A neural network approach to predict opioid misuse among previously hospitalized patients using electronic health records

PLoS One. 2024 Aug 28;19(8):e0309424. doi: 10.1371/journal.pone.0309424. eCollection 2024.

Abstract

Can Electronic Health Records (EHR) predict opioid misuse in general patient populations? This research trained three backpropagation neural networks to explore EHR predictors using existing patient data. Model 1 used patient diagnosis codes and was 75.5% accurate. Model 2 used patient prescriptions and was 64.9% accurate. Model 3 used both patient diagnosis codes and patient prescriptions and was 74.5% accurate. This suggests patient diagnosis codes are best able to predict opioid misuse. Opioid misusers have higher rates of drug abuse/mental health disorders than the general population, which could explain the performance of diagnosis predictors. In additional testing, Model 1 misclassified only 1.9% of negative cases (non-abusers), demonstrating a low type II error rate. This suggests further clinical implementation is viable. We hope to motivate future research to explore additional methods for universal opioid misuse screening.

MeSH terms

  • Analgesics, Opioid* / adverse effects
  • Electronic Health Records*
  • Hospitalization
  • Humans
  • Male
  • Neural Networks, Computer*
  • Opioid-Related Disorders* / epidemiology

Substances

  • Analgesics, Opioid

Grants and funding

The author(s) received no specific funding for this work.