Resampling methods for class imbalance in clinical prediction models: A scoping review protocol.

Area of Science:

• Medical Artificial Intelligence (AI)
• Machine Learning in Healthcare
• Clinical Prediction Modeling

Background:

• Class imbalance in clinical datasets (<30% positive cases) hinders medical prediction model sensitivity and fairness.
• Existing data-level (e.g., SMOTE) and algorithm-level (e.g., cost-sensitive learning) strategies lack comprehensive empirical evidence for effectiveness.
• This scattered evidence across diseases and models creates a methodological gap in reliable medical AI.

Purpose of the Study:

• To conduct a scoping systematic review with meta-regression on resampling strategies for imbalanced clinical data.
• To map and quantitatively summarize 15 years of research (2009-2024) on these strategies.
• To address the methodological gap concerning the reliable improvement of medical AI performance.

Main Methods:

• Systematic search of major databases (MEDLINE, EMBASE, Scopus, Web of Science, IEEE Xplore) and pre-print servers.
• Inclusion of primary studies applying resampling or cost-sensitive strategies to binary clinical prediction tasks with <30% minority class prevalence.
• Descriptive synthesis and random-effects meta-regression (logit-transformed AUC) to analyze moderator effects (imbalance ratio, strategy, model, sample size).

Main Results:

• A comprehensive catalog of clinical domains, sample sizes, imbalance ratios, strategies, model types, and performance metrics (AUC) will be compiled.
• Meta-regression will quantify the impact of various factors on model performance.
• Analyses will assess small-study effects and robustness using established statistical methods.

Conclusions:

• Determine when data-level versus algorithm-level balancing genuinely improves discrimination, calibration, and cost-sensitive metrics.
• Provide evidence-based guidance for selecting imbalance-handling methods in medical AI research.
• Inform reporting standards and identify research gaps, particularly in calibration and misclassification costs, for trustworthy AI in clinical practice.