Detection of Patients with Influenza Syndrome Using 7 Machine-Learning Models Automatically Learned from 47245 Emergency Department Reports
Abstract: The information available in Emergency Department (ED) reports has the potential to improve the detection of diseases; however most ED reports are normally locked from analyses due to their free-text format. This paper proposes an automatic approach to learn an influenza syndrome (IS) model from free-text ED reports. We used 7 machine learning algorithms (K2, NB, EMBC, SVM, LR, ANN, RF) to construct IS models in addition to the expert constructed model. We used two ED data sets: data set 1 (DS1) has 47245 ED cases and DS2 has 580 cases; and performed a two-step evaluation: 10-fold cross validation using DS1, and a single fold validation using DS1 as a training set and DS2 as a test set. The performance metrics used were AUROC and calibration. Results show that EBMC and NB algorithms outperformed others. We believe that the proposed approach can be applied to the modeling of other diseases.