Early detection of high blood pressure (BP) is of paramount relevance because hypertension is the main risk factor for many cardiovascular diseases (CVDs). This work evaluates the need of per-subject calibration when it comes to discriminate between normotensive (NTS) and hypertensive (HTS) subjects employing machine learning classifiers applied to electrocardiographic (ECG) and photoplethysmographic (PPG) recordings.
A total of 668 ECG, PPG and BP recordings from 51 subjects were analysed. After signal preprocessing and feature selection, 17 discriminatory features, such as pulse arrival and transit times, morphological features of the PPG signal and its first and second derivates were obtained to train machine learning-based classifiers. K-nearest neighbors (KNN), support vector machines (SVM) and ensemble classifiers were selected because they provided the best discriminatory results. The relevance of previous per-subject calibration before classification was evaluated by sequential validation, using both close and distant in time calibration measurements varying from less than 1h to more than 24h with respect to test measurements.
KNN classifier provided the best outcomes discriminating between NTS and HTS subjects with F1-score of 92.63%, sensitivity (Se) of 94.29% and specificity (Sp) of 92.80%, compared to SVM (F1-score of 90.24%, Se of 92.50% and Sp of 90.30%) and Ensemble classifiers (F1-score of 90.20%, Se of 90.36% and Sp of 92.24%). KNN classification accuracy for new subjects before calibration was 56.79%, but the inclusion of just one calibration measurement improved classification accuracy by 30\%, reaching gradually more than 97% with succesive calibrations close in time to the test measurement. Classification accuracy decreased with distance to calibration, but remained well above 83% even days after the last calibration.
NTS and HTS subjects discrimination can be significantly improved combining PPG and ECG recordings with previous per-subject calibration and, therefore, could be used for the prevention of hypertension and CVDs implementing these techniques in wearable devices.