AI deep learning can help predict hospitals’ acute dialysis needs
Key takeaways:
- During two prospective evaluations, forecasts had a mean absolute error of 3.1 patients per day.
- Using AI-forecast schedules would have reduced overstaffing by 67% and understaffing by 79%.
LAS VEGAS — The use of AI deep learning can help forecast acute dialysis needs and enhance staffing efficiency in hospitals, according to a speaker.
“Every time you use ChatGPT,” for example, “that is deep learning,” Michael Fralick, MD, PhD, a clinician scientist at Sinai Health in Toronto, told Healio. This type of machine learning involves training artificial networks to recognize patterns in large amounts of data, he said.
“Using a similar approach ... we wanted to answer the question: ‘Can we forecast? Can we predict how many dialyses there are going to be in the next 7 days?’”
Researchers analyzed data from 37,388 dialysis procedures across four hospitals in Canada from 2018 to 2024, in which AI was used to predict dialysis requirements for the upcoming week. The study included a retrospective and a prospective component, with retrospective data divided into separate training, validation and testing sets to avoid model overfitting.
Input for the AI model included calendar trends, historical dialysis data and weather conditions, with a main outcome of mean absolute error, measuring prediction accuracy.
Traditionally, acute dialysis is labor-intensive, Fralick said, and hospitals generally have scheduled a fixed number of specialized nurses, leading to patient-staff mismatch.
Using the AI model, “we could predict plus or minus two urgent dialyses. So, we thought, ‘That is good enough. Let us see what happens in real life.’ So, the nurse manager clicks a button and gets the forecast the same way you get a forecast for the weather for how many dialyses there will be over the next week,” he said. “And then if a nurse calls in sick, but it is going to be a slow day, they do not backfill and are able to save money.”
The mean daily dialysis volume in the retrospective data was 20.5 patients, according to the findings, with the model achieving a mean absolute error of 3.2. During two prospective evaluations in 2023 and 2024, forecasts had a mean absolute error of 3.1 patients per day.
Further, assuming hospitals maintain a staff of 10 nurses daily to meet an average demand of 20.5 patients, implementing AI-forecast schedules would have reduced overstaffing by 67% and understaffing by 79%, the researchers found.
Overall, “we think this approach [will] allow some smart scheduling, because most hospitals just say, ‘We are going to have 10 nurses available every day for urgent dialysis,’ but the demand might vary,” Fralick said.
For more information:
Michael Fralick, MD, PhD, can be reached at mike.fralick@mail.utoronto.ca.
Collapse
Lim B, et al. Using deep learning to forecast acute dialysis needs in hospital: A retrospective and prospective multicentre study. Presented at: Annual Dialysis Conference; March 13-16, 2025; Las Vegas.
Read more from AI in Medicine
View all
Read more about