Machine learning sharpens leukemia diagnosis but waits on rigorous validation

Acute myeloid leukemia, an aggressive cancer of the blood and bone marrow, has become a testing ground for computational medicine. A literature review published in Clinical and Experimental Medicine examines how bioinformatics and machine learning are being applied across the disease, from molecular diagnosis to prognostic modeling and treatment planning. The review, led by Atefeh Moghassem and colleagues, draws on publications indexed in PubMed and Google Scholar from January 2015 through 2025, focusing on studies that used computational methods for diagnosis, prognosis, biomarker discovery, or treatment-related applications in AML.

The authors describe AML as a hematologic cancer marked by diverse molecular and clinical features that shape diagnosis, prognosis, and response to therapy. Despite advances in molecular classification and therapeutic strategies, they note that accurate risk assessment, treatment optimization, and the clinical translation of emerging biomarkers remain difficult. This heterogeneity is precisely where computational approaches have found traction, because algorithms can integrate patterns across large datasets that are difficult to synthesize manually.

According to the review, current evidence shows that conventional machine learning algorithms, deep learning methods, and multi-omics analyses have facilitated the identification of molecular signatures, improved disease classification, enhanced risk stratification, and supported prediction of treatment-related outcomes. Multi-omics integration, which combines genomic, transcriptomic, and other molecular data layers, allows models to capture the biological complexity that underlies variable patient outcomes. Deep learning architectures, in turn, can learn hierarchical representations from high-dimensional data such as gene expression profiles, potentially revealing signatures that simpler statistical methods miss.

These capabilities matter clinically because AML treatment decisions have long depended on risk stratification. If computational models can refine which patients belong to which risk category, they could in principle support more precise therapeutic selection. The review frames this as an area of active progress: computational tools have moved from exploratory analyses toward contributions in diagnosis, prognostic prediction, and therapeutic decision-making, aligning computational outputs with the molecular classification frameworks that now guide AML care.

Yet the review is notably cautious about clinical readiness. The authors state that the clinical implementation of AML machine-learning models remains limited due to methodological challenges. They list retrospective study designs, batch effects, overfitting, insufficient external validation, data leakage, and limited assessment of clinical utility as recurring problems. Each of these threatens the reliability of published models. Retrospective designs mean most models are trained and tested on historical data rather than evaluated prospectively in clinical workflows. Batch effects, technical variation introduced during sample processing, can cause algorithms to learn artifacts of laboratory practice rather than biology.

Overfitting and data leakage compound the problem. Overfitting occurs when a model memorizes idiosyncrasies of its training data and fails to generalize to new patients. Data leakage, where information from the test set inadvertently influences training, can inflate reported performance metrics, making a model appear more accurate than it truly is. Insufficient external validation, testing on independent cohorts from other institutions, means many published models have not demonstrated that their performance holds beyond the data on which they were built. Limited assessment of clinical utility means that even models with strong statistical performance have rarely been shown to improve actual patient management.

The authors conclude that bioinformatics and machine learning represent powerful tools for advancing AML research and precision medicine, but that rigorous validation, transparent reporting, and prospective clinical evaluation are required before these approaches can be integrated into routine AML management. This positions the field at a familiar inflection point: the analytical methods exist and have produced promising results, but the evidentiary standards needed for clinical adoption have not yet been met at scale.

The review’s emphasis on transparent reporting suggests that the path forward involves not only better algorithms but also better study design. Prospective evaluation would test whether models trained on retrospective cohorts can perform in real-time clinical settings. External validation across institutions would establish generalizability. Until those steps are taken, the review indicates, machine learning in AML remains a research enterprise rather than a clinical standard, its contributions confined to discovery and hypothesis generation while routine management continues to rely on established molecular classification and clinical assessment.

Subject of Research: Applications of bioinformatics and machine learning in acute myeloid leukemia diagnosis, prognosis, and treatment

Article Title: Bioinformatics and machine learning in acute myeloid leukemia: advancing diagnosis, outcome prediction, and therapeutic strategies – a literature review

Article References: Moghassem, A., Khademi, M., Mohajerani, N., Ahmadi, A., Aldaghi, T., Gharehbaghian, A., & Nazari, E. (2026). Bioinformatics and machine learning in acute myeloid leukemia: advancing diagnosis, outcome prediction, and therapeutic strategies – a literature review. Clinical and Experimental Medicine. https://doi.org/10.1007/s10238-026-02371-5

Image Credits: AI Generated

DOI: 10.1007/s10238-026-02371-5

Keywords: acute myeloid leukemia, machine learning, bioinformatics, multi-omics, risk stratification, biomarker discovery, deep learning, external validation, precision medicine, prognosis, machine, learning