Multi-modal evolutionary ensemble classification in medical diagnosis problems

Expert systems for classification tasks in medical diagnosis systems require two properties. The true positives should be very high, as well as the true negatives, i.e. the system should correctly catch those who are ill, and correctly dismiss those who are healthy. The multi-modal evolutionary classifier uses a genetic algorithm to learn a reference vector for each class, and classification is done by measuring the distance of the new example to reference vectors. For complex datasets such as medical diagnosis, interactions between features are typically complex and the multi-modal classifier´s single reference vector is not able to capture this. In this work an extension to the algorithm is proposed, which learn sets of multi-modal classifiers using resampling and form an ensemble from these, using a genetic algorithm. The algorithm is evaluated on a sample of publicly available medical diagnosis datasets. While this is a work-in-progress, initial findings are that compared to the base classifier, using evolutionary learned ensembles improves accuracy in all cases, and is a direction for future work.