University of Michigan researchers have produced a new prediction model using longitudinal information and deep learning to better predict the return to work time for people with occupational injuries.
The prediction models currently used today make predictions about return to work times based on initial time of injury. Physicians then determine how to most efficiently distribute resources to ensure that everyone is recovering and returning to work in the most effective way possible. This puts a large strain on individuals in the healthcare system and provides lagging results.
“We created a model that can take in information daily and then produce a prediction as to if the patient is going to be at work or not in the future.”Erkin Otles, U-M Industrial and Operations Engineering (IOE) Student and Medical Scientist Training Program Fellow.
The new model uses longitudinal observation information and predictive analytics that could come from specific treatment notes, claims, or any other information that goes into a patient’s chart during their healing process. This is done through a technique called recurrent neural networks which use sequential data to produce a response, in this case, everyday.
Otles is the lead author on this paper. Co-authors on this paper are Principal Investigator Brian Denton, U-M IOE Department Chair, Jon Seymour, Peers Health CEO, and Haozhu Wang, Amazon Research Scientist. This work was conducted under a research contract with Peers Health.
“We show that using the longitudinal information for the same type of model has a better prediction ability than using information that’s only collected around the time of injury,” said Otles.
This prediction model will help insurance companies, worker’s compensation agencies, government organizations and employers better predict when employees with occupational injuries can return to work. The model also helps patients who need the most care get it in a timely manner.
“This model ensures that patients who are at risk of protracted recovery will be found as the people who might most benefit from additional management of their case,” said Otles. “Additionally, you have a new entity that is continually thinking about every patient instead of reviewing them on a case by case basis.”
Looking toward the future there are many paths for this model to be utilized and studied further; both in academia as well as in industry.
“There’s still a lot of additional research that can be done,” said Otles. “We can study the model in usage and conduct a trial with the people making decisions to determine whether it helps them make better decisions. We can also create tools around this model to help corporations fully utilize it and address a gap in occupational health.”
Click here to read the full article “Dynamic Prediction of Work Status for Workers with Occupational Injuries: Assessing the Value of Longitudinal Observations” on the Journal of the American Medical Informatics Association (JAMIA) website.
Peers Health is an electronic medical publisher applying advanced technologies to functional outcomes data.