Physiology-Based Identifiable Model-Quantity Subset Selection for Early Prediction of Intradialytic MAP Trajectories from Sparse Observations

Intradialytic hypotension (IDH) is a common and harmful complication of hemodialysis that is closely associated with changes in mean arterial pressure (MAP). However, mechanistic prediction of intradialytic MAP is hindered by sparse early-stage observations and model-quantity non identifiability in physiology-based models. This study develops a dimensionality-reduced predictive framework following a screen-select-estimate workflow for early prediction of intradialytic MAP trajectories from sparse MAP observations. Among 35 candidate model quantities considered in the screening procedure, the initial blood volume Vb0 and post-capillary venous resistance Rs3 are selected for estimation, whereas the remaining candidate quantities and treatment inputs are kept fixed or prescribed. Sensitivity analysis and Fisher information matrix (FIM)-based subset selection show that this two-quantity subset provides the best trade-off between physiological information content, identifiability, and numerical stability. Using MAP observations from the first 100 min of dialysis, the selected subset supports prediction of subsequent MAP trajectories across four representative patient types. At the end of the 240-min ultrafiltration period, corresponding to approximately 3.00 L prescribed ultrafiltration in the reference simulation, the model-derived terminal specific blood volume ranges from 66.48 to 69.19 mL/kg and RBV(240) ranges from 88.74% to 91.83% across the four phenotypes. Prediction-phase root-mean-square error (RMSE) ranges from 1.09 to 5.78 mmHg, and the maximum error ranges from 1.71 to 8.51 mmHg. This patient-type-level framework enables interpretable MAP trajectory prediction and provides a methodological basis for future patient-level evaluation of IDH-related MAP decline.