Opioid overdoses: language models trained on diagnostic history identify one in four at-risk patients within the top 5 per cent

OverdoseMoE combines several models adapted to ICD codes to estimate the risk of overdose over a six-month period based on one year’s worth of medical records.

Can an opioid overdose be anticipated based solely on a patient's diagnosis codes? A team of researchers continued the pre-training of two models, one based on the Mamba architecture (OODMamba) and the other on Qwen (OODQwen), on longitudinal sequences of ICD codes, before specializing them in the prediction of an overdose within 180 days of a year of history.

This adaptation to the language of diagnostics systematically beats generalist models: OODQwen achieves an AUROC of 68.56 and an AUPRC of 24.47. The authors go further with OverdoseMoE, which makes models of different sizes work together by weighting their opinions according to complementary strategies. The device goes up to an AUROC of 69.49 and an AUPRC of 25.17.

In practice, among the 5% of patients considered to be most at risk, about one in four (positive predictive value of 25.38%) actually overdosed, a concentration well above the baseline rate. The results are based on an independent cohort drawn from MIMIC-IV, a sign that the approach is being transposed from one hospital system to another.