Using Electronic Health Records and Machine Learning to Make Medical-Related Predictions from Non-Medical Data

摘要

Objectives: Administrative HIS (Hospital Information System) and EHR (Electronic Health Record) data are characterized by lower privacy sensitivity, thus easier portability and handling, as well as higher information quality. In this paper we test the hypothesis that the application of machine learning techniques on data of this nature can be used to address prediction/forecasting problems in the Health IT domain. The novelty of this approach consists in that medical data (test results, diagnoses, doctors' notes etc.) are not included in the predictors' dataset. Moreover, there is limited need for separation of patient cohorts based on specific health conditions. Methods: We experiment with the prediction of the probability of early readmission at the time of a patient's discharge. We extract real HIS data and perform data processing techniques. We then apply a series of machine learning algorithms (Logistic Regression, Support Vector Machine, Gaussian Naïve Bayes, K-Nearest Neighbors and Deep Multilayer Neural Network) and measure the performance of the emergent models. Results: All applied methods performed well above random guessing, even with minimal hyper-parameter tuning. Conclusions: Given that the experiments provide evidence in favor of the underlying hypothesis, future experimentation on more fine-tuned (thus more robust) models could result in applications suited for productive environments.