Improved Hypertension Detection Models Utilizing Pulse Rate Variability and Asymmetry

Pulse-rate variability (PRV) offers a compelling alternative for estimating blood pressure (BP). However, ongoing debates persist regarding its suitability for BP monitoring and resemblance to HRV, with room for improvement in existing PRV studies We recruited and classified five-minute electrocardiography (ECG) and photoplethysmography (PPG) recordings from 202 patients acquired from the MIMIC-II database into three categories: normotensive (NT), prehypertensive (PHT) and hypertensive (HT). We conducted PRV and pulse rate asymmetry (PRA) analyses using various time-domain, frequency- domain and non-linear indices. We performed three database splits (1. NT/PHT/HT, 2. NT/non-NT and 3. HT/non-HT) to test our models. To validate our findings, we compared them to heart rate variability (HRV) using Bland-Altman (BA) analysis and correlation. We employed multi-class (MCC) and single- class classification with a 10-fold cross-validation approach, reserving a 20% test set. Correlation results indicated high concordance (ρ ≥ 0.79, ρ < 0.05) across all features, except for LF/HF and SD1/SD2. The BA analysis also revealed a strong agreement between most PRV and HRV features (confidence interval > 90%, BA ratio < 10%). Our classification accuracy for MCC was notably high, reaching up to 95% for the three-class problem, 90% for NT/non-NT classification, and 93% for HT/non- HT classification, achieved using 6-, 7-, and 5-feature models, respectively. This study highlights the reliability of PRV in monitoring BP. Additionally, the incorporation of PRA into PRV analysis yielded significant improvements in results, surpassing previous studies that employed PRV for BP estimation within the same database.