(B-006) Prediction Models for Diagnosis and Prognosis of Leukemia: Systematic Review and Critical Appraisal

Background: Leukemia is a malignant hematological disease with a rising incidence rate. Clinical prediction models offer opportunities for early diagnosis, precision treatment, and prognostic assessment for leukemia patients. Numerous prediction models have been published for leukemia to date.

Objectives: To assess the methodological quality of prediction models for leukemia patients and to evaluate their risk of bias (ROB).

Methods: This systematic review was reported in accordance with transparent reporting of multivariable prediction models for individual prognosis or diagnosis: checklist for systematic reviews and meta-analyses. Studies that developed or validated multivariate prediction models to estimate the risk of individuals with leukemia and/or its pre-malignant conditions before August, 2023 were searched in the PubMed, Embase, Web of Science, and Cochrane Library databases. Two authors independently performed the literature search and study selection. Data were extracted using a standardized data extraction form designed according to the CHecklist for Critical Appraisal and Data Extraction for Systematic Reviews of Predictive Modelling Studies (CHARMS). Pairs of two reviews independently extracted data by Epidata and Excel software. The Prediction model Risk Of Bias ASsessment Tool (PROBAST) was used to assess the ROB for eligible studies.

Results: This review included 148 prediction models, comprising 6 diagnostic and 142 prognostic models. The primary outcomes for prognostic models were overall survival (n=36) and 5-year survival (n=20). The most common predictors were age (n=78), white blood cell count (n=43), platelet count (n=28), cytogenetic risk group (n=27), and blasts (n=27). Significant gaps in reporting on missing data proportions, imputation methods, and predictor variable handling were observed. Traditional models including Cox regression (n=108) and logistic regression (n=18) predominated, only 9 models employed machine learning. Seventy-seven models (52.0%) lacked internal validation, and 85 (57.4%) had no external validation. Only 64 (43.2%) models reported discrimination metrics and the area under the curve (AUC) ranged from 0.60 to 0.99, with 47 having an AUC above 0.7. Only 34 (23.0%) models reported calibration metrics. Overall, 139 models were rated as high ROB, primarily due to issues in predictor transformation, handling of missing values, model validation and performance.

Conclusions: Numerous prediction models for leukemia were developed but severely limited for practice for incomplete reporting and lack of external validation. Future model studies should focus on comprehensive reporting and model validation methods to enhance their methodological quality.