Abstract
Background:
Rare AML subtypes, such as AML with BCR::ABL1 fusion gene and Inversion(3)/translocation (3;3) [inv(3)/t(3;3)] account for <2% of cases and are associated with poor outcomes. Despite the growing application of Artificial Intelligence (AI) in malignant hematology, diagnostic tools for this subset are critically understudied. We conducted a systematic review to evaluate the validated AI models in these AML subsets.
Methods:
Using PRISMA guidelines, a comprehensive search of PubMed, Embase, Scopus, and IEEE Xplore from 2010 to June 2025 was conducted. Eligible studies were original research using AI/ML for prognostication in AML with BCR::ABL1 or inv(3)/t(3;3). Inclusion criteria included human data, validated models and reported prognostic measures (such as AUC and HR). Abstracts, reviews, non-human studies, and research without subtype focus or model validation were excluded. TRIPOD-AI and PROBAST were used to assess the quality of the study. Outcomes included validation status, model performance, and applicability to rare AML subtypes.
Results:
Of the 7 relevant studies, 3 concentrated on the target subtypes. One study differentiated AML with BCR::ABL1 (n=2) from blast-phase CML by a unique transcriptome profile that included CD25 overexpression and ID4 downregulation (p < 0.05). Another study used a texture-based AI model in a mixed AML/MDS population (n=92), showing moderate prognostic ability (AUC 0.71; HR 2.38 [95% CI: 1.40–3.95] in training; HR 1.57 [95% CI: 1.01–2.45] in validation; p < 0.045). No AI models were identified for AML with inv(3)/t(3;3) (n≈100). Broader AML AI models lacked external validation and subtype specificity, limiting their relevance to these rare subtypes.
Conclusions:
Validated AI models for prognostication targeting rare AML subsets remain scarce, despite reported 5-year OS being less than 10% and relapse rates above 90%. While molecular signatures for BCR::ABL1 show promise and texture based AI offers moderate predictive power, there are currently no validated AI models yet for inv(3)/t(3;3). This highlights the need for integration of subtype-enriched cohorts in future models, implementation of interpretable AI frameworks, and pursuing multicenter external validation. This will help in offering and advancing individualized risk stratification and guiding therapeutic decisions in this high risk population.
