Prediction of abnormal left ventricular geometry on the ECG through machine learning with a focus on obesity. (14th October 2021)

• Record Type: Journal Article
• Authors: Aggelaki, E; Marketou, M; Barmparis, G; Patrianakos, A; Kochiadakis, G; Vardas, P; Parthenakis, F; Tsironis, G
• Is Part Of: European heart journal. Volume 42(2021)Supplement 1
• Publication Date: 2021-10-14
• Subjects: Clinical; Cardiology -- Periodicals; Heart -- Diseases -- Periodicals; 616.12005
• DOI: 10.1093/eurheartj/ehab724.2386
• Languages: English
• ISSNs: 0195-668X

Abstract

Abstract: Background: Cardiac remodeling, an important aspect of cardiovascular disease (CVD) progression, is emerging as a significant therapeutic target. However, the ECG is not a sensitive method of detecting left ventricular hypertrophy (LVH), and as far as we know, it cannot detect changes in left ventricular geometry (LVG) at early stages, especially before LVH is present. Its sensitivity is particularly low for obese patients. Purpose: To use a machine learning (ML) classifier to detect abnormal LVG from EKG parameters/markers, even before it becomes LVH, and to propose some indicative markers useful for practitioners. We also looked at the results of our model for obese patients to test the markers in this population. Methods: We enrolled 594 consecutive subjects, aged 30 years or older (mean age: 61.6±12 years old) with and without essential hypertension and no indications of CVD. We tried to build a "clean" dataset through which we can target the clinical, anthropometric, and electrocardiogram measurements indicative of abnormal LVG. All patients underwent a full echocardiographic evaluation and were classified into 2 groups; those with normal geometry (NG) vs. those with concentric remodeling (CR) or LVH. Abnormal LVG was identified as increased relative wall thickness (RWT) and/or left ventricular mass index (LVMi). We analyzed the EKG waveforms deduced to single beat averages for each lead using custom software and extracted 70 markers. We then trained a Random Forest machine learning model to classify subjects with abnormal LVG and calculated SHAP values to perform feature importance and interaction. Results: The percentage of women was 56.5%, while 71.3% of all patients were hypertensive. Hypertension, age, body mass index divided by the Sokolow-Lyon voltage (BMI/S-L), QRS-T angle, and QTc duration were among the most important parameters (Figure, left panel) identified by the model as being predictive of abnormal LVG (AUC/ROC = 0.84, sensitivity = 0.94, specificity 0.61). Specifically for obese patients, whose prevalence in our population was 60.3%, our model performed well (sensitivity = 0.71, specificity = 0.92. When we tried our model without the the BMI/S-L parameter, the specificity dropped to 0.88. We also found that a cut-off point of 18 for the BMI/S-L marker predicted the patients who were more probable to have developed abnormal LVG (Figure 1). Conclusions: This study is the first to demonstrate the promising potential of ML modeling for the efficient and cost-effective diagnostic screening of abnormal LVG through ECG. We found specific clinical and ECG parameters that can predict early pathological changes of LVG in patients without established CVD and detect the population who will benefit from a detailed echocardiographic evaluation. Our model contributes to the development of human-centered and autonomous technologies and can optimize patient-management and treatment. Funding Acknowledgement: Type of funding sources: None.