Implementation of CNN-SVM with Index Pattern-Based Feature Selection on PPG Signals for Cuffless Hypertension Detection

<p>Hypertension is one of the leading causes of death worldwide and often goes undetected due to its minimal symptoms. Early detection is crucial, and one non-invasive method involves the use of photoplethysmogram (PPG) signals. However, PPG signals contain a large number of features, which can lead to information redundancy and decreased model performance. This study proposes a hypertension detection system based on a CNN-SVM combination, preceded by feature selection using position-based indices (odd, even, specific multiples) to reduce data dimensionality and accelerate computation. The PPG signal dataset was obtained from 216 patients at UNS Hospital. After preprocessing and feature selection, feature extraction was performed using a Convolutional Neural Network (CNN), followed by classification using a Support Vector Machine (SVM). The model was evaluated under three classification scenarios: binary classification (normal vs. prehypertensive-hypertension and normal-prehypertension vs. hypertension) and three-class classification (normal, prehypertension, hypertension). The best classification accuracy achieved was 93.10% for the normal vs. prehypertension-hypertension scenario, 88.38% for normal-prehypertension vs. hypertension, and 82.79% for the three-class classification. This approach demonstrates that the combination of CNN-SVM with simple feature selection can improve both accuracy and efficiency in PPG-based hypertension detection.</p>