A web-based tool utilizing machine learning algorithms for predicting illicit drug use in emergency departments

Background

Identifying illicit drug use through urine testing is time-consuming in the era of new psychoactive substances. This study aimed to develop a machine learning (ML) prediction model for early identification of illicit drug use in suspected emergency department (ED) patients.

Methods

Data from the Taiwan Emergency Department Drug Abuse Surveillance (TEDAS) database (2020–2023) was used. Six feature categories—demographics, triage data, referral source, symptoms, physical findings, and clinical characteristics—were included. The primary outcome was positive urine results for illicit drugs, confirmed by liquid chromatography-tandem mass spectrometry. Data were divided chronologically into training/validation and testing sets. Three supervised ML algorithms, including random forest, CatBoost, and light gradient boosting machine, were tested using K-fold cross-validation, and performance was evaluated by the area under the receiver operating characteristic curve (AUC) in the test set.

Results

The analysis included 13,615 urine test results from ED cases, identifying 3,185 positive cases (23.4%). A total of 9,529 cases (2020–2022) formed the training/validation cohort, and 4,086 (2023) were used for testing. Twenty features were used to construct the prediction model. The CatBoost classifier performed best, achieving an AUC of 0.846 (95% confidence interval [CI]: 0.831–0.859) in the testing cohort. A web-based tool and mobile apps were implemented to assist emergency physicians in predicting illicit drug use.

Conclusions

The machine learning model effectively predicts illicit drug use in ED patients and has been successfully implemented for free access. Further analysis is needed to assess post-implementation performance and its potential for use in other countries.