An improved method for assessing costs for patients admitted for treatment using neighborhood rough set and artificial neural network (ANN)
Assessing costs for patients admitted for treatment is considerably important in hospital system; however, the relationship between this outcome and its related factors is complex and high-dimensional. In medical science, ANN models are successfully applied in many fields. However, ANN models are considered as a black box model that has some disadvantages including heavy computational burden, proneness to overtraining, and the empirical nature of model selection. Thus, we developed a hybrid method to overcome these problems.
Classic rough set can extract knowledge from the raw data, but it cannot handle the numerical variables directly. Neighborhood rough set can compute attribute reduction from nominal variables, numerical variables and their mixtures when multi-variety data is available. The mean square error and other homogeneous criterions are commonly used to evaluate the ANN’s performance, but these criteria are correlated to model’s architecture and sample size. Therefore, we applied an information based criterion, the Akaike information criterion (AIC) to evaluate the ANN’s performance.
METHODS:
Sample data set was obtained from the records of 798 patients admitted for treatment of myocardial infarction, containing 24 condition attributes and 1 decision attribute (treatment costs). The superfluous factors were eliminated by the method of neighborhood rough set. Then, a three-layer back-propagation neural network (BPNN) was used to assess the treatment costs. The AIC was used to evaluate the performance of the proposed method.
RESULTS:
Based on neighborhood rough set, 12 factors were selected as input variables for the ANN to assess the treatment costs. The corresponding AIC is -5.20, which is lower comparing to the AIC (-4.32) of using ANN only.
CONCLUSIONS:
Neighborhood rough set can efficiently eliminate superfluous factors without loss of information. The proposed hybrid method based on neighborhood rough set and ANN outperformed the traditional ANN methods in estimating treatment costs, which can be applied in hospital system.