Quality-Aware PPG-Based Blood Pressure Classification for Energy-Efficient Trustworthy BP Monitoring Devices with Reduced False Alarms

Early detection of hypertension is crucial for significantly reducing the risk of serious health complications, which demands continuous or long-term blood pressure (BP) monitoring to ensure the timely management of hypertension. This paper proposes a new quality-aware photoplethysmogram (PPG) feature-based hypertension recognition framework to reduce false alarms and overall energy consumption by considering the affordable resource-constrained continuous BP monitoring devices. The proposed framework includes a high-pass filter (HPF), PPG signal quality assessment (PPG-SQA), PPGwaveform feature extraction (FE), and BP classification (hypertension and non-hypertension). In this paper, we present four SQA methods and nine machine learning based BP classification models, including logistic regression, decision tree, random forest, multilayer perceptron, k-nearest neighbours, XGBoost, AdaBoost, Bagged tree, and one-dimensional convolutional neural network (1D-CNN). Four SQA methods are based on the average magnitude difference function (AMDF/(SQA-M1)) features and the AMDF features with the total number of zero-crossings present in raw/original (SQA-M2), derivative (SQA-M3), and smoothed derivative (SQA-M4) PPG segment features. On three standard databases, results showed that the SQA-M4 method outperforms the other three SQA methods with an overall false alarm reduction rate (FARR) of 50.8% in the case of motion artifact (MA) corrupted PPG signals and FARR of 91.0% in the case of additive white Gaussian noise (AWGN)-corrupted PPG signals. The quality-aware PPG feature extraction is presented to extract twelve features, which are combined with nine ML classifiers to explore the best model with higher recognition accuracy. Among the nine BP-classify models, the random forest (RF)-based BP model achieved the highest accuracy of 92.1% with a specificity of 93.8% for the BP model with SQA approach as compared to the BP model without SQA approach. On the implementation of the Best model using the Raspberry Pi (R-Pi) platform, the quality-aware BP (QA-BP) model had FARR ranging from 32.4% to 62.2% for the NHT events and FARR ranging from 30.8% to 99.7% for the HT events in the case of random noise corrupted PPG signals, and the QA-BP model can reduce false alarms from 11.9% to 52.5% for NHT events and from 7.1% to 55.5% for HT events in the case of motion corrupted PPG signals. For the case of processing 60s noisy PPG signal, the QA-BPC method had an energy saving of 41.3% with display mode and an energy saving of 95.3% with the data transfer mode as compared to the BP classification method without the SQA method. The proposed QA-BP method had energy saving of 8.1% - 85.6% for the 10% - 90% noisy portions in 24-hour recordings and has great potential in the development of affordable, energy-efficient cuffless BP monitoring devices with reduced false alarms.