Can Unified Medical Language System-based semantic representation improve automated identification of patient safety incident reports by type and severity?

OBJECTIVE The study sought to evaluate the feasibility of using Unified Medical Language System (UMLS) semantic features for automated identification of reports about patient safety incidents by type and severity. MATERIALS AND METHODS Binary support vector machine (SVM) classifier ensembles were trained and validated using balanced datasets of critical incident report texts (n_type = 2860, n_severity = 1160) collected from a state-wide reporting system. Generalizability was evaluated on different and independent hospital-level reporting system. Concepts were extracted from report narratives using the UMLS Metathesaurus, and their relevance and frequency were used as semantic features. Performance was evaluated by F-score, Hamming loss, and exact match score and was compared with SVM ensembles using bag-of-words (BOW) features on 3 testing datasets (type/severity: n_benchmark = 286/116, n_original = 444/4837, n_independent =6000/5950). RESULTS SVMs using semantic features met or outperformed those based on BOW features to identify 10 different incident types (F-score [semantics/BOW]: benchmark = 82.6%/69.4%; original = 77.9%/68.8%; independent = 78.0%/67.4%) and extreme-risk events (F-score [semantics/BOW]: benchmark = 87.3%/87.3%; original = 25.5%/19.8%; independent = 49.6%/52.7%). For incident type, the exact match score for semantic classifiers was consistently higher than BOW across all test datasets (exact match [semantics/BOW]: benchmark = 48.9%/39.9%; original = 57.9%/44.4%; independent = 59.5%/34.9%). DISCUSSION BOW representations are not ideal for the automated identification of incident reports because they do not account for text semantics. UMLS semantic representations are likely to better capture information in report narratives, and thus may explain their superior performance. CONCLUSIONS UMLS-based semantic classifiers were effective in identifying incidents by type and extreme-risk events, providing better generalizability than classifiers using BOW.