ABSTRACT Introduction Timely surveillance of opioid‐related harm is critical to inform public health responses and policy evaluation. In Australia, where prescription and illicit opioids remain a leading cause of unintentional drug‐induced deaths, emergency departments (ED) are a vital point of contact for acute opioid poisonings. Existing surveillance systems rely on structured coding, yet much relevant information is recorded in free‐text fields, leading to underreporting or misclassification. This limits opportunistic identification of emerging patterns and weakens the evidence base for evaluating policy reforms. We aim to improve surveillance accuracy by applying natural language processing (NLP) to routinely collected ED data. Methods Using medical concept annotation tools, we will develop models trained on 15 years of Victorian Emergency Minimum Dataset (VEMD) records. These models will analyse both unstructured and structured fields to identify opioid poisoning presentations and be validated against a manually coded gold standard using standard performance metrics. In the second phase, we will incorporate additional unstructured clinical information such as discharge summaries from hospital electronic records, which are not available in the VEMD data, thereby allowing more comprehensive and accurate classification. Finally, we will assess the utility of NLP‐enhanced data in evaluating three major opioid policy changes. Discussion and Conclusions This study is the first to apply NLP at large scale to Australian ED data for drug poisonings. By improving the accuracy and consistency of opioid poisoning identification, this approach can strengthen routine surveillance and better inform timely policy and health system responses without increasing the workload for clinical staff.
Xia et al. (Sun,) studied this question.