The New York City medical examiner's office trained two language models on 9,263 reports written by investigators after each death in 2022 and 2023. The stronger model identified suspected overdose deaths with 92% sensitivity and 88% positive predictive value. Specificity was 76%, and both models did better than the office's existing surveillance tool and logistic regression on the same reports.
Why it is interesting: Both models read free text better than the office's existing surveillance tool, though specificity was the weaker measure.
Cite this item
BibTeX
Free sign-up. A digest of the new papers twice a week, grouped by domain.