Has Artificial Intelligence Arrived at Clinical Hematology? The Case of AML.

Artificial intelligence has arrived at the hematology clinic?

A new review in JAMA Oncology (Ansarian et al., 2025) surveys how AI systems are being woven into the management of acute myeloid leukemia (AML)—from image-based diagnosis to genomic risk modeling. 

The numbers are striking—up to AUROC 0.97 for AML detection and >99% accuracy for transcriptome classification.  But the authors also indicate that these tools are beginning to mimic clinical reasoning, by linking bone-marrow morphology with genetic signatures such as NPM1 status. 

The authors also highlight the growing role of federated learning, which lets hospitals train shared models without sharing patient data—an advance that could democratize access to high-performance AI even in resource-limited settings. Below is the abstract of Ansarian et al, positioning AML as a proving ground for AI in the cancer clinic.

### ABSTRACT

Acute myeloid leukemia (AML) is a severe hematologic cancer with complex genetic heterogeneity necessitating personalized treatment approaches. Artificial intelligence (AI) technologies may revolutionize risk stratification, diagnosis enhancement, and treatment planning in addressing critical gaps in AML management, particularly in low-resource health care environments.

Observations  This narrative review synthesizes existing AI applications in 3 primary areas of AML management. 

1. Machine learning algorithms integrating clinical, cytogenetic, and molecular data demonstrate greater prognostic accuracy than conventional European LeukemiaNet (ELN) guidelines. 
2. Deep learning approaches to image analysis yield excellent results for AML subtype identification from bone marrow smears (area under the receiver operating characteristic curve [AUROC]: 0.97) and genetic variant prediction (eg, NPM1 status [AUROC: 0.92]). 
3. AI-driven genomic analysis reveals novel prognostic signatures and therapeutic targets through advanced pattern recognition, with high-dimensional machine learning achieving greater than 99% accuracy in AML classification from transcriptomic data. 

Federated learning approaches enable multi-institutional collaboration with 96.5% accuracy in leukemia classification on heterogeneous datasets.

Conclusions and Relevance  AI technologies hold potential to improve AML treatment through enhanced risk stratification, early detection capabilities, and individualized treatment optimization.

##
AI CORNER
##

Chat GPT 5 writes,

• The review makes a persuasive case that AI could one day unify genomic, morphologic, and clinical data into a single risk framework for AML—but it also reveals how far the field still has to go.

Most cited studies remain retrospective, often trained on small or single-institution datasets, with limited external validation and uncertain generalizability across patient populations. Reported accuracies above 95% may partly reflect closed research settings rather than real-world complexity, where staining variation, scanner differences, and incomplete data routinely confound algorithms. Even the most “explainable” models still need prospective trials, workflow integration, and regulatory clarity before they can influence treatment decisions.

The technology’s direction is unmistakable—but don't overlook the gap between high-performance prototypes and practical, reproducible clinical tools. In that sense, this review captures a turning point for today: AI in leukemia care is not yet routine medicine—but it is no longer mere speculation either.