Artificial Intelligence-Based Optimal Classifier for the Prediction of Heart Ailment

Technical Summary: Artificial Intelligence-Based Optimal Classifier for the Prediction of Heart Ailment

Problem Statement
The prevalence of cardiac diseases is rising globally, with heart conditions responsible for approximately 30% of deaths worldwide. In India alone, over 30 million individuals suffer from heart conditions, placing significant strain on medical infrastructure, particularly in rural areas with limited diagnostic facilities. While various features contribute to cardiac abnormalities, identifying the most useful subset for accurate classification remains a challenge. The complexity of selecting optimal features and the difficulty in choosing the right artificial intelligence (AI) categorization model among numerous available options hinder the development of high-performing predictive systems.

Methodology
The study proposes a two-phase approach to predict cardiac ailments using the Cleveland Heart Disease Database from the UCI Machine Learning Repository.

1. Feature Selection via Statistical Analysis:
   The researchers utilized the Statistical Package for the Social Sciences (SPSS) to perform statistical analysis on the dataset. Specifically, independent sample two-tailed T-tests were conducted to determine the statistical significance of the features. Features with a p-value less than 0.05 were deemed significant. This process reduced the original set of variables to eight highly effective features: age, gender, chest pain type (angina), resting blood pressure, cholesterol, fasting blood sugar, resting ECG status, maximum heart rate, exercise-induced angina, ST segment during peak exercise, ST segment slope, number of major vessels, and type of defect.

2. Classifier Evaluation:
   The study employed MATLAB 2021a's Classification Learner application to evaluate the performance of various machine learning algorithms. A 10-fold cross-validation scheme was used to prevent overfitting. The researchers tested a comprehensive suite of 30 classifiers, including:

   • Decision Trees: Fine, Medium, and Coarse trees.
   • Discriminant Analysis: Linear and Quadratic.
   • Logistic Regression.
   • Naïve Bayes: Gaussian, Kernel, and others.
   • Support Vector Machines (SVM): Linear, Quadratic, Cubic, Fine, Medium, and Coarse kernels.
   • K-Nearest Neighbors (KNN): Fine, Medium, Coarse, Cosine, Cubic, and Weighted.
   • Ensemble Methods: Boosted Trees, Bagged Trees, Subspace Discriminant, and RUSBoosted Trees.
   • Neural Networks: Narrow, Medium, Wide, Bi-layered, and Tri-layered.

   The performance was compared using two feature sets: the full set of 13 commonly used features and the reduced set of 8 statistically significant features.

Key Contributions

• Statistically Driven Feature Reduction: Unlike studies relying on heuristic or automatic feature selection, this work utilizes T-tests to rigorously identify eight statistically significant features, ensuring a strong statistical basis for the model inputs.
• Comprehensive Comparative Analysis: The study evaluates 30 distinct classifiers, offering a broader comparison than many prior works that limit their scope to a few algorithms.
• Novelty in Kernel Selection: The research highlights the application of a Cubic SVM kernel for heart disease prediction, an approach not extensively investigated in previous literature on this specific dataset.

Results
The experimental results demonstrated that the reduced feature set (8 features) generally outperformed the full set (13 features) across most classifiers, indicating that the T-test selection effectively removed noise and reduced overfitting.

• Optimal Performance: The Cubic SVM classifier achieved the highest performance metrics among all tested models.
  • Accuracy: 90.28%
  • Area Under the Curve (AUC): 93%
  • Sensitivity (Se): 90.00%
  • Specificity (Sp): 90.63%
• Comparison with Other Models: While other models like Linear Discriminant (83.33% accuracy) and Quadratic SVM (83.33% accuracy) performed well, the Cubic SVM significantly outperformed them. The study notes that the Cubic kernel's ability to model complex, non-linear decision boundaries was crucial for capturing intricate relationships between the cardiac features.
• Benchmarking: When compared to prior studies (e.g., Kanti et al., Ahmad et al., Xu et al.), the proposed system achieved a higher accuracy (90.28%) and AUC (93%) than most existing approaches, which typically reported accuracies between 77% and 89%.

Significance and Claims
The paper claims that the combination of statistically significant feature selection (8 features) and the Cubic SVM classifier provides an optimal solution for predicting heart ailments. The authors assert that this approach offers:

• Enhanced Clinical Interpretability: By selecting features based on statistical significance (p < 0.05), the model retains clinical relevance and interpretability.
• Superior Predictive Power: The 93% AUC value indicates a highly effective classifier, suggesting that the Cubic kernel is better suited for the non-linear nature of cardiac disease data than the linear or Gaussian kernels commonly used in prior research.
• Robustness: The use of 10-fold cross-validation and the reduction of feature dimensionality helped mitigate overfitting, particularly important given the limited number of training samples in the dataset.

The study concludes that careful selection of these eight features, when applied to a Cubic SVM classifier, yields a high-performing predictive model for cardiac diseases, offering a valuable tool for physicians, especially in resource-constrained environments.