Non-Invasive Blood Pressure Estimation Using Photoplethysmogram Morphological Features and Machine Learning

Abstract

Continuous and unobtrusive blood pressure monitoring is vital for the early diagnosis and effective management of cardiovascular diseases. Conventional pneumatic cuffs cause physical discomfort during repeated measurements, while invasive catheterization poses substantial clinical risks. To overcome these limitations, this work proposes a non-invasive cuffless blood pressure estimation framework based on single-channel photoplethysmogram signals.  An automated signal processing pipeline segments individual pulse cycles and extracts 23 candidates morphological, area and derivative features. Then, screening by Pearson correlation identifies an optimal subset of twelve salient parameters. A strict subject-wise five-fold cross-validation scheme is used to systematically evaluate four representative machine learning regression models. The simulated data set contains 2,500 beat-to-beat samples, simulated according to benchmark hemodynamic distributions. XGBoost achieved the best performance in predicting systolic pressure, with MAE of 2.49 mmHg and correlation coefficient of 0.97. Random Forest performs slightly better for diastolic pressure (MAE = 4.21 mmHg) with a small difference with XGBoost (MAE = 4.28 mmHg, r=0.74). The simulated conditions’ numerical errors meet the benchmark thresholds of the ANSI/AAMI/ISO standard and the British Hypertension Society Grade A protocol. This does not mean clinical validation. These results validate the feasibility of lightweight, interpretable machine learning regression on wearable healthcare hardware for continuous cardiovascular monitoring.