Project: Machine learning strategies for the diagnostic and the prediction of late adverse effects related to childhood lymphoblastic leukaemia treatment.

Project presentation


Completed (2020-2022)




1 Computer science department, Université de Sherbrooke, Sherbrooke (QC), Canada


Acute lymphoblastic leukemia (ALL) is the most frequently diagnosed cancer in children. Even though childhood ALL presents a high survival rate, approximately two thirds of survivors present one or more late adverse effects such as obesity, dyslipidemia, osteonecrosis and hypertension during adulthood. The current research project focus on the implementation of tools to help the diagnosis and the early prediction of particular late adverse effects.

The first phase is dedicated to the implementation of an ameliorated model to estimate the maximal oxygen consumption of childhood ALL survivors. This estimate can further be used for the diagnosis of cardiorespiratory health conditions.

The second phase focuses on the development of an early obesity prediction model that uses clinical variables from the end of childhood ALL treatment as well as genomic variables.

For each of the phases mentioned above, we compare the performance of graph neural networks to other common models, precisely the multilayer Perceptron, a linear regression, a random forest and a decision tree with gradient boosting (XGBoost). In particular, the second phases presents the performance of a new graph neural network architecture used to encode the genome of the patients. The latter, is presented in the figure below.