Using Deep Learning to predict the mortality of Leukemia patients

Thumbnail Image
Muthalaly, Reena Shaw
Deep Learning , H2O , Leukemia , Cancer , Acute Lymphoblastic Leukemia , ALL
“If it were not for the great variability among individuals, medicine might as well be a science, not an art.” Sir William Osler, Canadian physician and McGill alumnus, quoted in 1892. Personalized medicine is the approach that tailors the treatment of patients based on their unique genetic makeup and genetic environment. When equipped with details of individual variation, physicians can separate patients into subgroups to predict which patients must be treated aggressively and which patients would respond to one drug rather than another. Personalized medicine is now being applied towards the prediction of mortality in childhood Acute Lymphoblastic Leukemia (ALL) patients. This is because individual children differ in the sensitivity of their leukemic cells and in their response to treatment-related toxicity. Currently, mortality prediction for childhood ALL is based on risk-stratification performed by expert doctors. The genotypic (actual set of genes), phenotypic (expression of those genes in observable traits) and clinical information is used to stratify children into various risk categories. The information collected by doctors for risk-stratification include response to certain drugs, clinical factors such as age and gender and biological measurements such as white blood cell count. The goal of this thesis is to automate the prediction of childhood ALL mortality using genotypic and phenotypic information of patients. We intend to achieve this using the Deep Learning algorithm, which is known to analyze non-linear and complex information, such as cell-interactions, effectively. We conduct this thesis using the Big Data software called H2O, using R. We first experimented with the standard Deep Learning parameters, and later adjusted the size and shape of the general network by optimizing the depth and retention factor of hidden neurons. We improved the accuracy by using the techniques of Out-of-Bag Sampling, Bagging and Voting. We later tuned H2O’s platform-specific network parameters. Later, the number of votes needed to obtain the highest accuracy of predictions, is also tuned. We built confusion matrices for each of our Deep Learning models to evaluate how well our models perform. Our results are promising as they are corroborated by the clinical dataset and they perform better than the physicians’ predictions.
External DOI