Machine learning strategies to predict late adverse effects in childhood acute lymphoblastic leukemia survivors

TL;DR

This paper introduces neural network models for predicting late adverse effects in childhood leukemia survivors, enabling earlier intervention and personalized follow-up, outperforming traditional models on small datasets.

Contribution

It develops graph-based neural networks for predicting health complications in childhood leukemia survivors, incorporating genomic data and improving prediction accuracy over existing methods.

Findings

Neural network models outperform linear and tree-based models on small cohorts.

Proposed VO2 peak prediction model does not require physical tests.

Created obesity prediction model using clinical and genomic variables.

Abstract

Acute lymphoblastic leukemia is the most frequent pediatric cancer. Approximately two third of survivors develop one or more health complications known as late adverse effects following their treatments. The existing measures offered to patients during their follow-up visits to the hospital are rather standardized for all childhood cancer survivors and not necessarily personalized for childhood ALL survivors. As a result, late adverse effects may be underdiagnosed and, in most cases, only taken care of following their appearance. Thus, it is necessary to predict these treatment-related conditions earlier in order to prevent them and enhance the survivors' health. Multiple studies have investigated the development of late adverse effects prediction tools to offer better personalized follow-up methods. However, no solution integrated the usage of neural networks to date. In this work, we developed graph-based parameters-efficient neural networks and promoted their interpretability with multiple post-hoc analyses. We first proposed a new disease-specific VO$_2$ peak prediction model that does not require patients to participate to a physical function test (e.g., 6-minute walk test) and further created an obesity prediction model using clinical variables that are available from the end of childhood ALL treatment as well as genomic variables. Our solutions were able to achieve better performance than linear and tree-based models on small cohorts of patients ($\leq$ 223) for both tasks.