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

Cardiovascular disease (CVD) is a leading cause of death globally, highlighting the need for effective and accessible diagnostic tools. Photoplethysmography (PPG), a non-invasive and inexpensive technique for measuring blood volume changes, offers a promising approach for CVD detection. PPG signals, however, are often noisy and high-dimensional, requiring sophisticated signal processing and machine learning techniques for accurate classification. This study addresses the challenge of improving CVD detection accuracy from PPG signals by combining several dimensionality reduction techniques with a diverse set of classifiers. The research aims to evaluate the performance of various machine learning models in classifying PPG signals as originating from individuals with or without CVD, ultimately contributing to the development of more effective and efficient diagnostic tools.

Existing research on PPG signal analysis for CVD detection has explored various approaches. Studies have utilized parameters such as independent component analysis (ICA), principal component analysis (PCA), entropy, and mutual information (MI) for feature extraction. Methods for removing signal artifacts, like those caused by motion or skin conditions, have also been developed, employing techniques such as waveform morphology and adaptive filters. Different classifiers have been applied to extracted features, including support vector machines (SVM), artificial neural networks (ANN), and K-nearest neighbors (KNN). However, there is ongoing exploration of techniques to improve classification accuracy and computational efficiency. Previous research has reported varying success rates depending on feature extraction methods, classifier selection, and the dataset used. This paper builds upon this existing work by incorporating multiple dimensionality reduction and classification techniques to potentially improve the accuracy of CVD detection from PPG signals.

This study employed a two-stage process for CVD detection from PPG signals obtained from the CapnoBase database. The database comprises PPG recordings from 41 subjects (20 with CVD, 21 normal), sampled at 300 Hz. Each recording was segmented into one-second epochs. **Stage 1: Dimensionality Reduction:** Five dimensionality reduction techniques were used: (1) Hilbert Transform (HT), (2) Nonlinear Regression (NLR), (3) Artificial Bee Colony-Particle Swarm Optimization (ABC-PSO), (4) Cuckoo Search, and (5) Dragonfly Algorithm. These methods aimed to reduce the dimensionality of the PPG data while preserving relevant information for classification. **Stage 2: Classification:** Twelve classifiers were applied to the dimensionally reduced data: (1) Principal Component Analysis (PCA), (2) Expectation Maximization (EM), (3) Logistic Regression, (4) Gaussian Mixture Model (GMM), (5) Bayesian Linear Discriminant Analysis (BLDC), (6) Firefly Algorithm, (7) Harmonic Search, (8) Detrended Fluctuation Analysis (DFA), (9) PAC Bayesian Learning, (10) KNN-PAC Bayesian, (11) Softmax Discriminant Classifier (SDC), and (12) Detrended SDC. The performance of each classifier-dimensionality reduction combination was evaluated using a 90/10 train-test split and assessed using metrics including accuracy, performance index, sensitivity, specificity, good detection rate (GDR), and error rate. Optimal parameters for each classifier were determined using a 10-fold cross-validation and mean squared error (MSE) as a stopping criterion. Statistical analysis, including mean, variance, skewness, kurtosis, Pearson correlation coefficient (PCC), and sample entropy, was performed on the dimensionally reduced data to characterize its properties.

The study found that the combination of the Hilbert Transform (HT) dimensionality reduction technique and the Harmonic Search classifier yielded the highest accuracy (98.31%) and good detection rate (GDR) (96.55%). Other combinations also demonstrated substantial accuracy. For example, the Nonlinear Regression (NLR) method coupled with the Harmonic Search classifier achieved 97.79% accuracy. The Cuckoo Search with Harmonic Search reached 96.095% accuracy. The performance of other classifiers varied depending on the dimensionality reduction technique employed. The Harmonic Search classifier consistently performed well across various dimensionality reduction methods, suggesting its robustness for this application. The analysis of MSE values further supported the superiority of certain classifier-dimensionality reduction pairings. The computational complexity was also analyzed and varied based on the chosen method. A comparison with previous studies revealed that the proposed approach outperformed many existing methods in terms of accuracy. Detailed statistical parameters for each dimensionality reduction technique were presented, demonstrating the differences in mean, variance, skewness, kurtosis, PCC, and sample entropy between normal and CVD cases. Normal probability plots revealed the presence of nonlinearity and outliers in both classes after applying the dimensionality reduction methods, justifying the use of nonlinear classifiers.

The high accuracy achieved by the Hilbert Transform and Harmonic Search classifier combination suggests that the HT effectively captures relevant features for CVD detection, and the Harmonic Search classifier is well-suited to handle the characteristics of this data. The consistent high performance of the Harmonic Search across different dimensionality reduction methods indicates its robustness and adaptability. The varied performance of other classifiers highlights the importance of carefully considering both dimensionality reduction and classifier selection. The superior performance of the HT and Harmonic Search combination warrants further investigation into the underlying reasons for its effectiveness. It also opens avenues for improving existing CVD diagnostic approaches. The study's findings demonstrate the potential of using machine learning to enhance the accuracy and efficiency of CVD detection using readily available PPG technology.

This research demonstrates the potential of integrating multiple dimensionality reduction techniques with a variety of machine learning classifiers to enhance CVD detection accuracy from PPG signals. The Hilbert Transform coupled with the Harmonic Search classifier showed superior performance. Future work should focus on exploring other dimensionality reduction and classification methods, investigating larger and more diverse datasets, and evaluating the performance of deep learning models for this task. Furthermore, integrating this approach into wearable technology for continuous, real-time CVD monitoring is a promising avenue of future research.

The study's limitations include the relatively small sample size from a single database. The generalizability of the findings to other populations and datasets requires further investigation. The choice of specific parameters for each dimensionality reduction and classification technique might influence the results. Future research should address these limitations by using larger and more diverse datasets, conducting a more exhaustive parameter optimization, and exploring alternative machine learning approaches.