Prediction of Relapse in Childhood Acute Lymphoblastic Leukemia

Abstract

Among childhood cancers, Acute Lymphoblastic Leukemia (ALL) is the most common type. Advancements in the treatment of this disease have increased the survival rate to 90 percent today. However, survivors are still at risk for relapse after remission, and once the disease relapses, the survival rate is much lower. A thorough analysis of data can potentially identify risk factors and reduce the risk of relapse in the first place. The main goal of this study is to identify patterns in those patients who experience a relapse, so these patterns can be used to predict relapse before it occurs. Four machine learning gradient-boosting algorithms and five tree-based models will be used on two datasets; one with data on patients who relapse and one with data on non-relapsing patients to achieve this goal. The performance of all four algorithms will be evaluated based on accuracy, specificity, and sensitivity. These measures will also be compared to tree-based models of the same measure. The results show that the LightGBM algorithm had the highest accuracy. The AdaBoost algorithm had the highest specificity. The XGBoost algorithm had the highest sensitivity. All four of the algorithms out-performed the tree-based models in both accuracy and specificity.

Conclusion

This study demonstrates the ability of gradient boosting classification models to predict and outperform tree-based models on small datasets when predicting the occurrence of relapse in patients with Acute Lymphoblastic Leukemia. Tree-based models underperformed against ADA, GBC, LGB, and XGB in all evaluated measures of performance; accuracy, specificity, and sensitivity. LightGBM performed best when measured by accuracy. AdaBoost performed best when measured by specificity. XGBoost performed best when measured by sensitivity. No single algorithm performed best in more than one measure, which indicates the robustness of using different algorithms when analyzing the same dataset.