Word2Vec inversion and traditional text classifiers for phenotyping lupus

Clayton A. Turner, Alexander D. Jacobs, Cassios K. Marques, James C. Oates, Diane L. Kamen, Paul E. Anderson, Jihad S. Obeid

Identifying patients with certain clinical criteria based on manual chart review of doctors' notes is a daunting task given the massive amounts of text notes in the electronic health records (EHR). This task can be automated using text classifiers based on Natural Language Processing (NLP) techniques along with pattern recognition machine learning (ML) algorithms. The aim of this research is to evaluate the performance of traditional classifiers for identifying patients with Systemic Lupus Erythematosus (SLE) in comparison with a newer Bayesian word vector method. We obtained clinical notes for patients with SLE diagnosis along with controls from the Rheumatology Clinic (662 total patients). Sparse bag-of-words (BOWs) and Unified Medical Language System (UMLS) Concept Unique Identifiers (CUIs) matrices were produced using NLP pipelines. These matrices were subjected to several different NLP classifiers: neural networks, random forests, naïve Bayes, support vector machines, and Word2Vec inversion, a Bayesian inversion method. Performance was measured by calculating accuracy and area under the Receiver Operating Characteristic (ROC) curve (AUC) of a cross-validated (CV) set and a separate testing set. We calculated the accuracy of the ICD-9 billing codes as a baseline to be 90.00% with an AUC of 0.900, the shallow neural network with CUIs to be 92.10% with an AUC of 0.970, the random forest with BOWs to be 95.25% with an AUC of 0.994, the random forest with CUIs to be 95.00% with an AUC of 0.979, and the Word2Vec inversion to be 90.03% with an AUC of 0.905. Our results suggest that a shallow neural network with CUIs and random forests with both CUIs and BOWs are the best classifiers for this lupus phenotyping task. The Word2Vec inversion method failed to significantly beat the ICD-9 code classification, but yielded promising results. This method does not require explicit features and is more adaptable to non-binary classification tasks. The Word2Vec inversion is hypothesized to become more powerful with access to more data. Therefore, currently, the shallow neural networks and random forests are the desirable classifiers.