79 Using artificial intelligence and machine learning to automate injury prevention surveillance data
Bibliographic record
Abstract
Abstract 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 many surveillance systems, data is input in a manual fashion with humans screening all charts. Natural language processing algorithms using transformer models can analyze text from triage notes of previous CHIRPP-positive patients to train models to predict future CHIRPP patients. We sought to automate these functions using machine learning and artificial intelligence. Objectives The objective of this initiative is to automate detection of injured patients using artificial intelligence. Once charts are identified, a further summarization model will automatically generate a narrative of the events surrounding the injury, as required for the surveillance system. Design/Methods ED notes were obtained from an electronic health record for all patients visiting the ED from October 2018 to December 2021 (n =203626). All charts were manually reviewed to determine if patients met the criteria for CHIRPP. CHIRPP-positive patients made up 24.8% (n = 50499) of the total presentations. Patient charts were randomly split into train (70%) and validation (30%) datasets. Pre-trained distilbert-base-uncased model from Huggingface was utilized. The best model according to AUROC was selected for further development. For note summarization, we fine-tuned a multi-task model (t5-small) from the Huggingface repository. We used human summaries from the same dataset (October 2018-December 2021) as training data (49,078 note, summary pairs). A 70/30 train/validation split was performed. Results Our model captured 90% of all positive cases while reducing the total number of charts for review by 77%, 75% and 63% for train, validation and test sets respectively. See AUC curve (figure 1). Based on 2022 volumes this is a reduction to 24,027 from 64,938 charts. For note summarization, when we compared the cosine similarities of human summaries to actual note text vs t5-small summaries to note text we saw remarkable similarity implying high model performance. Conclusion Our model has shown that high volumes of charts that were previously screened manually can utilize natural language processing to automate the task with minimal sacrifices in terms of data capture and accuracy, dramatically reducing cost while improving speed to reporting. Our summarization model also showed potential for further automating this process. This application to an injury prevention program has wide scale applicability and scalability.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.003 | 0.011 |
| Meta-epidemiology (narrow) | 0.001 | 0.000 |
| Meta-epidemiology (broad) | 0.001 | 0.001 |
| Bibliometrics | 0.004 | 0.002 |
| Science and technology studies | 0.001 | 0.000 |
| Scholarly communication | 0.001 | 0.001 |
| Open science | 0.001 | 0.001 |
| Research integrity | 0.001 | 0.001 |
| Insufficient payload (model declined to judge) | 0.002 | 0.002 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".