The cost of Type II Diabetes Mellitus: A Machine Learning Perspective

摘要

In this study, the burden of type II diabetes mellitus is investigated using machine learning methods. In particular, it is mainly aimed at obtaining an accurate quantification of the burden of diabetes by computing the number of indicators that provide the highest discrimination rate between normal people and patients. Assuming that the cardinality of the best-fitting feature set can be used to quantify the magnitude of the overall burden, several healthcare related features are extracted from demographic, diagnosis, medication and lab test records. Experimental results have shown that there are about 200 relevant indicators and the highest classification performance achieved in discriminating diabetic and normal people is remarkably superior to that of the baseline system. In other words, the burden of diabetes is not limited to a small group of complications, medications or lab tests. In the second phase of experiments, the relative effects of different indicators are evaluated by employing Lasso and Ridge regression algorithms. It is observed that the best set of indicators have different levels of effects in discriminating between diabetic and normal people.