Development of an artificial intelligence model for prediction of dry weight in chronic hemodialysis patients and assessment of its accuracy compared to standard bioelectrical impedance analysis

Abstract

Proper determination of dry weight (DW) is crucial for achieving positive outcomes in hemodialysis (HD) patients. However, the traditional clinical assessment of DW (C-DW) is often inaccurate. Recently, bioimpedance spectroscopy (BIS) analysis using a Body Composition Monitor (BCM) device has emerged as a gold standard method for determining DW (BCM-DW). Despite its accuracy, the high cost of the BCM device limits its accessibility. To overcome this challenge, the current study proposes a machine learning (ML) model, which is a part of artificial intelligence (AI), to assess DW using available clinical and laboratory parameters.
 
 Objective: To develop an ML model for predicting DW (ML-DW) and compare it with BCM-DW.
 
 Methods: The study consisted of a model development phase and a performance assessment phase. Retrospective data from chronic HD patients between 2017 and 2022 from two dialysis centers in Bangkok were retrieved. The parameters for this ML model included demographic, dialysis prescription, laboratory, and intradialytic time-varying data. The data utilized during the ML model development phase consisted of a training group for optimizing the parameters of the models and a validation group for determining when to stop the optimization. The final output of the model was ML-DW. The primary outcome of the study was the agreement comparison between ML-DW and BCM-DW.
 
 Results: All 56,000 time-varying data from 1,151 HD sessions were included in the ML model. The mean BCM-DW was 58.8?11.7 kgs, while the mean predicted ML-DW from the model was 59.5?11.3 kgs. The Bland-Altman plot showed the bias estimated by the mean difference was 0.78 kg, and the limit of agreement was -3.7 to 2.2 kg.
 
 Conclusion: This was the first study that developed a machine learning model aimed at predicting BCM-DW. Compared to other models, this one tried to explore the utilization of time-series data in the input variables. It also demonstrated external validation across different institutions. This study served as a proof-of-concept that machine learning can be a useful tool for DW prediction, but it is not yet a replacement tool for BCM. This warrants further model development that can be widely used in real practice.