Analysis of leukemia cancer classification with supervised machine learning and deep reinforcement learning based on gene expression monitoring (via DNA microarray)

Cancer is a progressively common risk that endangers human health around the globe, and an early diagnosis necessitates a positive prognosis. A substantial quota of the world population dies because of cancer, and the research community is alarmed that by 2025 a probable increase of cancer cases will reach a staggering 25 million cases. Several studies have already proven cancer genesis is mostly supported by an accretion of dangerous mutations, amplifying the recognition of cancer based on genome information. Moreover, according to health practitioners, cancer has been stated as one of the most lethal illnesses of the genome. It has been the pivot of research by doctors, pathologists, biologists, data scientists, and other life science and health professionals as it is a need of the hour to have an advanced automated method that can quickly identify cancer and its subtype, i.e., Acute Lymphocytic Leukemia (ALL) and Acute Myelocytic Leukemia (AML). This research aims to identify leukemia cancer subclasses using DNA Gene Expression. Much research can be found using gene expression data in the cancer classification domain. All the research can be categorized into four major categories: traditional data mining methods, Deep Learning (DL) methods, ML methods, and Classification using Deep Reinforcement Learning (RL); the most famous datasets are The Cancer Genome Altus and The Gene Expression Dataset. However, most of the literature utilized one type of model or some combination of different models. To the best of our knowledge, the optimal compassion of Machine Learning (ML) models, DL models, and Deep-RL models are still unknown or unavailable in one place. This thesis aims to explore and implement the models based on ML, DL, and Deep-RL and provide a comparative analysis. A comparative analysis of the three feature selection techniques was performed to show the importance of feature selection in Leukemia cancer prediction. Based on the analysis result, Logistic Regression (LR) gave the highest accuracy of 97% with the raw dataset, while with PCA version of the dataset Support Vector Machine (SVM), Gaussian-NB, and LR obtained an accuracy of 91.%. Furthermore, with the Random Forest (RF) importance dataset version, SVM achieved higher accuracy of 97% along with lasso regularization dataset version of DNN performed well and obtained a higher accuracy of 97%. Out of six different Deep-RL models with PCA dataset, model-3,5,6 achieved better accuracy of 88.24%. Although the PCA version of six Deep-RL configurations, model-1 reached an AUROC value of 72, which is higher than any non-PCA (raw dataset) version of six Deep-RL models. Moreover, with RF importance dataset out of six different configurations of Deep-RL model 6 obtained higher accuracy of 88.24%, and with lasso regularization dataset version six different configurations of Deep-RL model 1 achieved higher accuracy of 73.53% Keywords: Deep Reinforcement Learning, Deep learning, Machine Learning, Gene Expression; Leukemia Cancer Classification; PCA; Random Forest Importance, Lasso Regularization