O3203: Using Artificial Intelligence and Machine Learning to Automate Injury Prevention Surveillance

Background: The Canadian Hospitals Injury Reporting and Prevention Program (CHIRPP) is a pan-Canadian injury and poisoning surveillance system that tracks patients presenting to Canadian Emergency Departments (EDs). Like most injury surveillance systems, data is input in a manual fashion with humans screening physical or electronic notes. Our institution sees over 80,000 ED visits per year, and thus we sought to automate these functions using machine learning and artificial intelligence. Machine Learning natural language processing algorithms using transformer models can analyze text data from triage notes of previous CHIRPP-positive patients to train transformer models to predict future CHIRPP patients. This would minimize delays and allow for real-time reporting of injury data while also reducing resource expenditures.

Methods: Deidentified triage notes were obtained from an electronic health record system for all patients visiting the ED from January to December 2019 (n = 77,677). All charts were manually reviewed to determine if patients met the criteria for CHIRPP. CHIRPP-positive patients made up 20.1% (n = 15,638) of the total presentations. Patient charts were split into train (70%) and test (30%) datasets in ascending chronological order to best simulate prospective implementation of the model and enable reproducible results. Pre-trained sequence classification transformer models were obtained using the HuggingFace Library.

Results: Our model captured 98% of all positive cases while reducing the total number of charts for review by 52.3%. Based on 2019 volumes this is a reduction to 37,037 charts, down from 77,677. If capturing 90% of CHIRPP-positive cases was acceptable, our model can be adjusted to reduce the number of total charts for manual review from 77,677 to 21,749 (72% reduction). The model's performance can be adjusted based on desired sensitivities - to capture all cases, more charts would need to be manually reviewed (100% capture) vs less capture resulting in smaller false positives.

Conclusion: Our model has shown that high volumes of charts that were previously screened manually can utilize natural language processing to semi-automate the task with minimal sacrifices in terms of data capture and accuracy. This application to an injury prevention program has wide scale applicability and scalability to other health promotion and surveillance systems.