Automated CLL cell population detection using a weakly supervised approach and CLL MRD flow cytometry data

Minimal/measurable residual disease detection is routinely performed as part of post-diagnostic treatment plans for many types of cancer, for which multiparameter flow cytometry is one possible modality frequently used. We propose a machine learning approach for binary prediction of minimal residual disease status with flow cytometry data. Our method involves the projection of cells from the original feature space to a low-dimensional embedding in which cells are clustered by similarity, and the cluster-wise cell proportions are used for prediction as well as regression. This is a weakly supervised approach in that the only annotation required for the training set data is the percentage of neoplastic cells present in each case. We demonstrate the applicability of our method with respect to a cohort of chronic lymphocytic leukemia patient data to obtain high levels of accuracy. We contrast our approach with other proposed machine learning methods for application to minimal residual disease prediction.