We present a machine reading system for electronic medical record (EMR) question/answering (Q/A) developed using state-of-art deep learning models . Significant advantages of the system include its ability semantically map questions with candidate answers, only provide answers that are present in the medical records text, and not require re-training in new health care domains. The system was developed to translate colonoscopy pathology reports via a series of clinical questions to a form understandable by patients. As more medical records are made available via patient portals, such systems are needed to provide comprehensible and actionable knowledge to patients. We see the potential of using the underlying Q/A system to generate targeted summaries, refine records search, and facilitate EMR data mining. Accuracy exceeding 90% was achieved answering 12 clinical questions on a small test set (10) of colonoscopy pathology reports.
Describe the new knowledge and additional skills the participant will gain after attending your presentation.: Description of the problem
The ability to extract targeted knowledge from medical record text is one of the most significant challenges facing healthcare. Clinicians and ancillary personal are overwhelmed with medical records related data. Clinicians inability to process and comprehend the knowledge captured in the clinical text can significantly impact patient care.
Patients seeking access to their records are unable to comprehend them.
Current NLP systems are highly feature engineered, application specific, and difficult to adapt . We propose a different approach based on deep models of general language understanding and adapting these models to healthcare domains.
Jay Urbain (Presenter)
Milwaukee School of Engineering
Bradley Crotty, Medical College of Wisconsin