Temporal information extraction with the scalable cross-sentence context for electronic health records

Temporal information is essential for accurate understanding of medical information hidden in electronic health record texts. In the absence of temporal information, it is even impossible to distinguish whether the mentioned symptom is a current condition or past medical history. Hence, identifying the relationship between medical events and document creation time (DCT) is a critical component for medical language comprehension, which can link the mentioned medical information to the time dimension by marking temporal tags. Existing natural language processing (NLP) systems are typically based on the sentence where the medical event is located to extract the DCT relationship. Inevitably, the limited textual context can be insufficient as it is difficult to contain adequate document information. Introducing the surrounding sentences into models is a fitting way to enrich the information. However, in addition to document information, the added context can also bring noise to confuse the models. For effective utilization of the context, we design the DCDR (Dynamic Context and Dynamic Representation) model. Our model consists of two modules, i.e. the dynamic context mechanism and dynamic representation mechanism. The dynamic context mechanism is employed to bring the related texts into our model via the sliding windows and a scoring calculation. For the dynamic representation mechanism, a modified dynamic routing algorithm is adopted to filter the noise and generate an integrated representation for the whole context. Besides, the mentioned medical information is led into the routing process to enhance the dynamic representation module. The experiments show that our proposed model achieves improvement over existing models and achieves an F-score of 85.7% on the commonly used THYME corpus.