External Model Validation and Evaluation

For each neonate in the external data set, model-based population-predicted concentrations were computed by locking the final parameters of the original model through the NONMEM MAXEVAL = 0 POSTHOC command.¹⁷ These population-predicted concentrations were graphically compared with the corresponding observed concentrations for all levels, as well as separately for low and high concentration levels. This comparative analysis was conducted across all phases of controlled TH and during both the hypothermic and normothermic phases independently. The predictive performance was assessed using bias and precision, which were calculated using the following equations¹⁸:

where predicted refers to the model-predicted gentamicin concentrations, observed pertains to the measured gentamicin concentrations, and N represents the number of pairs. To comprehensively assess the predictive performance of the model across different phases of controlled TH, bias and precision were computed for gentamicin concentrations during both the hypothermic and normothermic phases. In addition, a distinction was made between low and high gentamicin levels because a higher bias or lower precision would have more significant implications for low gentamicin concentrations (typically trough levels) compared with high concentrations (typically peak levels).

Because gentamicin was often discontinued after a single dose, and blood samples were collected at fixed intervals regardless of dosing times, there were a limited number of true trough levels available for analysis. Consequently, rather than relying solely on trough levels, a cutoff of a gentamicin concentration of ≤1.5 mg/L was established. This approach allowed for a direct comparison between the actually measured low gentamicin concentrations and the population-predicted concentrations derived from the original PK model. High gentamicin levels were defined as samples taken within 2 hours after the preceding administered dose because all the highest gentamicin concentrations were measured within this time interval, aligning with a previous study's approach.¹⁹

To further evaluate predictive performance, a prediction-corrected visual predictive check (pcVPC) and a NPDE analysis were conducted, both with n = 1000 simulations.^20,21 Subsequently, if no apparent trends, imprecision, or bias were detected in the previous steps, the model building and external data sets were merged and jointly analyzed by refitting the merged data set to the original gentamicin PK model. A parameter obtained from the model refit was deemed accurate if it deviated by less than 20% from the original model fit. A covariate analysis on the merged data set was performed once again, using a forward and backward selection process. A decrease in the objective function value of ≥3.8 points was considered statistically significant in the first step, followed by a more stringent decrease in objective function value of ≥10.83 (P-value of <0.001) in the second part. Finally, a pcVPC of the merged data set was generated, and the robustness of the refitted model was assessed through a bootstrap analysis.