Using AI to identify left ventricular ejection fraction from the ECG: The SOLOMAX (SOciaL NetwOrk of MedicAl Experiences) project

Abstract

Aim: The interest in machine learning-based algorithms in the cardiovascular field is rapidly growing, especially for diagnostic and prognostic purposes. Recent evidence has demonstrated that certain electrocardiographic (ECG) parameters are predominantly associated with systolic function, estimated as left ventricular ejection fraction (LVEF) by echocardiography, albeit with still relatively low accuracy.
Consequently, this study aims to develop an AI-based model capable of predicting LVEF from ECG data in an Italian population.
Methods: Within the SOLOMAX project, we collected paired ECG-Echocardiography exams from 105 patients (64.82±16.02y;62.86%male). Precisely, we excluded patients with atrial fibrillation at the time of the ECG, PMK or electrostimulated rhythm, valve prostheses, previous cardiac surgery, O2 therapy or COPD, previous ablation or invasive electrophysiology procedures, currently hospitalized for Takotsubo or ACS, heart failure exacerbation, inotropic therapy, ACS over the last 3 months. We recorded anthropometric, clinical, biochemical, ECG, and Echocardiography parameters. The collected data was studied using AI-based techniques to create a new model to predict LVEF from ECG. Using an approach based on evolutionary algorithms, genetic programming was used. This approach solves a symbolic regression problem through genetic algorithms and provides a mathematical model of the relationship between ECG parameters and LVEF. The formula obtained was then used to build a simple explainable classifier, which provides a global interpretation of the link between ECG parameters and LVEF.
Results: The performance of the proposed approach and the reliability of the results were assessed using the k-fold cross-validation method and by estimating standard metrics derived from the confusion matrix associated with a binary classifier, that is, accuracy, sensitivity, specificity, precision, and F-Measure. The proposed approach consistently demonstrated its ability to distinguish patients with preserved LVEF from those with reduced LVEF. Each metric averaged across all experiments scored approximately 95%. Furthermore, in the expression generated by the AI model, the axes of the P, QRS, and T waves play a prominent role, as they are likely to provide a better interpretation of the three-dimensional cardiac geometry and, consequently, cardiac function.
Conclusions: AI applied to ECG data can be used to create cost-effective diagnostic and predictive tools for assessing LVEF. Indeed, the obtained formula highlights the relationship between ECG parameters and LVEF, as well as its complexity, which can aid in detecting heart diseases.