Protocol for a Scoped Systematic Review and Meta-Regression of Resampling Methods in Imbalanced Medical Datasets: Data-Level and Algorithm-Level Strategies in Clinical Prediction Models

Abstract

eview question / Objective Review Question (PICOTS) Population (P): Clinical datasets with binary outcomes where the minority class constitutes less than 30% of the total observations.Intervention (I): Resampling strategies including: Oversampling (e.g., SMOTE, ROS) Undersampling (e.g., RUS, NearMiss) Hybrid approaches (e.g., SMOTE+ENN) Algorithm-level strategies (e.g., cost-sensitive learning, focal loss) Comparator (C): Models trained on original (imbalanced) data without any correction Logistic regression models as reference classifiers Alternative balancing methods Outcomes (O): Discriminative performance (AUC, sensitivity, specificity, F1-score, accuracy) Calibration (Brier score, calibration slope) Reported misclassification costs (if applicable) Timing (T): Studies published from January 2009 to December 31, 2024.Setting (S): Clinical or healthcare prediction settings; primary research using retrospective or prospective study designs.Rationale Clinical prediction models are increasingly applied across various healthcare domains to support diagnosis, prognosis, and treatment decisions.However, a persistent methodological challenge in this field is class imbalance, where the number of observations in one outcome class (e.g., disease-positive cases) is much smaller than the other.This imbalance often leads to biased model performance, typically favouring the majority class and reducing the model's ability to detect minority events, frequently the most clinically important ones (e.g., rare adverse outcomes, disease onset).To address this, a wide array of resampling techniques has been developed.Data-level approaches such as oversampling (e.g., SMOTE), INPLASY

Ähnliche Arbeiten