Using artificial intelligence to identify drowning risk and protective factors present in media reports in metropolitan Houston, Texas

BACKGROUND: Natural language processing (NLP) can process unstructured text to provide information on injuries but has not been used for drowning. We tested the effectiveness of NLP in identifying risk and protective factors for drowning present in media and police reports. METHODS: This was a cross-sectional study of persons of all ages who unintentionally drowned in metropolitan Houston, Texas between 2016 and 2022. Drowning incident data were obtained from media reports (NexisUni using keywords 'drown' and 'submersion' to access online media archives in print blog, video and audio record formats) and from redacted police reports. Temporal, demographic and incident narratives were obtained manually. Information on swimming ability, supervision, alcohol use and life jacket use was extracted from report narratives using OpenAI's GPT-4o models which converted unstructured narratives into analysable structured data. Extracted data were validated against manually reviewed ground truth data in 20% of the cases. NLP was assessed for accuracy, precision, recall and F1 score. RESULTS: There were 190 reports of drowning (142 (74.7%) fatal and 48 (25.3%) non-fatal) between 2016 and 2022. The median age was 16 years (IQR: 5-41). Children aged 0-17 years constituted 45.8% of the cohort; 125 (65.8%) were males. Most drownings occurred in natural water (65.8%); the rest were in confined water (30.5%). The NLP metrics were recall (0.6-1.0), precision (0.7-1.00), F1-score (0.6-1.0) and accuracy (0.9-1.00). CONCLUSION: NLP is a novel method of extracting information on drowning risk and protective factors from media report narratives. This can be used to study the epidemiology of drowning.