AquaAI: development and internal validation of a Danish transformer-based model to identify drowning and aquatic incidents in prehospital medical records

BACKGROUND: Effective prevention of drowning and aquatic incidents requires timely and accurate surveillance supported by high-quality validated data. In Denmark, the use of the free-text fields in the Danish Prehospital Medical Record has proven effective in identifying potentially relevant cases for such surveillance. While these free-text fields contain rich contextual information, manual screening of all records is impractical. This study aimed to develop and internally validate a Danish natural language processing pipeline for identifying drowning and aquatic incidents (AquaAI) from routine prehospital records and prioritizing records for final manual validation by medical experts. METHODS: This nationwide retrospective cohort study was conducted using Danish prehospital electronic medical records from 2016 to 2024. Medical records were first retrieved using the Danish Drowning Formula, an iteratively developed trigger-word search algorithm, and expert-labelled as Drowning, Aquatic incident, or Non-relevant. A Danish transformer-based language model was then fine-tuned for three-class sequence classification as the second-stage classifier in this workflow and embedded in a hybrid pipeline with rule-based safeguards. Model development used a temporal split with training data from 2016 to 2021 and validation data from 2022. Final performance was evaluated on a temporally separated hold-out dataset comprising records from 2023 to 2024. Primary outcomes were class-wise sensitivity and a binary Relevant (Drowning + Aquatic incident) versus Non-relevant analysis to quantify case finding and workload reduction. RESULTS: The dataset comprised 40,876 medical records retrieved with the Danish Drowning Formula. On the hold-out dataset, AquaAI identified 91% of relevant cases and correctly filtered 89% of non-relevant records from manual review. Overall, this reduced the number of records requiring manual review by 77.5%. In the three-class analysis, sensitivity was 84% for Drowning, 83% for Aquatic incident, and 89% for Non-relevant. Only one drowning case (0.3%) was classified as non-relevant. CONCLUSIONS: A hybrid transformer-based pipeline operating as a second-stage classifier within a two-step retrieval workflow can identify drowning and aquatic incidents from prehospital free-text narratives with high case finding and substantial reduction of manual review. This approach may support national surveillance and targeted prevention initiatives.