Machine Learning Applied to the Early Diagnosis of Leukemia Using Biomarkers

Abstract

Leukemia is a rare and lethal blood cancer. One of the factors that increase the patient's chances of better treatment results is early diagnosis. The best attempt to discover leukemia usually is the image analysis exams, but this is costly, and sometimes it is late. Thus, this paper uses attributes of a complete blood count as input to Machine Learning algorithms to predict earlier and cheaper leukemia diagnoses. In this paper, we collected actual exam results. We developed a synthetic dataset with 1000 examples based on the distribution and limits of each attribute to classify a patient in positive or negative for leukemia. We tested different classifiers (Logistic Regression, Random Forest, XGBoost, and SVM) to predict sample classes. We show that it is possible with an accuracy of 96% to predict if a patient is likely to have leukemia based on its blood count.