Assessing the Generalizability of a Phonetic Text Clustering Method for Mining Clinical Data from the Electronic Health Records of Patients with Post-COVID-19 Syndrome (Preprint)

BACKGROUND Natural language processing (NLP) offers opportunities to extract meaningful information from unstructured clinical texts, but most methods rely on computationally expensive models and English-language datasets. We previously developed a phonetic text clustering (PTC) methodology to extract clinical information related to COVID-19. This study assesses its performance and generalizability in a heterogeneous dataset from a tertiary hospital in Brasília, Brazil. OBJECTIVE To assess the performance and generalizability of PTC methodology for extracting clinical information related to post-COVID-19 syndrome from unstructured electronic health records in a heterogeneous dataset from a tertiary hospital in Brazil. METHODS We conducted a cross-sectional validation study using electronic health records (EHRs) from 414 hospitalized COVID-19 patients (46,508 clinical annotations). PTC outputs were compared with manual annotation of 230 records by five clinical specialists. Performance was evaluated using accuracy, sensitivity, specificity, positive predictive value (PPV), and F1 score. Extracted clinical information was analyzed across temporal windows before, during, and after hospitalization. RESULTS PTC demonstrated high overall performance (accuracy 99%, sensitivity 93%, specificity 99%, PPV 89%, F1 score 91%). Most conditions and symptoms showed consistently strong metrics, with exceptions such as neurological alterations and heart disease. Hypertension, diabetes, and chronic kidney disease were the most frequently affirmed conditions. Fatigue, cough, weakness, headache, and muscle pain frequently persisted beyond 90 days, while loss of taste and smell were restricted to the acute phase. CONCLUSIONS PTC offers a transparent, low-cost, and scalable approach for extracting clinical information from an unstructured EHRs, particularly suitable for low-resource settings. Its strong performance and ability to characterize temporal symptom patterns support applications in monitoring and understanding post-COVID-19 syndrome.