FastPedia-ML: An Interpretable Machine-Learning Framework for Pediatric Leukemia Subtype Classification using Gene-Expression Data

Abstract

Pediatric Acute Myeloid Leukemia (pAML) is a heterogeneous disease with complicated genomic variants that make it difficult to subclassify the disease properly. The paper suggests a strong and explainable machine learning model applied to the classification of pediatric leukemia subtypes based on high-dimensional data in microarray gene expression. The framework combines ANOVA-based feature selection and variance-based filtering to minimize dimensionality, as well as adaptive SMOTE in order to deal with the imbalance of classes. The strategy of cross-validation is used in a nested way to guarantee the unbiased model evaluation and hyperparameters optimum. Three classifiers, which are Random Forest (RF), Support Vector Machine (SVM), and XGBoost are compared in the terms of weighted F1-score, MCC, and ROC-AUC. The experimental results on the GSE9476 dataset indicate that RF and SVM can be used to obtain perfect classification performance (F1-score = 1.000), whereas XGBoost can be used to obtain competitive results (F1 = 0.919). Statistical significance (p = 0.001) is proven by permutation testing. SHAP-based analysis also determines the biologically significant genes that correlate with the development of leukemia. The suggested framework has a high predictive power, robustness, and interpretability, which shows the possibility of using it in the context of precision medicine to diagnose pediatric leukemia.