Blood Pressure Estimation From Beat-by-Beat Time-Domain Features of Oscillometric Waveforms Using Deep-Neural-Network Classification Models

In general, existing machine learning based approaches, developed for systolic and diastolic blood pressure (SBP and DBP) estimation from oscillometric waveforms (OWs), employ features extracted from the OW envelope (OWE) alone and ignore important beat-by-beat (BBB) features which represent fundamental physical properties of the entire non-invasive blood pressure (NIBP) measurement system. Unlike the existing literature, this paper proposes a novel deep-learning based method for BP estimation trained with BBB time-domain features extracted from OWs. First, we extract six time-domain features from each beat of the OW, relative to the preceding beat. Second, using the extracted BBB features along with the corresponding cuff pressures, we form a feature vector for each OW beat and locate it in one of three different classes, namely pre-systolic (PS), between systolic and diastolic (BSD) and after diastolic (AD). We then devise a deep-belief network (DBN)-deep neural network (DNN) classification model as well as a novel artificial feature extraction method for estimating SBP and DBP from feature vectors extracted from OWs and their corresponding deflation curves. The proposed DBN-DNN classification approach can effectively learn the complex nonlinear relationship between the artificial feature vectors and target classes. The SBP and DBP points are then obtained by mapping the beats at which the network output sequence switches from PS phase to BSD phase and from BSD phase to AD phase, respectively, to the deflation curve. Adopting a 5-fold cross-validation scheme and using a data base of 350 NIBP recordings gave an average mean absolute error of 1.1±2.9 mmHg for SBP and 3.0±5.6 mmHg for DBP relative to reference values. We experimentally show that the proposed DBN-DNN-based classification algorithm trained with BBB time-domain features can outperform traditional deep-learning based methods for BP estimation trained with features extracted only from OWEs.