Combining machine learning and clinical rules to build an algorithm for predicting ICU mortality risk
Michael Krajnak, Joel Q. Xue, Willi Kaiser, William Balloni · Computing in cardiology · 2012
In this study we aim to develop a decision support application for predicting ICU mortality risk that starts with a clinical analysis of the problem that also leverages machine learning to help create an algorithm with good performance characteristics. By starting from a clear basis in clinical practice we hope to improve algorithm development and the transparency of the resulting system. We start with a general model structure for a fuzzy rule based system (FIS). The model can be specified by clinicians who identify the inputs and the rules. An optimizer based on a genetic algorithm generates the coefficients for the final solution. Using the 2012 PhysioNet/CinC Challenge data set we constructed a Phase 1 system using minimal clinical guidance. Our initial FIS's achieved scores of 0.39 for Event 1 and 94 for Event 2. In Phase 2 we updated the FIS based on clinician interviews. At the end of Phase 2 we achieved 0.40 for Event 1 and 60 for Event 2. We hope to show that machine learning techniques that are modeled on the clinical understanding of a problem can be competitive with more abstract machine learning approaches but may be preferable because of their explainability and transparency.