A Differential Evolution-Based Optimized Ensemble for Balanced and Imbalanced Medical Datasets.

Area of Science:

• Machine Learning
• Medical Informatics
• Computational Biology

Background:

• Class imbalance is a significant challenge in medical datasets, often hindering the accurate identification of minority classes, such as disease-positive cases.
• Traditional classifiers exhibit bias towards majority classes, leading to reduced detection rates for critical minority instances and decreased reliability in medical predictions.

Purpose of the Study:

• To introduce a novel ensemble learning framework, Optimized Ensemble by Differential Evolution (OEDE), designed to address class imbalance in medical datasets.
• To enhance the detection of high-risk or disease-positive instances within imbalanced medical data.

Main Methods:

• OEDE integrates three distinct base learners: Logistic Regression, Random Forest, and XGBoost, employing class-balancing techniques during training.
• Differential Evolution (DE) is utilized to optimize ensemble weights, maximizing the Area Under the ROC Curve (AUC) on a validation set.

Main Results:

• OEDE demonstrated significant performance improvements on imbalanced medical datasets, achieving a 70.08% AUC on the Thoracic dataset, outperforming baselines by over 19%.
• The framework reached a peak AUC of 97.89% on the Cervical Cancer dataset, consistently achieving competitive or superior AUC, F1-score, and Recall compared to traditional models.
• ROC curve analysis confirmed OEDE's enhanced discriminative capabilities in identifying minority classes.

Conclusions:

• The OEDE framework effectively enhances minority class detection in imbalanced medical datasets.
• Its robust and adaptable design positions OEDE as a valuable tool for healthcare risk prediction, particularly for identifying at-risk patient groups.