EMOCS: evolutionary multi-objective optimisation for clinical scorecard generation

Abstract

Clinical scorecards of risk factors associated with disease severity or mortality outcome are used by clinicians to make treatment decisions and optimize resources. This study develops an automated tool or framework based on evolutionary algorithms for the derivation of scorecards from clinical data. The techniques employed are based on the NSGA-II Multi-objective Optimization Genetic Algorithm (GA) which optimizes the Pareto-front of two clinically-relevant scorecard objectives, size and accuracy. Three automated methods are presented which improve on previous manually derived scorecards. The first is a hybrid algorithm which uses the GA for feature selection and a decision tree for scorecard generation. In the second, the GA generates the full scorecard. The third is an extended full scoring system in which the GA also generates the scorecard scores. In this system combinations of features and thresholds for each scorecard point are selected by the algorithm and the evolutionary process is used to discover near-optimal Pareto-fronts of scorecards for exploration by expert decision makers. This is shown to produce scorecards that improve upon a human derived example for C.Difficile, an important infection found globally in communities and hospitals, although the methods described are applicable to any disease where the required data is available.