to be derived from electrocardiography. Objective This study aimed to investigate the corresponding relationship between LVEF and HRV in patients with IHD via prediction models developed by machine learning techniques and compare the accuracy of these models. Methods In this retrospective real-world study, patients with IHD admitted to Jinan 4th People’s hospital between January 2019 and December 2020 were randomly selected and divided into a derivation cohort and a validation cohort in a 7:3 ratio. LVEF were measured by color Doppler echocardiography while HRV-related parameters were derived from a 24h Holter electrocardiography recordings. Patient clinical features and HRV parameters were identified and used in the subsequent modelling process. Logistic regression, naïve Bayes, decision tree, gradient boosting, extreme gradient boosting (XGBoosting) and random forests were respectively applied to estimate the correlation between HRV and LVEF. Model performances were evaluated by the coefficient R2 (the closer the result is to 1, the better) and the accuracy (a precise prediction was defined as the accuracy ≥ 90%). Results A total of 179 patients were included in this study with 125 (mean age 67.03 ± 16.12, male 72 [53.7%]) in the derivation cohort and 54 (mean age 62.51 ± 12.97, male 19 [48.7%]) in the validation cohort. Six variables including sex, age, SDNN, SDANN, rMSSD and pNN50 were found to be correlated with LVEF. Gradient boosting and random forests outperformed other machine learning models, showing the best prediction performance (> 90%) in the prediction of both LVEF percentage (%) and LVEF stratifications (< 40%, 40–50%, > 50%). These six indicators showed different contributions in different algorithms with sex as the least important indicator. Conclusion Gradient boosting and random forests machine learning models based on HRV- and patient characteristics-related parameters provided a new method in predicting LVEF in patients with IHD, which showed a potential in the diagnosis of IHD.