A Computational, Topological Approach to ICU Mortality Rate Prediction with Data Relationship Realization

The objective of this work is to predict the mortality of intensive care unit patients based on their physiological data and understand the relationships between physiological data. Such a model may be used to prioritize care when resources are limited or identify patients that will need significant care in the immediate future. This effort will take a novel approach applying computational topological analysis to classify patients. The algorithm predicting the patient outcomes is trained using an evolutionary algorithm. The dataset used is from the 2012 PhysioNet Computing in Cardiology Challenge. A set containing 4000 records with outcomes was used to train and test the prediction algorithm. The topology extraction algorithm, Mapper, was used to represent the high dimensional data as a 1-D graph of the set topology using a filter. The filter is trained using an evolutionary algorithm to maximize the positive prediction rate and sensitivity. The Event 1 score is the minimum of these two. This algorithm yielded an Event 1 score of 0.42 out of 1.00 for the PhysioNet Challenge. This is comparable to a currently used ICU classification system, SAPS-1 that achieved an event 1 score of 0.30.

Additional developments from this work include an optimized Mapper clustering function that runs in 120 seconds for the complete data set compared to the 2.2 month estimate using the original function. This allowed the rapid iteration needed for optimization in this algorithm. The algorithm developed in this thesis could be more generally applied to analysis and prediction in any feature space for generic problems.