Supplementary Material for: Building and prospectively evaluating a prediction model to forecast urgent dialysis needs across four tertiary hospitals

Background: Urgent dialysis is labour-intensive and expensive because it requires specialized nursing staff. Most hospitals schedule a fixed number of nurses daily for urgent dialysis needs, but daily dialysis demand fluctuates, leading to inefficiencies.

Methods: We developed statistical, machine learning, and deep learning models to predict the next seven days' dialysis needs. Our study included a retrospective (Apr 1, 2018 to Mar 31, 2023) and prospective component (Nov 1 to 30, 2023 and May 31 to June 27, 2024) across four hospitals (Hospital A for one hospital and Hospital B for three hospitals combined). To avoid model over-fitting, we divided our data into three sets: training, testing, and validation. The latter was performed prospectively during two silent deployment periods. The primary outcome measure was the mean absolute error (MAE).

Results: The mean daily dialysis volume in the retrospective data was 16.0 (standard deviation [SD], 5.7) for Hospital A and 4.5 (SD, 2.3) for Hospital B. The best performing models were autoregressive integrated moving average (ARIMA) and temporal convolutional network; both resulted in an MAE of 3.0 procedures for Hospital A and 1.5 procedures for Hospital B, compared to 4.4 and 1.9 respectively for the benchmark. During our two prospective evaluations, the mean daily dialysis volume was 16.8 (SD, 4.5) for Hospital A and 4.2 (SD, 2.5) for Hospital B. The ARIMA model resulted in the lowest MAE at 2.2 and 1.5 procedures, respectively.

Conclusions: Our multicentre, six-year study demonstrated that urgent in-hospital dialysis needs can be accurately forecasted.