Machine Learning Techniques for the Performance Enhancement of Multiple Classifiers in the Detection of Cardiovascular Disease from PPG Signals

Photoplethysmography (PPG) is a noninvasive, low-cost modality used in clinical diagnostics. This study applies machine learning to improve classifier performance for detecting cardiovascular disease (CVD) from PPG signals. Data from 41 CapnoBase subjects (20 CVD, 21 normal) were used. Signals were segmented into 1‑s segments, yielding 720 segments per patient and 29,520 total segments. To address high dimensionality, five dimensionality reduction (DR) techniques—three heuristic (ABC‑PSO, cuckoo search, dragonfly) and two transformation-based (Hilbert transform, nonlinear regression)—were applied. Twelve classifiers were evaluated: PCA, EM, logistic regression, GMM, Bayesian LDA, firefly, harmonic search, DFA, PAC‑Bayesian, KNN‑PAC‑Bayesian, softmax discriminant classifier (SDC), and detrend with SDC. Performance was assessed via accuracy, performance index, error rate, and good detection rate (GDR). The combination of Hilbert transform with the harmonic search classifier achieved the best performance with 98.31% accuracy and 96.55% GDR.