Patients' electronic medical records contain diagnostic codes, but they can also include information such as notes, test results, or billing data that indicate hazardous alcohol use.
To obtain contextual clues, the researchers programmed a natural language processing model to identify diagnostic codes and other indicators of risky alcohol use, such as drinks exceeding the recommended limits per week or a history of medical problems related to alcohol misuse.
Alcohol abuse during a surgical procedure is associated with higher infection rates, longer hospital stays, and other surgical complications. Of the patients studied, 4.8 percent had charts that included a diagnosis code related to alcohol use. With the help of contextual clues, the sample classified more than three times as at-risk, a total of 14.5 percent.
The model did as well as a panel of human alcohol use experts, matching their classifications to a subset of 87 percent of recordings.
The findings point to AI as a potential partner for doctors looking to identify patients in need of intervention or post-operative support, the researchers concluded.
The study “will lay the foundation for efforts to identify other risks with proper validation in primary care and beyond,” said VGvinoth Vaitheswaran, associate professor of health sciences learning at the University of Michigan Medical School and lead author of the paper. A message liberation. “Essentially, it's a way for the presenter to highlight what's already in the notes that other presenters have made, without them having to read the entire record.”
The researchers say they plan to release the model publicly, but note that it will need to be trained on medical records from private facilities.
Auto-detection of risky alcohol use before surgery using natural language processing
Alcohol: clinical and experimental research