In-host modeling challenges using population approach methods

TL;DR

This study highlights the limitations of relying solely on RSE<50% for parameter estimation in non-linear mixed effects models, emphasizing the importance of additional diagnostics for accurate in-host disease modeling.

Contribution

It demonstrates that RSE alone is insufficient for model validation, advocating for comprehensive diagnostics including correlation matrices and goodness-of-fit assessments.

Findings

RSE<50% does not guarantee accurate model fit or parameter identifiability.

Strong inverse correlations can indicate issues despite low RSE.

Model predictions often poorly fit actual patient data.

Abstract

Non-linear mixed effects models are widely used to estimate parameter estimates in the field of pharmacometrics across pharmaceutical industry, US regulatory agencies and academia. The preciseness of the parameter estimate is evaluated using relative standard error (RSE) with a threshold of <50% considered as 'precisely estimated'. Here we investigate the use of this metric alone in Monolix to calibrate a recently published in-host mathematical model for hepatitis D virus (HDV) with our own longitudinal data obtained from patients treated with HDV-entry inhibitor bulevirtide (BLV) monotherapy for up to 96 weeks. We identified substantial discordance between Monolix calibration output, measured longitudinal data and the HDV model despite the fact that Monolix parameters had a RSE <50%, suggesting that model parameters were estimated with precision. Surprisingly, while Monolix suggested precise parameter estimates based on RSE<50%, the correlation matrix in Monolix indicated a strong inverse correlation between BLV efficacy and the loss rate of HDV-infected cells raising identifiability issues. Furthermore, the fits failed to reproduce HDV kinetics accurately in the majority of patients (i.e., poor goodness of fit). Lastly, the estimated pretreatment serum HDV level varied significantly from measured observations. In summary, we demonstrate that even when RSE was <50%, other outputs such as the correlation matrix, confidence intervals and goodness of fit at both the individual and population level need to be checked for accuracy to accept or refine a proposed model.