Design and implementation of a natural language processing system at the point of care: MiADE (medical information AI data extractor).

BACKGROUND: Well-organised electronic health records (EHR) are essential for high quality patient care, but EHR user interfaces can be cumbersome for entry of structured information, resulting in the majority of information being in free text rather than a structured form. This makes it difficult to retrieve information for clinical purposes and limits the research potential of the data. Natural language processing (NLP) at the point of care has been suggested as a way of improving data quality and completeness, but there is little evidence as to its effectiveness. We sought to generate such evidence by developing an open source, modular, configurable NLP system called MiADE, which is designed to integrate with an EHR. This paper describes the design of MiADE and the deployment at University College London Hospitals (UCLH), and is intended to benefit those who may wish to develop or implement a similar system elsewhere. RESULTS: The MiADE system includes components to extract diagnoses, medications and allergies from a clinical note, and communicate with an EHR system in real time using Health Level 7 Clinical Document Architecture (HL7 CDA) messaging. This enables NLP results to be displayed to a clinician for verification before saving them to the patient's record. MiADE utilises the MedCAT library (part of the Cogstack family of NLP tools) for named entity recognition (NER) and linking to SNOMED CT, as well as context detection. MedCAT models underwent unsupervised and supervised training on patient notes from UCLH, achieving precision of 83.2% (95% CI 77.0, 88.1), and recall of 85.2% (95% CI 79.1, 89.8) for detection of diagnosis concepts. In simulation testing we found that MiADE reduced the time taken for clinicians to enter structured problem lists by 89%. We have commenced a trial implementation of MiADE at UCLH in live clinical use, integrated with the Epic EHR at UCLH. CONCLUSIONS: We have developed an open source point of care NLP system and successfully integrated it with the EHR in live clinical use at a major hospital. Simulation testing has shown that our system significantly reduces the time taken for clinicians to enter structured diagnosis codes.