The Explanation and Sensitivity of AI Algorithms Supplied with Synthetic Medical Data

Abstract

The increasing complexity and importance of medical data in improving patient care, advancing research, and optimizing healthcare systems led to the proposal of this study, which presents a novel methodology by evaluating the sensitivity of artificial intelligence (AI) algorithms when provided with real data, synthetic data, a mix of both, and synthetic features. Two medical datasets, the Pima Indians Diabetes Database (PIDD) and the Breast Cancer Wisconsin Dataset (BCWD), were used, employing the Gaussian Copula Synthesizer (GCS) and the Synthetic Minority Oversampling Technique (SMOTE) to generate synthetic data. We classified the new datasets using fourteen machine learning (ML) models incorporated into PyCaret AutoML (Automated Machine Learning) and two deep neural networks, evaluating performance using accuracy (ACC), F1-score, Area Under the Curve (AUC), Matthews Correlation Coefficient (MCC), and Kappa metrics. Local Interpretable Model-agnostic Explanations (LIME) provided the explanation and justification for classification results. The quality and content of the medical data are very important; thus, when the classification of original data is unsatisfactory, a good recommendation is to create synthetic data with the SMOTE technique, where an accuracy of 0.924 is obtained, and supply the AI algorithms with a combination of original and synthetic data.

Topics & Concepts