Abstract
Abstract Accurate and non-invasive blood pressure (BP) estimation plays a vital role in continuous cardiovascular health monitoring. This study presents BiSpectral MRRNet, a cascaded dual-stream architecture designed to extract deep features from the spectro-temporal representation of PPG and ECG signals. The model was trained on a hybrid dataset comprising 378 subjects, including primary data acquired using a custom ESP8266-based pulse sensor kit and secondary data sourced from the MIMIC-III and PulseDB Vital datasets, with proportional subject wise representation maintained during dataset reduction. Following initial training, the network was evaluated both as a standalone architecture and within a custom ensemble feature-level learning framework utilizing SVR, LightGBM, XGBoost, and Random Forest regressors. The BiSpectral MRRNet with z-score normalization and stacked ensemble achieved a mean absolute error (MAE) of 3.949 mmHg (SBP) and 2.392 mmHg (DBP) under random split evaluation, with correlation coefficients exceeding 0.94. The same model demonstrated strong generalization under subject wise evaluation, achieving an MAE of 3.780 mmHg (SBP) and 2.428 mmHg (DBP), confirming robustness across unseen subjects. The proposed framework satisfies AAMI and BHS clinical validation standards, attaining Grade A performance for both systolic and diastolic blood pressure estimation under random and subject wise evaluation. The trained model was further integrated into a cloud-based inference pipeline and deployed through a Flutter-based mobile application, enabling real-time cuffless BP estimation using wearable PPG acquisition. The proposed end-to-end framework demonstrates the potential for continuous, non-invasive cardiovascular monitoring in wearable healthcare applications.