Discovering the Predictive Value of Clinical Notes: Machine Learning Analysis with Text Representation

Hospital readmission shortly after discharge is threatening to plague the quality of inpatient care. Readmission is a severe episode that leads to increased medical care costs. Federal regulations and early readmission penalties have created an incentive for healthcare facilities to reduce their readmission rates by predicting patients at a high risk of readmission. Scientists have developed prediction models by using rule-based assessment scores and traditional statistical methods, and most have focused on structured patient records. Recently, a few researchers utilized unstructured clinical notes. However, they achieved moderate prediction accuracy by making predictions of a single diagnosis subpopulation via extensive feature engineering. This study proposes the use of machine learning to learn deep representation of patient notes for the identification of high-risk readmission in a hospital-wide population. We describe and train several predictive models (standard machine learning and neural network), to which several setups have not been applied. Results show that complex deep learning models significantly outperform (P < 0.001) conventionally applied simple models in terms of discrimination ability. We also demonstrate a simple feature evaluation using a standard model, which allows the determination of potential clinical conditions/procedures for targeting. Unlike modeling using structured patient information with considerable variability in structure when different templates or databases are adopted, this study shows that the machine learning approach can be applied to prognosticate readmission with clinical free text in various healthcare settings. Using minimum feature engineering, the trained models perform comparably well or better than other predictive models established in previous literature.