A machine learning approach for hypertension detection based on photoplethysmography and clinical data.

High blood pressure early screening remains a challenge due to the lack of symptoms associated with it. Accordingly, noninvasive methods based on photoplethysmography (PPG) or clinical data analysis and the training of machine learning techniques for hypertension detection have been proposed in the literature. Nevertheless, several challenges arise when analyzing PPG signals, such as the need for high-quality signals for morphological feature extraction from PPG related to high blood pressure. On the other hand, another popular approach is to use deep learning techniques to avoid the feature extraction process. Nonetheless, this method requires high computational power and behaves as a black-box approach, which impedes application in a medical context. In addition, considering only the socio-demographic and clinical data of the subject does not allow constant monitoring. This work proposes to use the wavelet scattering transform as a feature extraction technique to obtain features from PPG data and combine it with clinical data to detect early hypertension stages by applying Early and Late Fusion. This analysis showed that the PPG features derived from the wavelet scattering transform combined with a support vector machine can classify normotension and prehypertension with an accuracy of 71.42% and an F1-score of 76%. However, classifying normotension and prehypertension by considering both the features extracted from PPG signals through wavelet scattering transform and clinical variables such as age, body mass index, and heart rate by either Late Fusion or Early Fusion did not provide better performance than considering each data type separately in terms of accuracy and F1-score.