Causal Convolution-Transformer for Time Series Forecasting of Dry Weight in Hemodialysis Patients

Dry weight (DW) is a crucial parameter used to determine the ultrafiltration volume before a hemodialysis (HD) session. In clinical practice, it is defined as the lowest postdialysis body weight that does not lead to adverse symptoms during dialysis. However, the precise measurement of dry weight in the medical field remains a “missing piece in the puzzle”. Due to the complexity and instability of the existing methods for evaluating DW and significant inter-individual variations among patients, ensuring timely and accurate assessment of DW presents a significant challenge. In this study, we convert the assessment of dry weight into a task of time series forecasting and an advanced time series forecasting model, named the causal convolution-transformer, is proposed for the clinical forecasting of DW in HD patients. The proposed approach combines the exceptional performance of the transformer in time series data processing with the analytical capability of causal convolution for capturing local front and back information. We collect a total of 929,753 medical data samples from 3,276 HD patients in a hospital and extract 37 clinically relevant features associated with DW based on clinical expertise. After conducting experiments and performing a comparative analysis with clinical evaluation, the mean square error for the causal convolution-transformer model is 0.018 and the R-square is 0.963, which effectively accomplishes the task of forecasting DW.