An AI-Driven Pipeline for High-Dimensional Gene Expression Analysis in Leukemia Classification (ALL vs AML)
TL;DR
The findings demonstrate that integrating statistical feature selection, supervised dimensionality reduction and regularized machine learning can provide highly accurate and interpretable models for leukemia classification, highlighting the growing role of artificial intelligence in precision oncology and public health informatics.
Abstract
High-dimensional gene expression datasets present major analytical challenges in biomedical research because the number of variables greatly exceeds the number of available samples. This study proposes an artificial intelligence–driven analytical pipeline for the classification of Acute Lymphoblastic Leukemia (ALL) and Acute Myeloid Leukemia (AML) using microarray gene expression data from the Golub leukemia dataset. The aim of the study is to develop a robust, interpretable and leakage-free machine learning framework capable of supporting precision medicine and clinical decision-making. The dataset included 7,129 genes and 72 patient samples (47 ALL and 25 AML cases). An 80/20 train–test split was applied while preserving class proportions. Initially, variance filtering retained the 5,000 most informative genes from the training set. Differential expression analysis was then performed using the limma framework with Benjamini–Hochberg False Discovery Rate correction, identifying 734 statistically significant genes (FDR ≤ 0.05). Subsequently, supervised Principal Component Analysis was conducted on the selected genes, with the first principal component explaining approximately 41% of the total variance. The resulting components were used as inputs for ridge logistic regression with internal cross-validation. The proposed pipeline achieved excellent classification performance, with test Accuracy = 1.00, Sensitivity = 1.00, Specificity = 1.00, and AUC = 1.00, while 5-fold cross-validation produced a mean AUC of 0.983. Furthermore, permutation testing generated a mean AUC close to 0.50, confirming that the observed performance was not due to random chance or data leakage. Overall, the findings demonstrate that integrating statistical feature selection, supervised dimensionality reduction and regularized machine learning can provide highly accurate and interpretable models for leukemia classification, highlighting the growing role of artificial intelligence in precision oncology and public health informatics.
