Finding optimal decision scores by evolutionary strategies

Severeness of illness is often rated by physicians at admission time. For this purpose, medical scores have been developed as 'objective' rating methods. When considering their classification performance, it is not assumed that such an expert-driven score is an optimal one. Our aim is to design an optimized data-driven score. In particular, we compare classical scores with a new data-driven score for abdominal septic shock patients.

Medical scores are used as ratings for different aspects of a patient's health status. The medical score indicates either a more critical or a healthier condition. For example, physicians rate organ conditions for different organs. We consider four different scores, SOFA, APACHE II, SAPS II, and MODS. Beyond the use of such classical scores, we propose an evolutionary strategy, that is suitable for score design, to find optimized data-driven scores. A database of 282 patients is used to optimize a new score for abdominal septic shock patients. Classification performance is compared by a ROC analysis.

We give a general instruction for building optimized scores, i.e. we define individuals and operators for the evolutionary score design task. We apply this instruction to abdominal septic shock patient data. When compared to the SOFA score, it has similar classification performance, but it is more performant than APACHE II, SAPS II, and MODS. It can be used as a daily bedside score.

We argue that evolutionary strategies should be used for optimizing purposes in the medical score design process. Using abdominal septic shock patient data, we show that evolutionary score design is a feasible and performant method that can complement or replace expert knowledge, provided that qualitative data is available.