Over the past decade, the growing volume and diversity of clinical data from electronic health records (EHR) has increasingly affected evidence-based treatment, personalized medicine, and medical ‘Big Data’ analysis. Given the EHR from an individual patient, how to assign the personalized treatment? how to predict the most likely future event? For a cohort in which most of patients only have one encounter record, traditional machine learning methods such as logistic regression, random forest, and Multilayer Perceptron (MLP) with 1 hidden layer can be used to build the predictive model. We applied those methods in evaluating the vasopressor treatment for Subarachnoid Hemorrhage (SAH) patients. Compared with norepinephrine and dopamine, phenylephrine was shown to be the best vasopressor treatment. In the cohort that patients have long historical encounters, the deep learning technique, Recurrent Neural Network (RNN), has shown the state-of-the-art performance in prediction. We implemented RNN based model to predict heart failure for patients who have ever been diagnosed with unspecified essential hypertension. The model achieved high accuracy in prediction and provided Interpretability.
Duo Yu, PhD Candidate at Department of Biostatistics and Data Science, School of Public Health, University of Texas Health Science Center at Houston.