Hypertension Risk Assessment from Photoplethysmographic Recordings Using Deep Learning Classificators Jes´ us Cano 1 ,Vicente Bertomeu-Gonz´ alez 2 , Lorenzo F´ acia 3 , Roberto Zangr ´ oniz 4 , Ra ´ ul Alcaraz 4 , Jos´ e J Rieta 1* 1 BioMIT.org, Electronic Engineering Department, Universitat Politecnica de Valencia, Spain 2 Clinical Medicine Department, Miguel Hern´ andez University, Elche Spain 3 Cardiology Department, Hospital General Universitario de Valencia, Spain 4 Research Group in Electronic, Biomed. and Telecomm. Eng., Univ. of Castilla-La Mancha, Spain Abstract Regular monitoring of blood pressure (BP) is essen- tial to make an early detection of cardiovascular dis- eases caused by hypertension, a potentially deadly con- dition that do not present symptoms in its first stages. This study aims to investigate whether deep learning techniques can assess risk levels of BP using only photoplethysmo- graphic (PPG) recordings without the need of electrocar- diographic (ECG) recordings, as in many previous stud- ies. 15.240 segments from 50 different patients contain- ing simultaneous PPG and arterial blood pressure (ABP) signals were analysed. GoogleNet and ResNet pretrained convolutional neural networks (CNN) with the scalogram of PPG signals obtained by continuous wavelet transform (CWT) used as input images were employed for the clas- sification. The highest F1 score was achieved by dis- criminating normotensive (NT) patients from prehyper- tensive (PH) and hypertensive (HT), being 92.10% for GoogleNet and 93.91% for ResNet, respectively. In ad- dition, intra-patient classification using different data seg- ments for training and validation provided an F1 score of 90.28% with GoogleNet and 89.04% with ResNet. Time frequency transformation of PPG recordings to feed deep learning classifiers has been able to provide outstanding results in hypertension risk assessment without requiring either ECG recordings or feature extraction 1. Introduction Blood pressure (BP) is the most important biomarker for cardiovascular diseases, which is the leading cause of mor- tality worldwide and a major contributor to the reduction of quality of life [1]. Thus, early diagnosis and control of hypertension is essential for prevention. Moreover, most patients with severely elevated blood pressure have asymp- tomatic hypertension without signs or symptoms of end- organ damage. Without an early detection, it can derive in an hypertensive urgency with the presence of risk factors for progressive diseases as hearth failure and preexisting renal insufficiency or severe uncontrolled hypertension [2]. The most common noninvasive technique for BP mea- surement are based on uncomfortable arterial occlusion by inflatable cuffs where adequate accuracy is offered but only provide intermittent measurement and needs to be applied by professionals [3]. Recent advances in sen- sor technology have developed unobstructive cuffless de- vices to measure physiological parameters anytime. In this way, the use of photoplethysmographic (PPG) recordings is very promising for being noninvasive, with continuous measurement, low cost, simple and with a high correlation with arterial BP in frequency and time domain [4]. In this regard, many studies have applied artificial in- telligence technology in order to estimate or discriminate between blood pressure levels. Machine learning tech- niques combining electrocardiographic (ECG) and PPG signals get use of propagation theory with parameters as pulse transit time (PTT), pulse arrival time (PAT) and pulse wave velocity (PWV) to determine cardiovascular state [5]. Recent studies combine this propagation param- eters with PPG morphological feature extraction as inputs for the models [6]. In recent years, deep learning with powerful computational methods that eliminates feature extraction have shown an improvement in BP estimation from PPG signals [7] The present work proposes a method for hypertension risk assessment using the pretrained CNNs GoogleNet and ResNet and the scalogram of PPG signals by continuous wavelet transform (CWT) as inputs to feed this models without the need for ECG recordings or hand extraction of signals features for classification.