Acute myeloid leukemia (AML) is a fast-moving blood cancer that starts in bone marrow and is notable for its enormous biological diversity. Two patients who both appear to have AML may have completely different genetic profiles, different prognoses, and respond very differently to the same treatment — making truly personalized care challenging with traditional tools.
The current standard for deciding which patients need intensive treatment, stem cell transplants, or targeted therapies relies on the 2017 European LeukemiaNet (ELN) risk stratification, which groups patients into favorable, intermediate, and adverse risk categories based on selected genetic markers. While useful, this system has significant limitations: many patients in the same risk group have widely different actual outcomes, a problem called intra-stratum heterogeneity.
Machine learning (ML) — software that identifies patterns in large datasets without explicit programming — offers a potential solution. Unlike traditional statistical models with rigid pre-set rules, ML algorithms can automatically discover which combination of hundreds or thousands of variables best predicts a patient's outcome. As genomic testing has become cheaper and more common, the volume of molecular data available for each AML patient has grown dramatically, creating an ideal environment for ML tools.
This 2020 review from the Cleveland Clinic systematically surveys the emerging landscape of ML applications in AML, covering how these tools are being applied to diagnosis, prognosis, treatment decisions, and image analysis — and honestly assessing where challenges remain before clinical adoption can occur.
One landmark study reviewed by the authors used LASSO regression — an ML method that automatically identifies the most predictive variables — applied to transcriptomic (gene activity) data from over 12,000 patients pooled across 105 studies. Without any expert input or additional clinical data, the algorithm diagnosed AML with greater than 99% accuracy and distinguished the aggressive M3 subtype from others with 97.6% sensitivity and 99.5% specificity. Remarkably, model performance was not significantly affected when known AML-related genes were excluded, suggesting ML captures informative patterns beyond what humans have pre-selected.
For predicting how long patients would survive — arguably the most critical clinical question — a model using gradient boosted trees (GBT) achieved a concordance index (c-index) of 0.80 to 0.85 across five different patient datasets. This is far superior to the standard ELN criteria, which achieved a c-index of only 0.59 on the same data. A c-index near 1.0 reflects perfect prediction, while 0.5 is random chance, so this improvement represents a substantial clinical gain.
A separate study used genomic data combined with clinical information to build a model predicting whether individual AML patients would benefit more from receiving a stem cell transplant in their first remission versus waiting until relapse. The model, built on 1,540 AML patients, could potentially reduce unnecessary transplants in patients with favorable subtypes without harming overall survival — an example of ML guiding not just diagnosis but also major treatment decisions.
ML tools are not limited to genomic data. A recommender algorithm — similar in concept to systems used by streaming services to suggest content — was applied to AML-adjacent patients with myelodysplastic syndromes (MDS), identifying those at risk of resistance to standard therapy based on their pattern of gene mutations. Low-risk patients survived 22.8 months on average compared to 14.6 months for high-risk patients, demonstrating that ML can create meaningful risk distinctions even from relatively sparse genetic data.
Image analysis using convolutional neural networks (CNNs) — the same type of neural network that powers facial recognition technology — has been applied to bone marrow aspirate images in AML and related diseases. Models can classify cell types from bone marrow slides with sensitivity and specificity above 95%, potentially automating what currently requires highly trained pathologists. This capability could make expert-level diagnosis more accessible in resource-limited settings.
The authors also note emerging evidence that CNNs can extract clinically meaningful information from histopathology slides beyond simple classification — including predicting specific mutations and risk-stratifying patients based purely on tissue architecture. In solid tumors these capabilities are already demonstrated, and corollary applications in blood cancer pathology remain an active frontier worth exploring.
One major challenge highlighted by the review is overfitting — when an ML model learns patterns specific to its training dataset that do not generalize to new patients. A study of 26,266 patients undergoing stem cell transplant for AML found that model performance plateaued when as few as 6 variables were included, regardless of how many more were added. This suggests that the ceiling for predictive ability may not be broken simply by adding more data, but requires identifying genuinely new biological signals.
A second challenge is interpretability. Most high-performing ML models, especially deep learning approaches, operate as "black boxes" — they produce predictions without explaining their reasoning. This is a major barrier to clinical trust and regulatory approval. Methods like SHAP (SHapley Additive exPlanations) can dissect individual predictions and rank variable importance, helping translate what an ML model has learned into language clinicians can understand and critique.
External validity is also critical: a model trained on patients from large academic centers may not perform equally well in community hospitals with different patient demographics. Studies consistently show that models trained on multi-institutional data generalize better, underscoring the need for broad data collaboration before ML tools are deployed widely in AML care.
The review emphasizes that ML research in AML remains pre-clinical — promising and clearly advancing, but not yet ready for routine clinical implementation. Prospective validation in well-designed clinical trials, with diverse patient populations, is the next essential step toward realizing the potential these tools have already demonstrated in retrospective data.
The authors conclude that ML represents a genuinely transformative opportunity for improving AML management, particularly in risk stratification and treatment personalization. The ability to analyze thousands of molecular variables simultaneously — far beyond what traditional scoring systems can incorporate — means that ML models could eventually provide each patient with a precise individual risk estimate rather than a broad categorical label.
In the near term, ML-based diagnostic tools may help resolve ambiguous cases, standardize interpretation across institutions, and accelerate the diagnostic workflow. In the longer term, the integration of genomic, clinical, laboratory, and imaging data into unified predictive models could fundamentally change how oncologists approach treatment selection for AML.
The path forward requires multi-center data sharing, rigorous prospective validation, and continued work on interpretability so that clinicians, patients, and regulatory agencies can understand and trust the predictions these powerful tools generate. The authors are cautiously optimistic that this iterative process will eventually bring the benefits of ML to AML patients in the clinic.