Heart Rate Variability Machine Learning Models to Facilitate Elevated Blood Pressure Detection

Heart-rate variability (HRV) has high potential in hypertension (HT) detection, which is crucial in prevention from mortal HT-related diseases. Nevertheless, the few studies based on HRV so far present significant gaps that do not allow their implementation on monitoring devices. The present work aims to create simple and reproducible HRV-based models to facilitate HT or elevated blood pressure (BP) detection. Five- minute electrocardiography (ECG) recordings of 202 patients from the MIMIC database were extracted. Recordings were classified according to the patients’ BP at the extraction time as normotensive (NT- systolic BP < 120), prehypertensive (PHT- SBP < 140) and HT (SBP 140), with PHT and HT being associated with significant health risks. Time-domain, frequency- domain and Poincare HRV features were extracted and compared among the different classes with non-parametric tests. Multi-feature models for each class as well as for HT and elevated BP detection were created with the optimizable Ensemble technique, with 10-fold cross-validation and an 80 20% train-test set. All HRV features except for LF/HF and SD1/SD2 showed statistically significant differences among the three groups (p < 0.0001) as well as between HT-vs-all (p 0.0001) and elevated BP-vs-NT groups (p 0.0485). The multi-feature classification accuracy was between 88 95%. For elevated BP detection, a model with just four features achieved 95% accuracy and 100%, 93.75% and 89% sensitivity, specificity and F1-score, respectively. The presented model can be implemented in real-time ECG recording devices for timely detection of elevated BP.