A Cost Skew Aware Predictive System for Chest Drain Management

Abstract

Many real world classification problems are skewed in terms of the cost of potential misclassifications. Clinical health care for individuals following pulmonary resection involves chest drainage tube management where the decision to remove or maintain a patient’s chest drain post-surgery is one such skewed classification problem. This is because the associated cost of premature removal is significantly higher than delayed removal in terms of health risks, discomfort and economic factors. While recognition of a cost differential in a problem is straightforward, its implementation in a predictive system is not, because there is no simple way to quantify cost. We addressed this issue through the design of an evolutionary based optimization approach for cost matrices. In order to test our approach, we compared three different settings: one with no cost matrix, one where the cost matrix used is provided by the thoracic surgeons, and one where the cost matrix is optimized through our evolutionary algorithm. The results show that our optimization method for cost matrices yields a large improvement over the other two settings on most performance measures, including an almost 20% increase in overall accuracy. This is a surprising result since it suggests that cost matrices provided by experts may not be as useful as those derived by a computational optimization approach.