Abstract
Cardiovascular diseases (CVDs) remain a leading cause of global mortality, highlighting the need for accurate, real-time, and non-invasive blood pressure (BP) estimation. Pulse signals provide essential physiological and pathological insights into cardiovascular dynamics, offering a non-invasive and continuous approach to BP monitoring. However, conventional BP monitoring methods are often invasive, introducing procedural complexities that hinder widespread clinical application. To address this challenge, we propose a machine learning-driven framework for continuous arterial blood pressure estimation using photoplethysmogram (PPG) signals. We introduce a handcrafted feature extraction tool that systematically extracts and engineer's meaningful bio-signal features from the PPG based Arterial Pulse Waveforms (APW). Our methodology employs feature fusion, synergistically integrating PPG-derived handcrafted features—such as ascending and descending times, pulse rate, pulse width, intensity rates, slopes, areas, and intensity differences—with demographic attributes (gender, age, height, weight, and BMI) to refine predictive accuracy. This eliminates the dependency on idealized PPG morphologies, a fundamental limitation of conventional approaches. To enhance estimation fidelity, we developed a long short-term memory (LSTM) model trained on photoplethysmogram (PPG) pulse waveforms. Evaluated on the dataset, our model achieved a root mean square error (RMSE) of 20.75 mmHg for systolic blood pressure (SBP) and 10.76 mmHg for diastolic blood pressure (DBP). The model demonstrated moderate accuracy, with 21.48% of SBP predictions and 45.56% of DBP predictions falling within ±5 mmHg, and 43.70% of SBP predictions and 69.26% of DBP predictions within ±10 mmHg. These results suggest there is scope for further refinement to reach clinical-grade performance. This research contributes valuable insights toward improving AI-driven, non-invasive cardiovascular monitoring systems.