Adapted Machine Learning to Map Outpatient Neurological Conditions from a Large United Kingdom Urban Conurbation, between 2018 and 2024.
Keira Markey1, Rohan Ahmed2, David Rog2, Monty Silverdale3, James Lilleker4, Rajiv Mohanraj2
1University of Lancaster, 2Salford Royal Hospital, 3Dept of Neurology, 4University of Manchester
Objective:
To extract structured diagnostic data from free text outpatient neurology letters in Greater Manchester, United Kingdom (UK) and link to sociodemographic data to evaluate for potential health disparities. 
Background:
Diagnostic coding in outpatients of the United Kingdom National Health Service is currently poor. Clinical letters hold substantial information but are either unstructured or semi-structured. Natural language processing (NLP) and machine learning provide opportunity to extract clinical data into a structured format for analysis.
Design/Methods:

An open-source NLP tool, MedCAT, was integrated into an automated pipeline, and used to extract structured, retrospective diagnostic data from semi-structured neurology clinic letters in Greater Manchester, United Kingdom, between 1 January 2018 and 1 November 2024.

Results:

Successfully extracted diagnostic data were coded and linked to sociodemographic data for 125,273 unique neurology outpatients from a large UK conurbation. Headache (16.1%, n=26631) and epilepsy (14.3%, n=24880) were the commonest diagnoses made. Females residing in areas of highest social deprivation (IMD1) were more likely to be diagnosed with functional neurological disorder (ASRR[95% CI]: 1.78[1.73-1.83]) and headache (ASRR 95% CI]: 1.64[1.61-1.68] or if male with epilepsy (ASRR[95% CI]: 1.36[1.32-1.39]). In demyelinating conditions, females from lower social deprivation (IMD5) were seen in neurology clinics more (ASRR[95% CI]: 1.34[1.23-1.45]). Ethnicity was not recorded for 16.5% (n=17523), but people of Asian, Black and Mixed ethnicities had a lower likelihood of a neurology clinic compared to White people.

Conclusions:

Machine learning pipelines can be used to perform automated structural coding of outpatient data. For predominantly clinic-based specialities such as neurology this can be used to: identify diagnostic patient groups and explore potential health disparities. Such rich data, not currently available to decision makers, could support service planning, resource allocation and population health research.

10.1212/WNL.0000000000216123
Disclaimer: Abstracts were not reviewed by Neurology® and do not reflect the views of Neurology® editors or staff.