Extraction of Clinical Timeline from Discharge Summaries using Neural Networks

Abstract

Discharge summaries are a concise representation of the most important bits of information about a patient’s time in the hospital. Converting the free-text into a clinical timeline can facilitate accurate assimilation of information by physicians and the structured data can be used to populate knowledge bases, in clinical decision support systems, etc. Conventional methods for temporal evaluation of discharge summaries employ structured inference and extensive feature engineering. However, they also run the risk of overfitting to the training domain and thus, not being efficient in deployment. Novel methods of natural language processing leverage semantics from large corpuses and produce results with minimum feature engineering. This work explores the use of neural network architectures in clinical entity recognition and temporal evaluation. Recurrent neural networks are found to perform at par with conditional random field systems in clinical entity recognition, scoring 94.04% on the i2b2 2012 dataset. Moreover, they perform better for under-represented entity classes like ‘Occurrence’, ‘Evidential’ and ‘Clinical Department’ in a skewed dataset. The out-of-domain evaluation of conditional random fields and neural networks has favorable results on a corpus of ER visit, progress, consult and ICU notes from various medical centers. Neural networks are more agreeable to domain adaptation. This work also explores the use of convolutional neural nets for extraction of within-sentence temporal relations. Preliminary results show that convolutional networks might not be well suited to the task.