The metric sensitivity analysis further highlighted that method ranking is strongly influenced by the selected evaluation criterion. In this context, PR–AUC emerged as the most informative and discriminative metric, offering a balanced assessment of minority-class detection performance compared to accuracy and ROC–AUC. Cross-dataset and cross-classifier consistency analyses revealed that only a subset of balancing strategies generalizes well across different transcriptomic cohorts and learning paradigms, underscoring the importance of robustness evaluation in biomedical machine learning studies. Overall, this work provides practical guidelines for selecting effective and robust balancing strategies for genomic autism classification. The results support hybrid and boundary-aware oversampling approaches as reliable default choices for imbalanced transcriptomic data. Future research may extend this framework to additional omics modalities, larger and more diverse cohorts such as RNA-seq datasets, and the integration of explainable artificial intelligence techniques to enhance biological interpretability and clinical relevance.

The analysis of metric sensitivity showed that the choice of evaluation criterion has a major impact on how different methods are ranked. PR-AUC proved to be the most useful and distinguishing metric in this regard, providing a more balanced evaluation of how well minority classes are detected when compared to accuracy and ROC-AUC. When examining consistency across different datasets and classifiers, we found that only certain balancing strategies perform reliably across various transcriptomic datasets and machine learning approaches. This finding emphasizes how crucial it is to evaluate robustness in biomedical machine learning research. This study offers practical recommendations for choosing effective and dependable balancing strategies when classifying genomic autism data. The findings indicate that hybrid and boundary-aware oversampling methods serve as trustworthy default options for handling imbalanced transcriptomic datasets. Future studies could expand this approach to include other omics data types, larger and more varied cohorts like RNA-seq datasets, and incorporate explainable AI methods to improve biological understanding and clinical applicability.