Left ventricular hypertrophy (LVH) can arise from chronic elevation of left ventricular afterload or various other cardiomyopathies, the latter warranting specialized care with potentially family screening and sudden cardiac death prevention. Such diagnoses can be observed in patients with even mild LVH, stressing the need for a complete workup in all patients. The population of patients with left ventricular wall thickness of 12mm or more may prove especially large, making this need unrealistic to satisfy. Screening tools to determine the subset of patients who would benefit from this workup are therefore required.

To predict the hypertensive origin of LVH using machine learning.

We used a retrospective single-center population of patients with LVH, starting at 12mm of maximal left ventricular wall thickness. After splitting data in a training and testing set, we trained three different algorithms (decision tree, random forest, and support vector machine) based on first-line clinical, laboratory, and echocardiographic variables. Model performances were validated on the testing set. Classification thresholds were adjusted to ensure specificity>95% for the last two algorithms.

The population consisted of 591 patients, 372 (63%) had arterial hypertension and 141 (24%) had an etiological diagnosis of hypertensive cardiomyopathy for their LVH. All models exhibited very good areas under receiver operating curves (Fig. 1): 0.82 (0.77–0.88) for the decision tree, 0.90 (0.85–0.94) for the random forest, and 0.90 (0.85–0.94) for the support vector machine. After threshold selection, the last model had the best balance between specificity of 0.96 (0.91–0.99) and sensitivity of 0.31 (0.17–0.44). All algorithms relied on similar most influential predictor variables, among which: hypertension, systolic blood pressure at admission, estimated glomerular filtration rate, and maximal left ventricular wall thickness. Online calculators were developed.

Machine learning models were able to predict the hypertensive origin of an LVH with good performances. Implementation in clinical practice could reduce the number of etiological workups needed in patients presenting with LVH.