Single-lead ECG based detection of arterial hypertension using artificial intelligence algorithms

Abstract Background Unawareness of arterial hypertension is an important contributing factor to the inadequate control of the disease and the absence of appropriate antihypertension treatments. Population screening programs have shown that more than 50% of hypertensives were unaware they had hypertension. Wearable devices, a ubiquitous accessory in our days, can record single-lead electrocardiograms that are then analyzed by machine learning (ML) algorithms. The recorded signal, most often of lead I, could potentially reveal an undiagnosed hypertension. Purpose The purpose of this study is to create and train a ML algorithm that detects the existence of arterial hypertension using features derived from a single-lead ECG (lead I), Methods We enrolled 1220 consecutive individuals without cardiovascular disease. The participants were classified into hypertensive and normotensive group. Hypertensive patients were recruited from the outpatient clinics of the respective centers. Normotensive healthy individuals were referred either for the investigation of atypical chest pain or for the modification of risk factors for cardiovascular disease such as hyperlipidemia. We trained a Random Forest (RF) model to classify subjects into 2 groups: hypertensive or normotensive. A feature-based machine learning algorithm, such as a RF, can also provide an interpretation of its results in the form of feature importance and interactions between features. Results The prevalence of hypertensive people in our study was 66.5%. Our RF model was able to distinguish hypertensive from normotensive patients on a hold-out test set, with an AUC 0.85, accuracy 81%, sensitivity 80% and specificity 83% (both using a threshold of 0.5), and an F1 score 0.86. When predicting on the subset of the test set containing people younger than 59 years old, it achieved AUC of 0.75, sensitivity of 50% and specificity of 87%. Age, BMI, the area under the T wave divided by the QRS complex area, the area under QRS segment adjusted for BMI, the QT corrected interval, and the PQ interval, were the most important anthropometric and ECG-derived features in terms of the success of our model (Figure 1). Conclusions Our findings indicate that one-lead ECG based detection of arterial hypertension is possible by using ML algorithms. This study demonstrates the huge potential of artificial intelligence to transform smartwatches ECG signals into diagnostic tools.Figure 1