Prediction models for different types of leukemia: a systematic review and critical appraisal.

Abstract

Objectives: To systematically review and evaluate the methodological quality and risk of bias (ROB) of leukemia prediction models essential for clinical decision-making.

Methods: We reviewed 148 prediction models published before August 2023 from PubMed, Embase, Cochrane Library, and Web of science databases. Two reviewers independently screened articles and extracted data using CHARMS criteria. ROB was assessed using PROBAST. Models were categorized by leukemia subtype and analyzed for methodological characteristics.

Results: A total of 61 acute myeloid leukemia (AML) models primarily predicted survival (82.0%), diagnosis (4.9%), or death (4.9%) using predictors including age, cytogenetic risk, and white blood cell count. Among the 22 chronic myeloid leukemia (CML) models, the focus was on survival (72.7%) and time to treatment (19.0%), utilizing blast percentage, age, and platelet count. A total of 21 chronic lymphocytic leukemia (CLL) models primarily predicted survival (71.4%) using IGHV status, Rai stage, and age. The methodological shortcomings including incomplete reporting, methodological limitations, and high ROB were consistent across different leukemia subtypes. Traditional statistical methods predominated (Cox regression 72.9%, logistic regression 12.2%), with only nine machine learning models. Critical methodological limitations included lack of internal validation (52.0%) and external validation (57.4%). Only 43.2% reported discrimination metrics (AUC 0.60-0.99), with 28.0% achieving AUC > 0.7. Calibration was reported in only 23.0% of models. High ROB affected 93.9% of studies, primarily due to inadequate data handling and validation.

Conclusions: Existing leukemia prediction models have limited clinical utility due to methodological shortcomings and high ROB. Future research should prioritize transparent reporting, rigorous validation, and external validation to enhance clinical applicability and generalizability.