Prediction Using Note Text: Synthetic Feature Creation with word2vec

word2vec affords a simple yet powerful approach of extracting quantitative variables from unstructured textual data. Over half of healthcare data is unstructured (1) and therefore hard to model without involved expertise in data engineering and natural language processing. word2vec can serve as a bridge to quickly gather intelligence from such data sources. In this study, we ran 650 megabytes of unstructured, medical chart notes from the Providence Health & Services electronic medical record through word2vec. We used two different approaches in creating predictive variables and tested them on the risk of readmission for patients with COPD (Chronic Obstructive Lung Disease). As a comparative benchmark, we ran the same test using the LACE risk model (2) (a single score based on length of stay, acuity, comorbid conditions, and emergency department visits). Using only free text and mathematical might, we found word2vec comparable to LACE in predicting the risk of readmission of COPD patients.