Acute Myeloid Leukemia (AML) is one of the most complex and dangerous blood cancers. It is characterized by rapid proliferation of abnormal myeloid cells in the bone marrow that crowd out healthy blood cells. Despite decades of research, curative outcomes are achieved in only a minority of patients, and treatment decisions require integrating vast amounts of genetic, clinical, and imaging data.
AML is not a single disease but a collection of genetically distinct subtypes - each with different mutations, different prognoses, and potentially different optimal treatments. Identifying which subtype a patient has and what therapy they should receive requires synthesizing cytomorphology (microscopic cell appearance), cytogenetics (chromosome abnormalities), molecular genetics (specific gene mutations), and immunophenotyping (surface protein profiles).
Machine learning (ML) - a branch of artificial intelligence in which algorithms learn patterns from data rather than following explicit rules - is ideally suited to handling this complexity. ML can simultaneously process thousands of variables across thousands of patients, finding patterns invisible to human clinicians and generating predictions with consistent, objective accuracy.
This comprehensive review surveys the current state of ML applications across the full spectrum of AML management - from initial diagnosis to risk stratification, treatment selection, and transplant decision-making - while also frankly discussing the pitfalls and regulatory challenges that must be addressed before these tools reach everyday clinical practice.
Support Vector Machines (SVMs) work by finding the optimal mathematical boundary (hyperplane) that separates distinct groups of data points in high-dimensional space. In AML, SVMs excel at classifying subtypes from complex genomic datasets and distinguishing leukemic cells from normal cells in flow cytometry data.
Random Forests (RF) work by building hundreds of decision trees, each trained on a random subset of the data. Each tree votes on a classification, and the majority vote wins. This ensemble approach is robust to overfitting and can rank which variables (genes, clinical factors) most strongly influence the prediction, making results interpretable.
Deep Neural Networks (DNNs) are modeled loosely on the structure of the brain, with layers of interconnected computational nodes that learn progressively more abstract features from raw input data. DNNs excel at image recognition (analyzing bone marrow microscopy images) and can automatically extract features without human preprocessing.
Clustering methods like k-means and FlowSOM group data points by similarity without pre-specified categories. These unsupervised approaches can discover previously unknown patient subgroups and identify rare cell populations in flow cytometry data, potentially revealing new disease subtypes or therapy targets.
One of ML's most powerful diagnostic applications is analyzing bone marrow and peripheral blood microscopy images. Deep neural networks can be trained to recognize leukemic cells from images, performing cell segmentation (separating individual cells), feature extraction (measuring cell shape, nucleus size, staining intensity), and classification (assigning cell type) - all steps previously done manually by expert hematologists.
At the genomic level, Warnat-Herresthal et al analyzed over 12,000 samples from more than 100 different studies, combining transcriptomic and genomic data with ML to develop classifiers that detect AML accurately, quickly, and at low cost. Such large-scale approaches demonstrate ML's ability to work across diverse data sources and institutions.
ML can pinpoint specific dangerous mutations. FLT3-ITD, a mutation in the FLT3 tyrosine kinase gene present in about one-third of AML patients, is associated with poor prognosis. Support vector machines and random forests can identify genes that predict FLT3-ITD mutation status, helping to flag high-risk patients who need more aggressive therapy.
Flow cytometry - a technique that measures thousands of cells per second based on how they scatter laser light and emit fluorescence from labeled surface proteins - is another pillar of AML diagnosis. ML tools including SVM, random forests, and self-organizing map-based FlowSOM automate flow cytometry analysis, detecting residual disease (minimal residual disease, or MRD) with higher sensitivity than manual analysis.
Current standard AML risk stratification uses the European LeukemiaNet (ELN) 2017 classification, which divides patients into favorable, intermediate, and adverse risk groups based on specific mutations and cytogenetic abnormalities. While clinically useful, this system misses the full genetic complexity of individual tumors.
Multiple ML models have been shown to outperform the ELN classification. Fleming et al used random forests and decision trees on more than 2,000 non-APL AML cases and reported a lower error rate than ELN 2017. Shreve et al's ML model integrating clinical, cytogenetic, and mutational data showed significantly better performance at predicting individualized patient outcomes.
Deep neural networks can predict AML prognosis from cytogenetics, mutational status, and patient age, and have been used by Gerstung et al to build 'knowledge bank' models that accurately predict relapse, remission, and overall survival. These tools allow oncologists to move from population-level risk groups to personalized outcome predictions for each individual patient.
ML can also refine the understanding of specific mutations. Patkar et al identified genomic aberrations in NPM1-mutated AML (one of the most common AML mutations) and developed a scoring system that subdivides NPM1mut AML into three prognostic subgroups - revealing clinically meaningful heterogeneity within what was previously treated as a single entity.
ML can predict whether a patient will respond to treatment before therapy even begins. Using RNA sequencing data from pediatric AML patients, ML models can predict likelihood of complete remission after induction chemotherapy - helping oncologists identify upfront which patients need alternative or intensified regimens rather than standard chemotherapy.
In drug discovery, deep neural networks are proving especially valuable. Several studies have used DNNs to identify new therapeutic targets in AML, including SMARCA4 as a marker of sensitivity to topoisomerase inhibitors mitoxantrone and etoposide, novel FLT3 inhibitors mapped via drug discovery algorithms, and STAT3 inhibitor candidates for AML and myelodysplastic syndrome (MDS).
ML supports better clinical trial design and patient recruitment. Algorithms can scan electronic health records to identify eligible patients who might otherwise be missed, assess patient characteristics against inclusion/exclusion criteria automatically, and predict the likelihood of treatment success or failure in advance - improving trial efficiency and reducing costs.
For allogeneic stem cell transplantation - still the only curative option for many AML patients - ML algorithms developed by Shouval et al on 25,000+ leukemia patients from the European Society for Blood and Marrow Transplantation can predict 100-day overall survival after transplant. This helps oncologists select appropriate conditioning regimens and immunosuppression strategies for individual patients.
Measurable Residual Disease (MRD) refers to tiny amounts of leukemia cells remaining in a patient after treatment - too few to cause symptoms but enough to eventually cause relapse. Detecting MRD with high sensitivity is crucial for deciding whether to continue treatment, switch therapies, or proceed to stem cell transplantation.
Flow cytometry-based MRD assessment requires expert analysis that is not available at all hospitals. ML tools, particularly SVMs, have been shown to automate and standardize MRD assessment from flow cytometry data, making high-quality MRD testing accessible to centers that lack specialized expertise.
Collaborative multi-center ML models trained on large standardized MRD datasets could provide uniform, automated MRD assessment verified by field experts - bringing this critical prognostic tool to community hospitals and smaller cancer centers that currently cannot perform it reliably.
As AML treatment options expand with new FDA-approved agents including targeted inhibitors of FLT3, IDH1, IDH2, and BCL-2, MRD-guided treatment decisions will become increasingly important. ML can help determine not just whether MRD is positive but what specific residual cell populations are present, potentially guiding selection between the growing menu of targeted agents.
Overfitting is one of ML's biggest risks. A model that performs brilliantly on its training data may fail catastrophically on new patients if it has 'memorized' patterns specific to the training set rather than learning generalizable rules. Careful separation of training and validation data, and testing on independent patient cohorts, is essential to catch overfitting.
Most existing ML studies in AML are retrospective - built from historical data where the outcomes are already known. While useful for discovering patterns, retrospective models must be tested in prospective studies (where the model makes predictions before outcomes are known) to prove they work in real clinical practice.
Models trained on large public datasets like The Cancer Genome Atlas or Beat-AML may perform poorly when applied to regional patient populations with different demographic characteristics, treatment protocols, or data collection methods. Local validation in the target clinical environment is essential before clinical deployment.
Hidden biases can corrupt ML results in unexpected ways. For example, if microscopy images systematically include a scale bar in one position for AML samples and a different position for normal samples, the ML may learn to classify based on scale bar position rather than actual cell morphology - a problem that would be invisible without careful examination of what features the model actually learned.
The review argues that basic ML literacy should become part of medical training. Physicians who understand how different models work - and critically, what they cannot do - will be better equipped to integrate ML recommendations into patient care, identify when an algorithm may be failing, and push back against inappropriate clinical use of poorly validated tools.
ML does not replace clinical judgment - it augments it. An ML system provides probability estimates and risk scores, but the physician must contextualize these within the full clinical picture, the patient's preferences, and factors the algorithm cannot access. Human-AI collaboration is the goal, not automation of clinical decision-making.
Equitable access is a potential benefit of ML that is often overlooked. Well-designed ML tools could bring expert-level hematologic diagnostic capabilities to community hospitals and underserved regions where specialist expertise is unavailable, reducing geographic disparities in AML care quality.
Strict regulatory oversight and legal frameworks are essential for safe ML deployment in clinical practice. Adaptive regulation that keeps pace with the rapid development of ML technology, combined with mandatory prospective validation before clinical use, will be necessary to ensure patient safety as these tools transition from research to practice.
The vision articulated in this review is an integrated ML workflow that automatically analyzes data from all diagnostic modalities - microscopy images, flow cytometry, cytogenetics, molecular genetics, and clinical parameters - and provides the clinician with an integrated summary: the likely AML subtype, risk stratification, predicted outcomes with different treatments, and suggested therapeutic options.
Such integrative tools require close collaboration between hematologists, computational scientists, data engineers, and software developers - teams that rarely work together today. Building these partnerships and establishing shared data infrastructure is as important as developing the algorithms themselves.
The development of prospective basket trials - clinical trials testing targeted therapies against specific mutations rather than specific cancer types - will be facilitated by ML that can identify patients with disease-driving mutations across different cancers simultaneously. This represents a fundamental shift from histology-based to biology-based cancer medicine.
The ultimate goal is precision medicine for AML: every patient receiving a treatment matched to the specific biology of their individual leukemia, monitored in real time by ML-powered MRD surveillance, with treatment adjustments guided by algorithmic predictions. The technology to achieve this is emerging; the clinical validation and regulatory frameworks still need to catch up.