PREPARE ALL: An Artificial Intelligence Tool for Predicting Relapse in Children With Acute Lymphoblastic Leukemia

Area of Science:

• Pediatric Oncology
• Machine Learning in Healthcare
• Hematologic Malignancies

Background:

• Acute Lymphoblastic Leukemia (ALL) is a significant concern in pediatric oncology.
• Accurate prediction of relapse is crucial for timely treatment adjustments and improved outcomes.
• Current methods for relapse prediction may not fully leverage complex clinical and laboratory data.

Purpose of the Study:

• To develop and validate the Pediatric Relapse Prediction and Risk Evaluation for Acute Lymphoblastic Leukemia (PREPARE-ALL) tool.
• To integrate clinical expertise with machine learning (ML), specifically Extreme Gradient Boosting (XGBoost), for relapse prediction.
• To compare the sensitivity of ML-based predictions against individual clinician assessments in pediatric ALL.

Main Methods:

• Development of the PREPARE-ALL tool using data from the ICiCLe ALL-14 pretrial cohort across five centers.
• Inclusion of 33 clinical and laboratory features in the model.
• Utilized an 80:20 train-test split for model validation with XGBoost.

Main Results:

• The XGBoost model achieved a sensitivity of 68.5% in detecting relapses among 2,252 pediatric ALL patients.
• Key predictors of relapse included high hyperdiploidy, BCR-ABL1 fusion positivity, and measurable residual disease status.
• PREPARE-ALL demonstrated higher recall (68.5%) compared to clinical judgment (approximately 31%-36%).

Conclusions:

• The PREPARE-ALL tool effectively predicts relapse in pediatric ALL, identifying twice as many relapses as clinicians.
• It serves as a practical decision-support tool for early relapse triage and treatment planning.
• Timely therapeutic adjustments guided by PREPARE-ALL can potentially improve outcomes in pediatric ALL patients.