138 Done in 16 Seconds: extracting research ready data from NHS clinical letters

Abstract

Background Letters dictated by healthcare professionals in routine patient care form an invaluable dataset but are difficult to access and interpret. The UK MS Register (UKMSR) previously outlined (ABN2019) usage of Natural Language Processing Algorithms (NLP-A) to harvest and transform written language into analysable data in databases. We expanded the variables captured, increased the number of donating hospitals and compared the results to the previous NLP-A. Aim Apply the new NLP-A to a random letter selection and evaluate output and results. Methods A random, seeded, selection algorithm chose 100 letters from a corpus of 2690 consented in/outpatient letters from 13 Trusts. Letters were reviewed by human domain experts for Date Of Birth, NHS Number, Gender, Clinic Date, Postcode, MS Type and Expanded Disability Status Score (EDSS). NLP-A was applied and assessment made against the same variables. Results Run time was 15.4s, Sensitivity and Specificity > 98% in all cases except Clinic Date (Sensitivity 87%, Specificity 20%), MSType (Sensitivity 84%, Specificity 98%) and Postcode (Sensitivity 100%, Specificity 66%). Low specificity in Clinic Date illustrates disagreement on criteria between reviewer and NLP-A. These results represent a 5% increase (in common variables) in Sensitivity and Specificity over the 2019 algorithm. Conclusion We have improved the ability to accurately and rapidly identify required variables from the UKMSR minimum dataset using NLP-A. We are continuing to implement this on a widespread basis in the UKMSR.

Ähnliche Arbeiten