Comparative Analysis of Machine Learning Techniques for Imbalanced Genetic Data

Abstract

Advancements in genome sequencing technologies have significantly increased the availability of genomic data. The use of machine learning models to predict the pathogenicity or clinical significance of genetic mutations is crucial. However, genetic datasets often feature imbalanced target variables and high-cardinality, skewed predictor variables. These attributes complicate machine learning modeling processes. This study addresses these challenges in both regression and classification tasks. In this study, we systematically explored the impact of various data preprocessing techniques, feature selection methods, and model choices on the performance of machine learning models trained on imbalanced genetic data. We evaluated the performance metrics using fivefold cross-validation. Our key findings demonstrate that the regression models are robust to outliers and skew in predictor and target variables. Similarly, in classification tasks, class-imbalanced target variables and skewed predictors minimally impact model performance. Among the models tested, random forest was the most effective model for both imbalanced regression and classification tasks. Our key contributions are as follows: we address a significant research gap by focusing on imbalanced regression, a problem that is sparsely explored compared to class-imbalanced classification. We identify the techniques that improve prediction performance and provide practical insights into handling genetic data. Additionally, we provide a foundation for future research to further optimize machine learning approaches in genomics. This study uses a genetic dataset as a case, but our findings are applicable to imbalanced data in other fields.