ExHyptNet: An explainable diagnosis of hypertension using EfficientNet with PPG signals

Background: Hypertension is a crucial health indicator because it provides subtle details about a patient's cardiac health. Photoplethysmography (PPG) signals are a critical biological marker used for the early detection and diagnosis of hypertension. Objective: The existing hypertension detection models cannot explain the model's prediction, making it unreliable for clinicians. The proposed study aims to develop an explainable and effective hypertension detection (ExHyptNet) model using PPG signals. Methods: The proposed ExHyptNet model is an ensemble of multi-level feature analyses used to detect and explain hypertension predictions. In the feature extraction stage, recurrence plots and EfficientNetB3 architecture are employed to extract deep features from the PPG signals. Then, features are explained using a Gradient-weighted Class Activation Mapping (Grad-CAM) explainer in the explainable stage. In the last stage, XG-Boost and extremely randomized trees (ERT) classifiers are used to make the qualitative and quantitative analysis for evaluating the performance of the proposed ExHyptNet model. Results: The performance of the ExHyptNet model is evaluated on two public PPG datasets: PPG-BP and MIMIC-II, using holdout, stratified 10-fold cross-validation, and leave-one-out subject validation techniques. The developed model yielded a 100% detection rate for the classification of normal and multi-stage hypertension classes using three validation techniques. The proposed work also demonstrates a detailed ablation study using hyper-parameters, pre-trained models, and the detection of several PPG categories. Conclusion: The developed ExHyptNet model performed better than the existing automated hypertension detection systems. Our proposed model is practically realizable to clinicians in real-time hypertension detection as it is validated on two public PPG datasets using different validation techniques.