Cost-sensitive feature selection in medical data analysis with trace ratio criterion

Feature selection and classification are important tasks in medical data mining. However, different misclassifications of medical cases could lead to different losses. This paper proposes a framework for medical data classification and relevant feature selection by the combination of the trace ratio criterion and a novel cost-sensitive linear discriminant analysis classifier approach. The proposed multi-class cost-sensitive linear discriminant analysis classifier uses linear discriminant coefficients as conditional probabilities to estimate the posterior probabilities of a testing instance, calculates misclassification losses via the posterior probabilities, and predicts the class label that minimizes losses. Experimental results showed that the proposed scheme have comparable or even lower total cost and higher accuracy than state-of-the-art cost-sensitive classification algorithm.