Validation of COPD etiotypes by AI-text mining of electronic health records

<bold>Introduction:</bold> A Lancet Commission paper (Stolz et al., 2022) proposed a taxonomic classification for COPD: "Genetic", "Infection", "Early-life events", "Tobacco" and "Environmental exposure". The definition of these etiotypes is based on clinical ground and may result in new paradigms for the management and treatment of COPD. However, so far these etiotypes are neither evidence-based nor validated. <bold>Aim:</bold> We saught to identify and validate these COPD etiotypes with data from electronic health records (EHRs) in two independent Swiss lung centres. <bold>Methods:</bold> EHRs from 2 COPD cohorts were extracted including 356 in- and out-patients from the University Hospital Basel and 211 out-patients with stable COPD at the Cantonal Hospital St. Gallen. The concordance between the textual information in the EHRs and the description of the COPD etiotypes was investigated using text mining and correspondence analysis. <bold>Results:</bold> The etiotype term usage and predictions from both cohorts are shown in Figure 1. The etiotype distribution differed significantly between the 2 cohorts with most patients classified as "Infection" in the first cohort compared to the second (67% vs. 40%, p<0.001) whereas the "Smoking" etiotype was more prevalent in the second cohort compared to the first (34% vs. 12%, p<0.001). <bold>Conclusion:</bold> Etiotype prediction from EHRs is feasible and reveals the multidimensional nature of COPD. Most patients expressed signals from several etiotypes. <fig><object-id>erj;64/suppl_68/OA4650/F1</object-id><object-id>F1</object-id><object-id>F1</object-id><graphic></graphic></fig>