Comparison of seven machine learning models in hypertension classification using photoplethysmographic and anthropometric data

This study presents an algorithm for classifying individuals into four hypertension categories (healthy, prehypertension, Stage 1, and Stage 2) using indices computed from photoplethysmographic (PPG) and anthropometric data. The dataset includes 219 individuals (115 women, 104 men, ages 21-86), with resting PPG signals, body mass index (BMI), age, weight, height, and resting heart rate. Key features (PPGAI, Ab, and Ad indices) were computed from the PPG signal. After dimensionality reduction through stepwise linear regression, the most informative predictors of hypertensive stages were identified for model training. Seven machine learning models, including Support Vector Machine (SVM), K-Nearest Neighbours, Logistic Regression, Random Forest, Naive Bayes, Linear Discriminant Analysis, and Quadratic Discriminant Analysis, were evaluated using leave-one-out cross-validation and the most accurate one was selected for final classification. The Linear SVM showed the best performance, correctly classifying 71.3%, 67.1%, 38.2%, and 55% of healthy, prehypertensive, Stage 1, and Stage 2 subjects, respectively. However, in a preliminary screening scenario aimed at prompting clinical follow-up for positive cases, the algorithm flagged 76.5% of prehypertensive, 97.1% of Stage 1, and 100% of Stage 2 individuals as belonging to one of the three hypertensive categories. Nonetheless, additional training data are needed to improve the model's accuracy.