PII-064 - MODELING MORTALITY OF PEDIATRIC PATIENTS UNDERGOING HEMATOPOIETIC STEM CELL TRANSPLANTATION USING SUPERVISED MACHINE LEARNING.

A. Dunn¹, A. Cao¹, J. Gobburu¹, J. Long-Boyle², R. Goyal¹; ¹University of Maryland, Baltimore, MD, USA, ²University of California, San Francisco, San Francisco, CA, USA.

Background: Currently, there is an unmet need for the identification of factors that may increase mortality in pediatric patients undergoing hematopoietic stem cell transplantation (HSCT). A machine learning (ML) method for risk factor discovery by modeling mortality was developed for this purpose.

Methods: Patient-related, treatment-related, and survival outcome data was retrospectively collected from pediatric patients who underwent HSCT and analyzed using Pumas v2.1. Due to low sample size and class imbalance, a total of 200 training and testing samples were prepared using stratified bootstrapping to predict 1-year mortality. Ten ML algorithms were trained on the samples. The final model was selected based on sensitivity, specificity, and accuracy, and feature importance was obtained using the mean absolute Shapley value for each feature across all samples.

Results: Of the 68 subjects analyzed 25% were decedents. Gaussian Naïve Bayes algorithm demonstrated optimal performance, with a median sensitivity of 83%, specificity of 32%, and accuracy of 43%. The risk factors identified as of highest importance included donor relation, degree of mismatch, serotherapy regimen, requiring retransplant, and days to absolute neutrophil count exceeding 500 cells/µL.

Conclusion: To the best of our knowledge, this is the first time a ML approach was used to predict 1-year mortality and identify risk factors for pediatric patients undergoing HSCT. Due to the low sample size, a median sensitivity of 83% justified the adequacy in model performance. This model suggests that receiving cells from an unrelated donor and the degree of donor cell mismatch are the greatest risk factors for mortality.