Developing the Interpretability of Deep Artificial Neural Network on Application Problems

Abstract

In recent years, the use of electronic health records (EHR) has increased dramatically. Mining hidden knowledge in “big data” from EHR has become a subject worthy of exploration. On the other hand, many recent applications used deep artificial neural network (ANN) to analyze EHR data and yielded great performance. Accordingly, this study developed functional models using deep ANN, and tried to validate effectiveness of this method in regression analysis and classification problem. Based on datasets downloaded from the UC Irvine Machine Learning Repository, the output mean squared error value 0.840 was within the range of one variance for the regression analysis. Similarly, the prediction accuracy 73.0% on the testing data was reported for the classification problem. Another focus of this study was identifying critical attributes using the layer-wise relevance propagation (LRP) algorithm to improve interpretability of deep ANN. According to evaluation outcomes, the identified features would match with those recognized by univariate analysis. In summary, effectiveness of deep ANN and LRP on application problems has been validated in this study.