Neural networks to predict clinical events from cytometry data

From a computational perspective, clinical decision making requires classifying a patient into classes that respond similarly to treatment. Cytometry characterises individual cells in patient specimens by size, complexity or surface markers. The result is a matrix per specimen with an unordered number of rows refecting the cells and a defined number of cell-parameters as columns. These data can be used for classification, hence also for clinical decision making. Classically, predictive modelling uses cell subpopulation quantities as a predictor to classify each sample. The cell subpopulations are defined by binning cells with similar characteristics. Sequential gating, manually defining cuts in the parameters, or automated clustering are possibilities for binning. In contrast to this two-step procedure, this thesis investigates the application of neural networks to classify cytometry samples without prior cell subpopulation identification. I provide a Cell Cloud Classification python package (CCC): A modular framework for neural networks to classify and investigate cytometry samples. The investigation includes a novel, decision-tree based concept to explain the predictions. I study the impact of hyperparameters by classifying public cytomegalovirus samples. The impact was measured using the classification AUC. Here, using the mean proved superior to maximum pooling. There were an optimum number of layers, but the number of nodes was irrelevant. The number of drawn cells, but neither replacement nor focus on rare cells were relevant. Surprisingly, network initialisation affected performance, but validation and test performance correlated. I study CCC in binary and multiclass simulations. Additionally, classifying T-cell based flow cytometry samples of (non-) starved human blood resulted in reliable but non-classical discriminative cell subpopulations. In summary, I present a python package to classify cytometry samples using neural networks and show its application to real and simulated datasets.