Enhanced Detection of Relapse Risk in Pediatric ALL Using Multivariable Predictive Modeling

The Pediatric Relapse Prediction and Risk Evaluation for Acute Lymphoblastic Leukemia (PREPARE-ALL) tool was developed as a machine learning–based clinical decision–support tool for predicting relapses in children with ALL. According to findings from an Indian study published in JCO Clinical Cancer Informatics, PREPARE-ALL’s predictive ability to identify relapses is stronger than that of clinicians.

Researchers gathered data from the Indian Collaborative Childhood Leukemia group (ICiCLe) ALL-14 pretrial cohort, which included 2,252 pediatric patients with B-precursor acute lymphoblastic leukemia from five centers. Of these patients, 25.1% relapsed.

The machine learning classification model achieved a sensitivity of 68.5%, a positive predictive value of 31.3%, a negative predictive value of 82.8%, an accuracy of 54.8%, and a specificity of 50.3%.

A total of 14 features associated with relapse were assessed in sensitivity analyses; the most common predictors of relapse were high hyperdiploidy and BCR-ABL1 fusion positivity, measurable residual disease after induction therapy, sex, age, elevated white blood cell count, and risk group.

In the validation data set, the final model achieved a recall of 68.5% compared with 31% to 36% based on three clinicians’ judgments. Interclinician agreement showed variability in assessments for relapse compared with the more consistent results of the AI model.

Accuracy was lower in intermediate- and high-risk subgroups, but recalls were reported for all groups with a rate of 52.9% for patients with standard risk, 44.8% for those with intermediate risk, and 82.5% for high-risk patients.

“By prioritizing sensitivity, the model addresses the key challenge of missed relapses and serves as a scalable adjunct for risk-adapted therapeutic planning, rather than replacing clinician judgment,” the study authors, including co-senior author Venkatraman Radhakrishnan, MD, MBBS, MSc, DM, of the Department of Medical Oncology, the Cancer Institute (WIA) in Chennai, India, wrote in their report.

Model Methods and Selection

The researchers intended for the classification framework to reduce noise and bias and identify only the most important features associated with relapse in pediatric patients with acute lymphoblastic leukemia.

They tested the performance of 10 different machine learning models—linear machine learning methods of logistic regression, support vector machines, and Gaussian Naive Bayes; decision trees and random forest trees; boosting algorithms, such as eXtreme Gradient Boosting (XGBoost); and k-nearest neighbor distance-based learning—to predict relapse and find the most reproducible approach. XGBoost was chosen as the final model due to its superior and balanced performance; although its accuracy was lower than that of other models, its recall performance was prioritized because the researchers decided that a missed relapse would be more consequential than a false positive.

Data sets were split 80:20 for training and validation with fivefold cross-validation for model selection. For comparison, three clinicians evaluated the validation data set to determine concordance with the AI model predictions.

DISCLOSURE: Dr. Saravanan declared being an inventor on a granted patent for the PREPARE-ALL relapse prediction algorithm, but did not receive royalties for it. For full author disclosures, visit ascopubs.org.

AI in Practice: Have you tried out any similar predictive AI models in your own practice? We want to hear about your experiences! Connect with us by email.

ASCO AI in Oncology is published by Conexiant under a license arrangement with the American Society of Clinical Oncology, Inc. (ASCO^®). The ideas and opinions expressed in ASCO AI in Oncology do not necessarily reflect those of Conexiant or ASCO. For more information, see Policies.