Small, imbalanced, and noisy datasets in healthcare research: A systematic review

Background: The proper application of algorithms to health data can aid patient care and support medical decisions, yet its reliable deployment is contingent on data quality. In practice, clinical datasets are structurally characterised by class imbalance, limited sample size, and noise. When not properly managed, it can result in biased predictions towards the majority class Objective: This review provides a comprehensive overview of data strategies, algorithm techniques and evaluation metrics that handle imperfect datasets. Methods: Articles were retrieved from four databases, and a total of one hundred and twenty-six articles were included and systematically analysed. Results: Initial statistics highlight the recent concern to better manage imbalanced data for reliable predictions of the minority class. At a data level, synthetic minority oversampling technique (SMOTE)-based techniques showcased higher prevalence. Despite the reduced dataset size, deep learning models were largely implemented, raising generalisability concerns. Accuracy was the most frequently reported metric for model evaluation despite its well-documented inadequacy under skewed class distributions. To consolidate current practices, this review proposes the interconnected layers of adaptivity, complexity & explainability (InterLACE) framework, a decision flow that integrates dataset characterisation with the selection of data and model-level interventions best suited to mitigate data complexity challenges. Conclusion: This review compiles widely used approaches for managing classification complexity and proposes the InterLACE framework as a structured decision system that promotes more transparent model development, emphasising data quality and information integrity, particularly in the clinical use of personal and biosignal data, where interpretable artificial intelligence (AI) models are essential.