Challenges and recent advances in methods for handling imbalanced multiclass classification problems: a methodological review

Abstract Background Class imbalance is a common challenge in real-world health science applications, including medical diagnosis, rare disease detection, and ICU mortality prediction, where one or more classes are underrepresented. Although several methods address imbalance in binary classification, multiclass imbalance remains particularly challenging due to multiple minority classes, often leading to biased performance and reduced predictive accuracy. Despite several advancements, most classification models struggle to identify patterns in imbalanced data, limiting their effectiveness in real-world applications. Methods A structured literature search was conducted to identify methodological studies on imbalanced multiclass classification, including algorithmic strategies and advances in performance evaluation. Articles published up to 2024 were retrieved from Scopus and Web of Science using predefined keywords. Studies were screened through titles and abstracts based on predefined inclusion and exclusion criteria, with additional backward citation searching for methodologically relevant studies. In total, 75 studies were included in the final methodological review to synthesize key challenges and recent advances. Results Despite the introduction of several metrics for assessing multiclass imbalance, the Imbalance Ratio (IR) remains the most commonly used measure for quantifying imbalance severity. Existing balancing techniques mainly rely on distance-based, cluster-based, and distribution-based approaches, reflecting methodological diversity. In multiclass settings, various decomposition strategies, classification algorithms, and performance metrics have been proposed to address imbalance; however, repeated use of imbalance-handling mechanisms, such as class weight adjustments across decomposition, training, and evaluation stages, may introduce bias. The effectiveness of these strategies depends on data characteristics including dimensionality, sample size, distribution, number of classes, and imbalance severity. Notably, insufficient reporting of these characteristics in many studies limits the assessment of feasibility and generalizability across diverse data settings. Conclusions This review synthesizes the strengths and limitations of existing methods for handling imbalanced multiclass classification, offering practical insights for improving model robustness and predictive performance. Effective management of class imbalance supports several Sustainable Development Goals by promoting equitable decision-making and enhancing reliable analysis across diverse health, societal, and environmental challenges, making it essential for developing robust and generalizable models across diverse domains.