Predicting Patient's Choices of Hospital Levels Using Deep Learning and Representation Improvements

In countries that enable patients to choose their own healthcare providers, a common problem is that the patients do not go to unsuitable hospital levels. This might cause problems such as overwhelming tertiary facilities with mild condition patients, and resulting in limited the treatment for acute and critical patients. Our aim is to predict patients' choices of hospital levels to support the evaluation during policy making. We proposed a deep neural network (DNN) framework, which involves an improvement of the representation for insurance data, a DNN design to make accurate predictions, and a model interpretation to further understand the decision of the model. This study used the 5-year nationwide insurance data of Taiwan. With the combination of autoencoder (AE) and DNN, the prediction results achieved an accuracy of 0.94, area under the receiver operating characteristics curve (AUC) of 0.88, sensitivity of 0.93, and specificity of 0.97 with highly imbalanced data. The result shows that changing data representation had a positive effect on the prediction model. The model interpretation results show that past medical experiences and recommendations of acquaintances are most important. Deep learning technology acts as a feasible tool that provides additional evaluation besides using traditional statistical methods.