Can synthetic data overcome the privacy and fidelity bottleneck in Pharmacometrics? A comparative benchmark using a daptomycin population pharmacokinetic model

Abstract Introduction The sharing of individual patient data is essential for advancing pharmacometrics but is strictly limited by privacy regulations (e.g., GDPR). While synthetic data generation offers a legally compliant alternative, its structural impact on complex nonlinear mixed-effects (NLME) modelling remains largely unexplored. This study aimed to benchmark five generative artificial intelligence algorithms by evaluating the balance between data privacy and the preservation of structural PK properties and clinical dosing guidance. Material & methods A daptomycin two-compartment PopPK model was used to simulate a reference cohort of 500 patients. Five generative algorithms (Modified AVATAR, Gaussian Copula, Synthpop, TVAE, and CTGAN) produced 100 independent synthetic datasets each. A two-stage evaluation framework was applied: first, a statistical indistinguishability test based on logistic regression (AUC ROC) was used as a macroscopic pre-selection criterion to determine algorithm eligibility for NLME modelling and privacy risk assessment. Privacy risk was independently quantified using the Anonymeter framework (Singling Out and Linkability attacks). Eligible algorithms were further evaluated on PK parameter recovery bias and clinical dosing simulations. Results Deep learning architectures (TVAE, CTGAN) were excluded at the pre-selection stage due to both biologically implausible covariate generation and high macroscopic detectability (mean AUC ROC = 0.837 and 0.986, respectively). Synthpop, AVATAR, and Gaussian Copula all passed the indistinguishability threshold (AUC ROC = 0.475 +- 0.033, 0.490 +- 0.013, and 0.619 +- 0.031, respectively) and proceeded to NLME evaluation. However, attack-based privacy assessment revealed that Synthpop carried an unacceptable singling-out risk (0.035), disqualifying it from privacy-preserving data sharing. AVATAR and Gaussian Copula demonstrated acceptable privacy profiles (singling-out = 0.004 and 0.001; linkability = 0.010 and 0.003, respectively). At the structural level, Gaussian Copula injected stochastic noise inflating residual error (+157.0%) and V1; (+25.9%), blunting predicted Cmax and predisposing to empirical dose escalation and risk of toxicity. AVATAR acted aSs a smoothing filter, deflating V2; (-48.3%) and underestimating CL (-12.9%). Forward clinical simulations confirmed directionally opposed prediction errors: Gaussian Copula consistently underestimated Cmax across standard and renally impaired profiles (-14.5% and -16.0%, respectively), predisposing to empirical dose escalation, whereas AVATAR- and Synthpop-derived models overestimated Cmax and Cmin in the obese infected patient (+14.7% and +8.2%, respectively), compounding the accumulation risk already present in this profile. Conclusion While no generative algorithm currently offers a perfect solution, AVATAR and Gaussian Copula represent the most viable candidates, being the only methods to satisfy both macroscopic indistinguishability and attack-based privacy criteria. These findings highlight the necessity of a structured, two-stage validation framework and suggest that, when coupled with therapeutic drug monitoring, synthetic datasets could significantly enhance multicentre collaboration while maintaining strict regulatory compliance