Monitoring Acute Lymphoblastic Leukemia Therapy with Stacked Denoising Autoencoders

For acute lymphoblastic leukemia treatment monitoring, the ratio of cancerous blood cells, called Minimal Residual Disease (MRD), is in practice assessed manually by experts. Using flow cytometry, single cells are classified as cancerous or healthy, based on a number of measured parameters. This task allows application of machine learning techniques, such as Stacked Denoising Autoencoders (DSAE). Seven different models’ performance in assessing MRD was evaluated. Higher model complexity does not guarantee better performance. For all models, a high number of false positives biases the predicted MRD value. Therefore, cost-sensitive learning is proposed as a way of improving classification performance.