Optimisation of natural language processing for real-time structured clinical data capture in electronic health records

Information stored in electronic health records (EHRs) can play an important role in supporting clinical decision making (for example, it can help clinicians select the most appropriate medication for a patient), resulting in improved quality of patient care in the NHS. Much of the information in EHRs is recorded as "free text", that is, in ordinary language without any restriction on format, as this is the natural way in which people communicate. Although computers can be used to interpret free text, they cannot always get it right. However, if a standardised, structured format was used for recording information from the outset, this problem could be avoided. However, it can be very time-consuming and cumbersome for clinicians to enter the information in a structured way. This can mean that information is incomplete, or that clinicians are so busy on the computer that they do not have time to listen to their patients.

We are currently developing a system called "MiADE" which can analyse text entered by clinicians; this method is called "natural language processing" (NLP). We are testing out the MiADE system within a specific EHR system called Epic at University College London Hospitals (UCLH). The scope of the current work, which is funded by the National Institute of Health Research, is to extract just basic information about diagnoses. The information extracted includes the diagnosis code, and whether a diagnosis is confirmed, suspected or negative. More detailed information such as diagnosis date, cause, manifestations and evidence cannot be handled by Epic's existing user interface. This limits the potential benefits of the current system.

The aim of this new project is to improve the design of artificial intelligence systems to make it easier for clinicians to record information and thus support clinical decision making. The project will extend the scope of the MiADE system, with the aim to enable future systems to be as useful, effective and easy to use as possible.

The first objective is to develop more advanced NLP methods that can capture more details about diagnoses. We will develop new NLP methods to extract information about date, cause, manifestations and evidence for a diagnosis from free text. We will use EHR data from UCLH to develop and test the NLP methods. We will then use the new methods to study patients with fluid overload due to impaired heart function ('heart failure') in UCLH data. We will extract information from clinical notes about the subtype and cause of their heart failure, and the severity of their symptoms.

The second objective is to create and test an experimental user interface to enable clinicians to interact with the NLP system more easily. The user interface will allow clinicians to enter new information about diagnoses in a structured way, and integrate it with information already in the record. We will enable clinicians to test the user interface with simulated patients to ensure that it is easy to use.

The third objective is to find out how necessary it is for NLP systems to 'learn' from local data provided by the healthcare environment in which they are going to be used. We will compare the performance of different NLP systems on patient records from Great Ormond Street Hospital (GOSH), a specialist children's hospital. We will compare the performance of an NLP system trained on GOSH data, an NLP system trained on UCLH data, and a commercial NLP system that was developed without access to hospital data.

Overall, this project will provide an evidence base for improving the way that EHR systems can use NLP to make it easier for clinicians to record detailed information at the point of care. This will support the wider adoption of NLP integrated within EHR systems, resulting in improved patient safety and quality of care. It will also improve the usefulness of health records for research, which will benefit future patients.