Leukemia is a blood cancer caused by the uncontrolled growth of abnormal white blood cells that crowd out healthy cells in the bone marrow and blood. It is divided into four main types: acute myeloid leukemia (AML), acute lymphocytic leukemia, chronic myelogenous leukemia, and chronic lymphocytic leukemia — each requiring different treatments. Identifying the correct subtype is therefore essential for giving patients the right therapy.
Traditional leukemia diagnosis relies on manual microscopic inspection of blood and bone marrow samples, blood cell counts, and expert interpretation. These methods are time-consuming, require highly skilled personnel, and are subject to variability between observers. As leukemia care improves, there is a growing need for automated, accurate, and fast diagnostic systems that do not depend on expert availability.
Microarray technology allows researchers to measure the activity level — or expression — of thousands of genes simultaneously across many samples. For leukemia, gene expression patterns differ between subtypes in ways that reflect their underlying biology. In principle, this data should allow machine learning algorithms to classify leukemia subtypes accurately, but the data presents a major technical challenge: each sample has 22,283 gene measurements but there are only a small number of patient samples to learn from. This imbalance, where there are far more variables than samples, is called the curse of dimensionality.
This 2025 PLOS ONE study addressed this challenge directly by proposing a feature selection approach to reduce 22,283 gene measurements down to only the 25 most discriminating ones, and then testing multiple machine learning and deep learning models on this reduced dataset to classify five leukemia-related cell types.
The study used the CuMiDa GSE9476 dataset — a curated, publicly available leukemia gene expression dataset that has been extensively cross-checked against thousands of published genomics articles to ensure data quality. It contains 64 samples representing five classes: AML, Peripheral Blood (PB), Peripheral Blood Stem Cells CD34 (PBSC_CD34), Bone Marrow, and Bone Marrow CD34. Each sample has measurements for 22,283 genes.
The researchers developed a custom feature selection algorithm based on maximizing class separability — essentially asking which genes have the most different expression patterns between the five leukemia classes. The algorithm computes a separability score for each gene by measuring how much the gene's values differ between classes while remaining consistent within each class. The top 25 genes by this score were selected for all subsequent classification work.
The data was split with 60% for training and 40% for testing, and pre-processed by normalizing gene expression values to zero mean and unit variance. Five classification methods were then applied to the 25-gene reduced dataset: Random Forest (RF), Linear Regression (LR), Decision Tree (DT), Support Vector Machine (SVM) with PCA, and Long Short-Term Memory (LSTM) deep learning networks.
LSTM (Long Short-Term Memory) is a type of deep learning architecture originally designed for sequential data — it contains memory cells with input, output, and forget gates that allow it to selectively remember or discard information across long sequences. Though typically used for time-series data, LSTM's ability to capture complex long-range dependencies in sequential gene feature vectors makes it powerful for high-dimensional biological classification tasks.
The results showed a clear performance hierarchy among the five classifiers. Linear Regression and Random Forest achieved 92.3% accuracy, with precision, recall, and F1-score all around 0.93. These are respectable results, demonstrating that even relatively simple machine learning approaches can classify leukemia subtypes well when given the right 25 genes.
Support Vector Machine (SVM) outperformed both, achieving 96.15% accuracy, 97.11% precision, and an F1-score of 0.9634. SVM finds the mathematical boundary that best separates different classes in high-dimensional space — even after dimensionality reduction to 25 features, SVM's boundary-finding approach proved more effective than the ensemble or regression-based methods.
The LSTM deep learning model, trained over 100 epochs, achieved perfect scores across all metrics: 100% accuracy, 100% precision, 100% recall, and a perfect F1-score of 1.0. This means the LSTM model correctly classified every single test sample into its correct leukemia subtype without a single error. The confusion matrix showed no misclassifications whatsoever.
5-fold cross-validation — where the data is split into five subsets and each is used as a test set in turn — confirmed the robustness of the results. The LSTM achieved 100% cross-validation accuracy, while SVM achieved 96%, Random Forest 96.14%, and Linear Regression 89.64%. The consistency between single-split and cross-validated results for LSTM is particularly impressive given the small sample size of 64 patients.
The key insight of this paper is that the performance advantage comes primarily from combining effective feature selection with deep learning. When all 22,283 genes are included, even powerful algorithms struggle: there is simply too much noise and redundant information for the classifier to identify the truly informative patterns. By reducing to 25 genes with high class-separability, the authors removed noise while retaining the most biologically meaningful signal.
LSTM's perfect performance suggests it is particularly well-suited to capturing complex, nonlinear interactions among the 25 selected genes that simpler models like linear regression cannot represent. LSTM's memory mechanisms allow it to model how combinations of gene expression levels — not just individual genes — determine leukemia subtype. This kind of higher-order pattern recognition is exactly what clinical leukemia subtyping requires.
Compared to earlier published studies on the same GSE9476 dataset, the authors' approach outperformed or matched all competing methods. Previous best results used SVM with 10 features (94% accuracy) or ensemble models with 9 features (94%). A 2023 study using 25 features and linear programming achieved 98.44%. This paper's LSTM with 25 features achieved 100%, representing a new state-of-the-art on this benchmark dataset.
The authors acknowledge that the small dataset size (64 samples) is a limitation and that perfect 100% accuracy on such a small test set may not fully generalize to larger, more diverse patient populations. However, the consistency of results across multiple classifiers and cross-validation folds supports the validity of the feature selection approach as a methodology even if LSTM's perfect accuracy needs confirmation on larger cohorts.
This 2025 study demonstrates that automated leukemia subtype classification from gene expression data is achievable with high accuracy using a relatively small number of carefully selected genes and modern deep learning methods. The ability to correctly classify all five leukemia cell types from just 25 gene measurements has clear implications for developing faster, cheaper, and more accessible diagnostic tools.
If validated on larger and more diverse patient cohorts, this approach could reduce dependence on manual bone marrow examination and expert pathology interpretation — making early, accurate leukemia subtyping available in settings that currently lack the specialized resources for conventional diagnosis. The feature selection algorithm itself is computationally lightweight and could be applied quickly in a clinical informatics pipeline.
The authors also suggest that the methodology is generalizable beyond leukemia — the same feature selection approach and LSTM classification framework could in principle be applied to other cancers with high-dimensional microarray or sequencing datasets. The fundamental challenge of identifying a small number of highly discriminating features from thousands of genetic measurements is shared across nearly all cancer genomics classification problems.