Artificial intelligence-based prognostic models in acute myeloid leukemia: systematic review and meta-analysis

Area of Science:

• Hematology
• Medical Informatics
• Biostatistics

Background:

• Machine learning (ML) and deep learning (DL) show promise for improving survival prediction in acute myeloid leukemia (AML).
• Comparative benchmarks for artificial intelligence (AI)-based survival prediction models in AML are not well-established.
• Existing studies often lack standardized reporting and rigorous external validation.

Purpose of the Study:

• To systematically review and meta-analyze the performance of AI models for overall survival (OS) and relapse-free survival (RFS) prediction in AML.
• To compare the performance of gene-centric versus nongenetic AI models.
• To assess the risk of bias and optimism bias in AML survival prediction models.

Main Methods:

• Systematic literature search using PRISMA 2020 guidelines (PubMed, Scopus, Web of Science, January 2018-March 2025).
• Inclusion of studies developing or externally validating AI models for AML survival prediction, reporting Area Under the Receiver Operating Characteristic Curve (AUC).
• Data extraction on study design, population, features, algorithms, and AUCs; risk of bias assessment using PROBAST; random-effects meta-analysis.

Main Results:

• Included 24 studies (137 model cohorts, ~51,055 patients); 74% had low risk of bias, with statistical analysis being the weakest domain.
• Pooled validation AUC across 73 independent cohorts was 0.769 (95% CI, 0.742-0.795), with substantial heterogeneity (I² = 95.7%).
• Validation AUCs increased with prediction horizon (5-year AUC = 0.833). Nongenetic models showed a trend towards higher AUC (0.776) than gene-centric models (0.741).

Conclusions:

• AI prognostic models for AML demonstrate moderate discrimination with modest optimism bias but significant heterogeneity.
• Limited prospective validation necessitates standardized reporting and rigorous external evaluation of these models.
• Performance varies by prediction horizon and feature type, highlighting the need for careful model selection and validation.