How to determine when data incompleteness compromises the correct training of data-enhanced mechanical models of haemodialysis? A data quality criterion.

Abstract

Hemodialysis is essential for managing end-stage renal disease, with rising patient numbers increasing the need for individualized treatment. To support this, hybrid models-integrating mechanistic frameworks with data-driven calibration-have increasingly replaced purely mechanistic or statistical approaches. These models rely on complete, high-quality clinical data; however, incomplete values-particularly hematological data-are common and may compromise model reliability. In this study, we evaluated the impact of incomplete data acquired during dialysis on the training and predictive performance of hybrid models, aiming to develop a method for identifying when session data can no longer be considered suitable for model calibration. We developed an approach to handle missing data, used available complete sessions to simulate realistic scenarios of data acquisition loss, and assessed its impact. Based on the extent of missing data, 386 dialysis sessions from 52 patients were classified into three categories of completeness corresponding to different training robustness. The impact of less robust training on predictive ability was then evaluated. The outcome of this process is a method to identify the sessions in which the lack of data impedes proper training of the model. A quality criterion is proposed to assess the robustness of the model training. Based on this, the prediction is deemed either reliable or flagged for caution. In such cases, the clinician is advised to repeat a training session to ensure a robust patient-specific prediction. The same approach applies to all the hybrid models describing the patient-specific response to dialysis.