Alcohol Status Standardization from Clinical Real World Data with Transformer Architectures

Most electronic health records (EHRs) lack structured and standardized documentation for lifestyle-related risk factors. In this work, we address this problem by showcasing methods to standardize the reporting of the alcohol consumption status of patients by applying a state-of-the-art transformer architecture on an anonymized clinical dataset in German via a multi-class classification schema, using the terminology standard SNOMED CT. We report on this task, leveraging a pre-trained transformer model in German with a macroaveraged F1-measure of 0.85, taking into account various context lengths of the source under investigation. The model performs slightly better than a fastText baseline model, reaching a macroaveraged F1-measure of 0.82. We identified four error types that substantially influence performance metrics: negations, single-word compounds, abbreviations, and multilingual expressions, particularly Latin ones.