Cost-Sensitive Neural Network for Optimizing Medical Expenditures of Populations
Kiarash Zahirnia, Mehdi Teimouri · 2018
For critical and costly medical issues hospitals, governments, and insurance companies often incur asymmetrically large costs for misdiagnosis in the form of false negatives. Therefore, hospitals are incentivized to perform many medical tests to prevent misdiagnosis. But since each test itself incurs a cost, a critical question occurs: which tests need to be performed for which patients to minimize the total cost across all patients, including the cost of the medical tests and cost of misdiagnosis. Using machine learning techniques, we can learn to minimize this cost for future patients based on previous training data. Conventional machine learning methods cannot handle the asymmetric effect of misdiagnosis, and do not choose which features are more important for each individual instance. For this scenario, cost-sensitive machine learning should be used. In this paper, a new method of cost-sensitive machine learning is proposed. The proposed method uses two different strategies to make decision for patients with unknown features. We evaluate the proposed method by comparing to some of the most common cost-sensitive learning methods and show superior results in many cases. We apply these methods to nine real-world medical datasets.