ECG and PPG-Based Hypertension Screening Under Non-Hypertensive Blood Pressure Recordings

Jesús Cano1, Vicente Bertomeu-Gonzalez2, Lorenzo Fácila3, Jos� Moreno-Arribas2, Raul Alcaraz4, José J Rieta5, Universitat Politecnica de Valencia, 2Clinical Medicine Department, Miguel Hernandez University, 3Cardiology Department, General University Hospital Consortium of Valencia, 4University of Castilla-La Mancha,, Universitat Politecnica Valencia


Blood pressure (BP) fluctuates throughout the day in every subject, mainly due to circadian oscillations as well as a response to physical and mental stimuli. However, BP variability is larger in hypertensive (HTS) patients than in normotensive (NTS) subjects, being proportional to the increase in mean BP. This study aims at investigating whether machine learning (ML) classifiers can detect the pathology of hypertension regardless of absolute BP values. The goal is to identify HTS patients from non-HTS recordings and NTS subjects from non-NTS recordings using photoplethysmographic (PPG) and electrocardiographic (ECG) recordings.

A total of 803 simultaneous PPG, ECG and invasive BP recordings from 51 subjects were analyzed. 668 were coherent BP segments, with high BP for HTS patients and normal BP for NTS subjects, and 135 were incoherent segments, with normal BP for HTS patients and high BP for NTS subjects. The relationship between BP and the PPG was evaluated with features representing arterial wave propagation and morphological PPG features, such as pulse area, peak-to-peak time or pulse arrival time. Up to 37 classification models as Decision Trees, Nearest Neighbors, Support Vector Machine or Ensemble Classifiers were evaluated to classify incoherent segments.

Using the obtained discriminant features of coherent segments for training and the set of incoherent segments for validation, the classification model that provided the best outcomes was K-nearest neighbors with F1-score of 88.30%, sensitivity (Se) of 79.81% and specificity (Sp) of 96.77%, whereas Bagged Trees classifier provided F1-score of 79.26%, Se of 75% and Sp of 93.55%. High Sp values reflected correctly identified NTS individuals when BP reached prehipertensive or HTS values.

Combining PPG and ECG recordings with ML-based methodologies would be of high interest for hypertension screening, so that HTS patients and NTS subjects could be properly discerned even in the case of incoherent or altered BP values. This method could be used as a support for clinical decision-making when diagnosing hypertension.