Identifying anaphylaxis using weakly-supervised prediction models and natural language processing

Researchers developed and tested an algorithm to identify anaphylaxis, a rare and acute condition that is difficult to accurately recognize using claims data alone. The algorithm combined data from electronic health records and insurance claims from two healthcare systems (Kaiser Permanente Washington and Vanderbilt University Medical Center). They used automated natural language processing to analyze clinical texts. The best-performing model achieved an AUC value of 0.931 at Kaiser Permanente Washington. The model's sensitivity for detecting anaphylaxis was very high, ranging from 0.78 to 1.0. The new approach was simpler and more cost-effective than previous methods because it did not require manual data curation. The algorithm can be easily deployed across different healthcare facilities.