Performance discrepancy mitigation in heart disease prediction for multisensory inter-datasets.

Area of Science:

• Cardiology
• Computer Science
• Biomedical Informatics

Background:

• Heart disease is a leading cause of global mortality, necessitating improved early prediction methods.
• Existing machine learning (ML) models for heart disease often perform poorly on unseen datasets due to the inter-dataset discrepancy problem.
• This discrepancy arises when models trained on one dataset are tested on another, distinct dataset.

Purpose of the Study:

• To mitigate the inter-dataset discrepancy problem in heart disease prediction using ML.
• To systematically evaluate the impact of preprocessing techniques on model performance across multiple datasets.
• To identify optimal strategies for feature selection and classification in inter-dataset scenarios.

Main Methods:

• Utilized five distinct heart disease datasets, exploring all training and testing combinations.
• Implemented a comprehensive preprocessing pipeline including SMOTE-Tomek for imbalance handling, Random Forest (RF) for feature selection, and Principal Component Analysis (PCA) for feature extraction.
• Incorporated missing value imputation (RF regression), log transformation, outlier removal, normalization, and data balancing.
• Evaluated eight different classifiers: Support Vector Machine, K-Nearest Neighbors, Decision Tree, RF, eXtreme Gradient Boosting, Gaussian Naive Bayes, Logistic Regression, and Multilayer Perceptron.

Main Results:

• Random Forest (RF) demonstrated superior performance for both feature selection and classification in both intra- and inter-dataset setups.
• RF achieved up to 100% accuracy in certain configurations and 96% accuracy during inter-dataset feature selection.
• The proposed preprocessing pipeline significantly improved ML model performance, reducing inter-dataset discrepancies without complex model development.

Conclusions:

• Effective data preprocessing is crucial for overcoming inter-dataset discrepancies in heart disease prediction models.
• Random Forest is a highly effective tool for feature selection and classification in this domain.
• Addressing inter-dataset discrepancies opens avenues for creating robust, generalizable ML models by combining diverse datasets.