The Center for Epidemiology Versus Arthritis at the University of Manchester in the UK is a leading medical research organization devoted to improving the lives of people living with musculoskeletal conditions through innovations in research methodology and clinical practice. Through their collaboration with National Health Service (NHS) regional hospitals and the Department of Computer Science, they have developed innovative methods for unlocking the power of clinical and epidemiological data in order to improve the lives of patients living with musculoskeletal diseases.
In an effort to protect extremely vulnerable populations during the COVID-19 pandemic, the UK’s Scientific Committee issued social distancing measures to “shield” these groups from infection. In response, national scientific societies released a risk stratification tool to guide clinicians in identifying vulnerable patients based on their immunosuppressive therapies and diagnoses. This exercise resulted in correctly identifying 2 million people in the UK who qualified for shielding in a very short amount of time. It was a challenging task, however, as certain types of information needed to identify such patients were only available in semi-structured data within hospital clinic letters, which required the manual review of all patients.
Researchers at the University of Manchester sought to study how a natural language processing (NLP) approach could help clinicians identify patients for shielding in the Rheumatology Department at a local hospital. This approach would use text-mining to extract diagnoses and medications from the free-text portion of semi-structured outpatient letters. The extracted information would then be fed into an algorithm to determine whether a particular patient should be recommended for shielding. The accuracy of this algorithm would then be compared to manual review by rheumatologists.
To automate their algorithm, diagnoses would have to be linked to a standard terminology (SNOMED CT®). In theory, NLP could automatically link diagnosis terms to the right codes, but the group would have to contend with the fact that clinicians often vary in the terminology and level of specificity that they use to document diagnoses.
Researchers therefore needed a way to determine which diagnosis terms should be linked to which codes. Additionally, they needed to be able to preserve the level of specificity present in complex diagnostic concepts when mapping to SNOMED CT to ensure that only the most appropriate codes would be used to drive their shielding algorithm.
The research group had already experienced success using IMO’s terminology services to underly their NLP-based research. They were therefore eager to continue collaborating with IMO to develop their COVID-19 shielding algorithm. In particular, the Manchester group wanted to leverage IMO’s robust and highly-granular clinical interface terminology to drive automatic diagnostic term coding.
IMO’s precise problem terminology has over 1.6 million diagnostic terms mapped to 36 reference code systems, including SNOMED CT, which ensures that every diagnosis term has comprehensive mapping. Through the use of over 5 million continuously maintained and expertly vetted proprietary lexicals, IMO can easily handle multiple variations in clinical terminology used to document diagnoses.
Initial results were very promising, with the research group’s screening algorithm detecting almost 90% of patients that had been manually identified by rheumatologists. This high rate of recall suggests that their algorithm was sensitive enough to identify those patients who would benefit most from shielding, which is critical when trying to reach out to vulnerable patients during a rapidly evolving pandemic.
“The development and deployment of such an automated algorithm has several advantages,” said lead author and Clinician Scientist, Meghna Jani, PhD, “including reducing the time needed to manually review a large number of records in order to make informed, timely decisions.”
Recognizing the important role that IMO’s clinical interface terminology played in their research, Professor of Computer Science, Goran Nenadic, PhD, said that by “embedding a terminology server like IMO within our NLP application, we were able to bridge the gap between our text-extraction process and running our screening algorithm.” Dr. Nenadic further explained that “without IMO, we would have had to rely on a much more time-consuming code look-up process which could have also ended up being less accurate.”
The work between IMO and the University of Manchester highlights the potential
of IMO’s clinical interface terminology to enable cutting-edge NLP that could lead
to much-needed clinical decision-making tools. These tools could then be used to help identify vulnerable patient-cohorts for proactive measures in response to public health emergencies.
The group at the University of Manchester is excited to continue working with IMO to drive further applications of their NLP research. Some future projects include:
Adapting their platform for other clinical use cases requiring rapid risk stratification of patients for public health initiatives such as vaccination roll outs
Broadening text-extraction methods to include free-text, unstructured narrative in clinical notes
Developing tools to describe different disease prevalence rates for different populations using documented diagnoses
For more on how a robust clinical terminology can help improve NLP, click here.
SNOMED and SNOMED CT are registered trademarks of SNOMED International.