Natural language processing and geospatial analysis of 50,847 individuals to identify ethnicities and neighbourhoods with high risk of acute valvular decompensations
N Nwegbu, C Oh, N Ahluwalia, S Law, C Hayward, S Williams, C Zhong, G Lloyd, A JoshiAbstract
Background
Acute valvular decompensations (AVD) in patients without previously diagnosed heart valvular disease (HVD) occurs over 100 times a year in our hospital network[1]. Delayed presentation may be influenced by social and environmental determinants of health (SEDOH)[2]. Natural Language Processing (NLP) techniques, including Named Entity Recognition (NER) have advanced significantly, allowing analysis of echocardiogram free-text reports at scale and to a high degree of accuracy. Geospatial statistical methods can integrate individuals’ clinical, social and geographic exposures[3].
Purpose
To identify geosocial predictors of AVD, identfying groups that may benefit from targeted echocardiography to detect early disease.
Methods
Regex, Long-Short Term Memory[4] [LTSM] and Large Language Models[5] [LLM] approaches to NER were validated in a set of 1000 expert-annotated echocardiography reports. The most accurate inference technique analysed echocardiogram reports from 50,847 individuals. AVD was defined as having a first-ever echocardiogram during an acute admission to hospital, and undergoing valvular intervention within a year. Logistic regression modelled individual-level risk for severe HVD. EHR demographic data and two environmental exposures (PM10 and neighbourhood greenspace proportion) were integrated based on the patients’ Lower Layer Super Output Area (LSOA) information. Fitted probabilities from the individual-level logistic regression models were aggregated to the LSOA level to derive a mean predicted risk for each area.
Results
RegEx achieved the highest accuracy in identifying HVD in our validation set (F1 score 0.90 vs. LSTM 0.83 vs.LLM 0.83). Of 50,847 echocardiograms analysed, 1973 patients were identified as having severe HVD. Of these, 172 were acute, undiagnosed valvular decompensations (76 aortic stenosis, 41 aortic regurgitation, 72 mitral regurgitation, 7 mitral stenosis). 1801 had chronic, severe valvular disease (792 aortic stenosis, 380 aortic regurgitation, 822 mitral regurgitation, 99 mitral stenosis). Being male (OR 1.8, CI 1.31-2.51, p = 0.0003) or “Black-Any other” or “Asian or Asian British–Bangladeshi” ethnicity (OR 2.46, CI 1.05-5.75, p = 0.04; OR 1.75, CI 1.10-2.78, p = 0.02), was associated with increased risk of AVD. Figures 1A&B show the regional relative mean risk of AVD and chronic severe HVD by LSOA, Figure 1C shows the normalized difference in risk across regions (Moran’s I value 0.305 (p-value < 0.001), indicating that the risk of AVD is significantly different to the background prevalence of chronic valvular disease.
Discussion
AVD does not follow the same geographic pattern as chronic disease. Patients of “Black-Other” and “Asian-Bangladeshi” ethnicity were more likely to experience AVD. We identify regions where targeted diagnostics could add value. These efforts should ensure cultural congruence with the relevant patient group.Figure 1