Application of Deep Learning to Electrocardiography for Hypertension Detection

Early detection of hypertension (HT) is crucial for effective HT management in order to avoid health consequences. The so far existing methods, however, show significant limitations in providing continuous and timely data. Considering the contribution of deep learning (DL) in healthcare and the robustness of electrocardiography (ECG) technology when it comes to cardiovascular-related research, their combination is a promising alternative approach for improved HT detection. The present study investigates this assumption. 20520 2.5-s ECG recordings from 171 patients of the MIMIC database are utilized, classified between normotensive (NT), prehypertensive (PHT) and HT categories corresponding to sysolic blood pressure (BP) <120 mmHg,  <140 mmHg and ≥140 mmHg, respectively. 80% of the recordings were used as the training set, trained by InceptionResnet-v2 CNN with given parameters. ECG signal conversion to images is performed with recurrence plots, using m = 3 and τ = 8 ms for phase-space reconstruction. The results were assessed in 3 different scenarios, by calculating the performance metrics on the test set. Scenario A: NT/PHT/HT. Scenarios B and C: NT/non-NT (PHT and HT combined) and HT/non-HT. Results on scenarios A and B indicate high performance, with accuracies about 97% and balanced sensitivity and specificity metrics (96.77 − 97.42%). When HT/non-HT classification was assessed, the performance dropped (accuracy: 82.01%, imbalanced sensitivity and specificity per class: 69.9 − 98.9%). The reported results indicate the high potential of DL and ECG combination in alleviating the HT detection issues and assisting the overall effort against elevated BP.