The VA faces rising prevalence of PTSD among its veterans from US conflicts in Iraq and Afghanistan, and a consistently back-logged inventory of claims from veterans seeking care, compounded by the expanded PTSD eligibility conditions entitling veterans to VA health services. Temporally sensitive surveillance of the sequelae of post traumatic stress disorder (PTSD) and changes resulting from treatment interventions is a prerequisite to evaluating the efficacy of PTSD treatment programs. Assisted clinical surveillance of the vast amount of multi- dimensional historic data stored in EHRs has the potential to support physicians in the diagnosis and treatment of their patients and clinical researchers seeking patterns of events and their timing. A surveillance model of the natural history of PTSD could move clinical understanding of the disorder beyond the atemporal check-list of symptoms to identifying patterns in symptom progression. Information for a temporally oriented analysis of a patient's PTSD care is likely captured in clinical notes prepared by VA clinicians. However, a system of text processing with a temporal component to gather the information relevant to the care of veterans with this clinical problem does not currently exist. A barrier to medically relevant interpretation of documented events has been the absence of a comprehensive system for capturing the contextual features necessary. The Natural Language Processing developments proposed here are to erect the framework for extracting information pertinent to the timing and sequence of PTSD symptoms, treatments and changes from both narrative documents and structured data elements within a patient's EHR in order to implement a patient- oriented event database, queriable for select event sequences. A tool that delivers a useable time-line of events will make it possible to query the patient-level database on the basis of temporal patterns of selected clusters of clinical or personal history events. We propose methods to increase the number of events anchored by a time reference. Our methods include using separate occurrences referring to same event in order to acquire time-stamps for a greater number of events, and strategies for decreasing interval lengths that constrain the temporal interpretation of unanchored events, and using both of these to increase the precision of the event sequence assignments. We propose three areas for discovering instances of co-referring events: 1) local document level co-reference occurrences, 2) narrative document events referencing structured data events, 3) cross-document narrative event co-references. In addition, we propose database update and extend criteria to iteratively populate a relational database of temporally ordered events. Our proposed method of processing new sets of events with each iteration of document analysis is designed to improve the accuracy of both the anchoring and sequencing processes by inserting time-stamped events into event sequences which are anchored only by time constraint intervals. With these development pieces in place, we will be equipped to provide detailed symptom and treatment chronologies for patients under care for PTSD in VA facilities. We will test the validity and accuracy of 60 such patient chronologies by comparing the sequences of events generated by our system called the Med-TARSQI Chronology to humanly constructed chronologies of the 60 cases. We will also test and evaluate performances of each of the document level NLP modules in the derivation of the chronologies delivered by the Med-TARSQI Chronology.
Tracking a veteran's PTSD symptoms and treatments over time is currently a matter of reading through narrative text to assemble a working time-line in the clinician's mind. We have developed Med-TTK which is a temporal reasoning NLP tool for ordering relevant events contained within the EHR. The tool development is in concert with work being carried out in multi-site Consortium of Health Informatics Research (CHIR) research on extracting and time-stamping PTSD symptom and treatment mentions contained within single documents of medical narrative text. The proposed work will link across multiple patient narrative documents to construe sequences of PTSD symptom occurrences and treatments over time within a queriable event database. This research designed includes the means for constructing and maintaining a comprehensive medical chronology across medical episodes of interest by text mining the EHR. The chronology will be composed of sequences of events contributing to PTSD, reported PTSD symptoms, interventions designed to treat those symptoms.