External Validation and Bayesian Forecasting of Rivaroxaban Population Pharmacokinetic Models in Older Chinese Patients with Nonvalvular Atrial Fibrillation.
Chen J, Huang W, et al. • Drug design, development and therapy • 2026
Published rivaroxaban population pharmacokinetic models demonstrated heterogeneous transferability in older Chinese patients with nonvalvular atrial fibrillation, and Bayesian forecasting using a single trough concentration provided only modest, model-dependent improvement in concentration prediction without consistently overcoming limitations of the underlying models.
Key Findings
Results
None of the six published rivaroxaban PopPK models consistently met predefined acceptance criteria when externally validated in older Chinese patients with NVAF.
Six published rivaroxaban PopPK models were evaluated using an independent cohort from the RIVA-GAP study.
Predictive performance was assessed using prediction error metrics reflecting bias, precision, and prediction coverage, together with normalized prediction distribution error (NPDE) diagnostics.
135 patients with 257 rivaroxaban concentration observations were included for external validation.
Predictive performance varied substantially among the six models, and none consistently met predefined acceptance criteria.
Results
Model A showed comparatively lower prediction bias among the evaluated models but still demonstrated limited precision.
Model A had a median prediction error (MDPE) of 33.64%, indicating systematic overprediction bias.
Median absolute prediction error (MDAPE) for Model A was 53.72%, reflecting poor precision.
F30 (fraction of predictions within 30% of observed concentrations) for Model A was only 31.52%.
Despite being the best-performing model for bias, Model A still fell short of typical acceptance thresholds for clinical application.
Results
Bayesian forecasting using paired steady-state trough-peak concentrations reduced prediction bias for Model A but did not substantially improve precision.
122 patients with paired trough-peak samples were included for Bayesian forecasting.
Bayesian forecasting reduced the MDIPE (median individual prediction error) of Model A from 33.64% to 12.38%.
Despite bias reduction, MAIPE (median absolute individual prediction error) remained at 37.33% after Bayesian forecasting.
IF30 (individual fraction within 30% of observed concentrations) was 41.8% after Bayesian forecasting for Model A.
Similar model-dependent effects were observed across the other five models.
Results
Bayesian forecasting using a single trough concentration provided modest and model-dependent improvement in peak concentration prediction.
Bayesian forecasting was performed using paired steady-state trough-peak concentrations to evaluate changes in individual peak concentration prediction.
Improvement in predictive performance following Bayesian updating was not consistent across all six models.
The approach did not consistently overcome the limitations of the underlying base models.
The improvement in bias did not translate to clinically acceptable precision across the evaluated models.
Results
Published rivaroxaban PopPK models demonstrated heterogeneous transferability when applied to older Chinese patients with NVAF.
The study population consisted of older Chinese patients with nonvalvular atrial fibrillation (NVAF), a population distinct from those in which the original models were developed.
Performance differences among models suggest that demographic, genetic, and clinical characteristics of older Chinese NVAF patients may not be adequately captured by existing models.
The external validation cohort was drawn from the RIVA-GAP study (ChiCTR2300074934, registered 21 August 2023).
No single model demonstrated uniformly acceptable bias, precision, and coverage metrics across all predefined criteria.
What This Means
This research suggests that existing mathematical models used to predict rivaroxaban (a blood thinner) drug levels in the body do not work well when applied to elderly Chinese patients with a common heart rhythm disorder called nonvalvular atrial fibrillation. The researchers tested six previously published models against real blood concentration measurements from 135 patients, and found that all six models performed poorly — none reliably predicted drug levels within acceptable ranges of accuracy. The best-performing model still overestimated concentrations by about 34% on average and was only correct within 30% of the actual measured level about 31% of the time.
The study also tested whether a technique called Bayesian forecasting — which uses a patient's own measured trough (lowest) drug level to refine predictions of their peak (highest) drug level — could improve accuracy. While this approach did reduce systematic bias somewhat (for example, cutting one model's average error from 33.6% to 12.4%), the overall precision remained poor, and the improvement was inconsistent depending on which underlying model was used. This means that simply having one drug level measurement is not enough to reliably personalize dosing predictions using current models.
This research matters because rivaroxaban is widely used to prevent strokes in patients with atrial fibrillation, and getting the dose right is important — too little risks stroke, too much risks bleeding. The findings highlight a gap: existing pharmacokinetic models were largely developed in non-Chinese or younger populations and may not accurately reflect how older Chinese patients process this drug. This suggests that new models specifically developed and validated in this patient group may be needed before model-guided precision dosing can be reliably implemented in this population.
Chen J, Huang W, Chen G, Ge B, Zhao Q, Wang H, et al.. (2026). External Validation and Bayesian Forecasting of Rivaroxaban Population Pharmacokinetic Models in Older Chinese Patients with Nonvalvular Atrial Fibrillation.. Drug design, development and therapy. https://doi.org/10.2147/DDDT.S632231