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

The study addresses early and accurate detection of cardiovascular disease (CVD) using photoplethysmography (PPG), a widely accessible and noninvasive technique for monitoring blood volume changes. PPG-based assessment is suitable for wearable and remote monitoring but is susceptible to motion and skin artifacts, and PPG datasets can be high-dimensional. The research objective is to enhance detection performance of multiple classifiers by first reducing PPG data dimensionality and then evaluating diverse machine learning classifiers on the reduced features, aiming for high accuracy with low error rates. The work is motivated by the global burden of CVD and the potential for PPG to enable large-scale, real-time screening and continuous monitoring.

Prior studies have explored PPG for various cardiac applications, including artifact removal, heart rate variability analysis, coronary artery disease (CAD) detection, arrhythmia classification, and risk stratification. Methods span time/frequency domain features, higher-order statistics, SVD-based annotations, and deep neural networks. Reported accuracies vary: ANN (up to 94.7%), KNN (81.5%), ELM (up to ~90% sensitivity/specificity), SVM and GMM (up to ~99% in limited settings), softmax regression (~94.44%), SDC with SVD/statistics (~97.88%), DCNN (~85%), decision trees with HRV (~94.4%), and Naïve Bayes (up to 94.44%). Prior work also examined metaheuristic optimization for dimensionality reduction and classification. This study extends the literature by systematically comparing five DR methods with twelve classifiers on a unified PPG dataset and protocol, highlighting the synergy between Hilbert transform features and harmonic search classification.

Dataset: CapnoBase database comprising raw PPG recordings with varied morphologies and annotated respiratory signals. A total of 41 records (20 CVD, 21 normal) were analyzed. PPG was treated segment-wise: 1‑s segments, 720 per patient, totaling 29,520 segments. The paper notes sampling at 200 Hz in multiple sections (yielding 144,000 samples per patient) and also mentions 300 Hz with 200‑sample segments; the primary analysis is conducted on 1‑s segments with subsequent dimensionality reduction. Preprocessing: Noise/artifact removal using independent component analysis (ICA). Dimensionality Reduction (DR): Original per-patient matrix (200×720) reduced to 100×720 using five DR methods: (1) Hilbert transform (HT), (2) Nonlinear regression (NLR, Levenberg–Marquardt-based least squares), (3) ABC‑PSO (hybrid artificial bee colony and particle swarm optimization with defined fitness F = aφ_k + b(|T|−|K|)/|T|), (4) Cuckoo search with Lévy flights, and (5) Dragonfly algorithm (separation, alignment, cohesion, attraction, diversion dynamics). Statistical characterization: On DR outputs, extracted mean, variance, skewness, kurtosis, Pearson correlation coefficient (PCC), and sample entropy to assess class separability and nonlinearity. Classifiers: Twelve methods were evaluated: PCA, EM, logistic regression (threshold 0.5), GMM (EM-estimated parameters), Bayesian linear discriminant classifier (BLDC), firefly algorithm classifier, harmonic search (HS), detrended fluctuation analysis (DFA), PAC‑Bayesian, KNN‑PAC‑Bayesian, softmax discriminant classifier (SDC), and detrend with SDC. Classifier-specific parameters were tuned (e.g., HS class targets 0.85/0.1; PCA threshold 0.72; EM likelihood/convergence settings; SDC λ=0.5). Targets for classes were chosen with T_CVD=0.85 and T_Normal=0.1 based on normalized feature means. Training/testing: 90% training, 10% testing with 10‑fold training/testing procedure. Mean square error (MSE) used as stopping criterion with threshold 1e−5 or max 1000 iterations. Performance metrics from confusion matrices included sensitivity, specificity, accuracy, performance index (PI), good detection rate (GDR), and error rate. Computational complexity: Big‑O estimates reported per classifier/DR pairing to assess efficiency trade-offs.

- Best overall: Hilbert transform + harmonic search achieved 98.31% accuracy, PI 96.485%, GDR 96.55%, and lowest average error rate 3.38%. - Across DR methods, harmonic search consistently outperformed other classifiers; second-best accuracy 97.79% with NLR + harmonic search; third-best 96.095% with cuckoo search + harmonic search. - Under cuckoo search DR (example detailed table): harmonic search accuracy 96.095%, sensitivity 92.185%, specificity 100%, PI 91.29%, error rate 7.81%; GMM achieved sensitivity 100% but specificity 84.38% (accuracy 92.19%); firefly showed lowest performance (accuracy 78.275%, PI 20.76%, sensitivity 56.15%). - MSE analysis: For CVD cases, ABC‑PSO + DFA gave the minimum testing MSE of 4.00×10−8; for normal cases, Hilbert + harmonic search yielded 9.00×10−8. Maximum MSE observed was 6.60×10−4 for cuckoo search + PCA (CVD cases). - Robustness: Most classifiers (except logistic regression) maintained accuracies ≥80% across DR methods; DFA often ranked second in accuracy (e.g., 95.575% with ABC‑PSO). - Computational complexity: Lowest reported complexity O(m log m) for logistic regression and firefly under Hilbert DR; highest O(m^7) for KNN‑PAC‑Bayesian with ABC‑PSO. Some high‑complexity pairings (e.g., ABC‑PSO + DFA ~ O(m^5)) still achieved high accuracy (~95.58%).

Results indicate that dimensionality reduction is essential for handling high‑dimensional PPG data and improving classifier performance. The Hilbert transform enhances discriminative phase-related properties that align well with the harmonic search classifier’s pitch adjustment mechanism, enabling superior separation of CVD vs. normal segments. Statistical analysis of DR outputs showed low PCC and evidence of nonlinearity and outliers, supporting the use of nonlinear classifiers. Harmonic search demonstrated robust performance across all DR methods, while logistic regression and PAC‑Bayesian were comparatively weak, particularly under certain DR settings. Segment-based analysis (1‑s segments) produced strong performance metrics, highlighting feasibility for fine-grained screening. Given PPG’s accessibility and low hardware requirements, the approach shows promise for large-scale, real-time and remote CVD screening using wearables, with favorable trade-offs between accuracy and computational cost for selected method combinations.

The study demonstrates that combining transformation- or heuristic-based dimensionality reduction with appropriate classifiers markedly improves CVD detection from PPG segments. The best-performing pipeline—Hilbert transform with harmonic search—achieved 98.31% accuracy and 96.55% GDR. Nonlinear classifiers benefited from the DR-extracted features, and DFA consistently provided strong secondary performance. The findings support PPG-based, segment-level CVD screening suitable for continuous, real-time monitoring via wearable devices. Future work will focus on improving performance via heuristic hyperparameter optimization and exploring deep learning architectures (CNNs and deep neural networks) to automatically learn relevant features and further reduce detection latency.

- Classifier variability: PAC‑Bayesian learning and logistic regression did not achieve high accuracy across all DR methods, indicating limited robustness in this setting. - Segment length trade-offs: Very short (second-to-second) detection may increase false alarms; longer segments (e.g., 30 s to 1 min) can improve accuracy but risk overfitting and reduced generalizability. - Inconsistencies in sampling descriptions exist (200 Hz vs. 300 Hz mentioned), though the main analysis is based on 1‑s segmentation; beat-to-beat analysis was not performed. - Potential overfitting risk due to tuned parameters and k‑fold procedures; external validation on independent datasets was not reported.