BePCon: A Photoplethysmography-Based Quality-Aware Continuous Beat-to-Beat Blood Pressure Measurement Technique Using Deep Learning

Research on noninvasive blood pressure (NIBP) measurement using electrocardiogram (ECG)/ photoplethysmogram (PPG) and their combinations has been most popular in ambulatory health monitoring. The real challenge is motion artifact (MA) corruption in the PPG, which makes the blood pressure (BP) measurement unreliable. This article presents BePCon, a deep learning-based model for beat-to-beat (BtB) BP measurement using a temporal convolutional network (TCN). At first, the signal quality assessment (SQA) of PPG is done by a self-organizing map (SOM). Next, the time-domain, statistical, wavelet, and stacked autoencoder features from current and previous good quality PPG cycles are extracted. A recursive feature elimination (RFE) selects optimum set of 20 features from each cycle before being fed to the TCN to predict the systolic BP (SBP) and diastolic BP (DBP) of current beat. While evaluated over 150 data records from PhysioNet MIMIC-II/III waveform database, BePCon achieves standard deviation (SD) and mean absolute error (MAE) of 3.24 and 2.38 mmHg, respectively, for the SBP and 1.73 and 1.23 mmHg, respectively, for the DBP. An improvement of accuracy by a factor of 19.56% for SBP and 24.61% for DBP is obtained over without SQA. BePCon also complies with Association for Advancements of Medical Instrumentation (AAMI) and British Hypertension Society (BHS) Grade A standard and improvement over published works on BtB BP measurement using MIMIC-II/III waveform database. A standalone implementation with a single core 1-GHz ARM v6 controller supported by 512-MB RAM shows low latency (~2.5 s/beat) and low memory requirement (~32.22 kB/beat). This establishes that BePCon has the potential for real-time ambulatory BtB BP measurement.