Identifying QT prolongation from ECG impressions using Natural Language Processing and Negation Detection

Electrocardiogram (ECG) impressions provide significant information for decision support and clinical research. We investigated the presence of QT prolongation, an important risk factor for sudden cardiac death, compared to the automated calculation of corrected QT (QTc) by ECG machines. We integrated a negation tagging algorithm into the KnowledgeMap concept identifier (KMCI), then applied it to impressions from 44,080 ECGs to identify Unified Medical Language System concepts. We compared the instances of QT prolongation identified by KMCI to the calculated QTc. The algorithm for negation detection had a recall of 0.973 and precision of 0.982 over 10,490 concepts. A concept query for QT prolongation matched 2,364 ECGs with precision of 1.00. The positive predictive value of the common QTc cutoffs was 6-21%. ECGs not identified by KMCI as prolonged but with QTc>450ms revealed potential causes of miscalculated QTc intervals in 96% of the cases; no definite concept query false negatives were detected. We conclude that a natural language processing system can effectively identify QT prolongation and other cardiac diagnoses from ECG impressions for potential decision support and clinical research.