(303r) Integrating Three-Fluid Nozzle Technology as a QbD Tool for Enhancing Advanced Respiratory Drug Delivery Formulations: A Statistical Modeling Approach

Dry powder inhalers, DPIs, have been gaining increased attention as an effective treatment for both local and systemic diseases, owing to their distinct advantages in terms of better physical and chemical stability. Conventional DPI formulations are traditionally composed by physical mixtures of coarse carriers with one (or more) micronized active pharmaceutical ingredients (API), typically with aerodynamic particle size below 5 µm. However, despite being a well-developed approach, these formulations still present some challenges and carrier-free composite particles, where one or more APIs and excipients are embedded into a single solid phase arise as alternative.

This approach improves uniformity and dispersibility, having the potential to increase therapeutic activity. Composite particles can present different morphologies, namely encapsulated particles, where the API core is surrounded by an outer layer of an excipient. The preparation of such particles within the inhalable range requires particle engineering methods to ensure correct lung deposition and increased bioavailability. Among those, spray drying (SD) has been shown to be a preferred method to prepare encapsulated composite particles as it enables fine tuning of powder properties and is an easy-scalable process.

A three-fluid external nozzle (3FN) is a novel pneumatic atomization strategy that features a three-layered concentric structure. In this design, two liquid feeds flow through independent channels, thus overcoming solubility and solvent incompatibility issues and eliminating the need for co-solvent systems that solubilize all components in a single solution such as those often found for traditional atomization strategies (traditional approach). The three-fluid nozzle enables swift and scalable manufacturing of formulations incorporating encapsulated active pharmaceutical ingredients (APIs). This advancement holds significant advantages for lung drug delivery, offering enhanced therapeutic outcomes, targeted delivery, controlled release, and improved patient compliance. Using this process tool one aqueous solution containing a shell-forming agent is fed to the outer channel of the nozzle while an organic solution of the API is fed to the innermost channel.

The success of dry powder formulations is tightly linked to its formulation properties, however, when manufactured by means of spray drying the interrelations between product’s critical quality attributes (CQAs) and manufacturing potential critical process parameters (pCPPs) pose a challenging task. On that frame, Quality by Design (QbD), a systematic framework endorsed by regulatory authorities, has gained significant momentum for its potential to streamline pharmaceutical development and manufacturing processes. Embracing the principles of QbD, it is pivotal to have a systematic understanding of the pCPPs and CQAs that can significantly enhance the manufacturing robustness and efficiency. To reach that goal a statistical approach is often employed, by means of design of experiments (DoE), reducing the number of required tests that maximize the interactions between pCPPs and CQAs. A DoE as a QbD tool enables to establish mathematical relationships between cPPs and CQAs by means of empirical modelling tools, such as statistical regression models.

The goal of this study was to employ a QbD strategy to develop empirical models for CQAs of DPIs, consisting of encapsulated composite particles prepared using a 3FN. A DoE was performed where the investigated CQAs were particle size, reported as the Dv50, and aerosolization parameters, namely fine particle fraction FPF and emitted dose, ED. Spray drying parameters such as the outlet temperature (T_out), the atomization flow rate (F_atom) and the ratio between the feed flow of the organic and aqueous solution (M_ratio) were considered as pCPPs, while other parameters were kept constant for a model formulation of trehalose:leucine (80:20 % w/w) containing Itraconazole as model drug.

The DoE was based on a three-factor full factorial design and a central point. An exploratory data analysis (EDA) enabled the identification of relationships and patterns in the data. The empirical models were developed by fitting a multivariate linear regression (MLR) by least squares. Stepwise variable selection was applied to ensure only significant variables were included, thus assuring model simplicity. The stopping criterion for the variable selection was minimization of the Akaike information criterion (AIC). Moreover, outlier identification was performed by Hotelling’s T² distribution analysis.

The developed models were evaluated with a hypothesis test, whereby the null hypothesis must be rejected, for a given significance level α (α=5% in this study). To evaluate the models’ significance an ANOVA analysis was performed. Model fitness to the data was evaluated by the coefficient of determination , and the adjusted coefficient of determination, , as the latter prevents the selection of over-fitted models. Multicollinearity was checked by the variation inflation factor (VIF), considering a VIF threshold of <5. Lastly, to validate the models and assess their predictive capacity, model predictions were compared with a new set of independent experimental data.

From the EDA, linear relationships could be identified, it could also be observed that, likely ED was independent on the studied pCPPs because all tests yielded similar results. Therefore, no regression model was developed for this CQA, as it is likely a cause of the selected formulation without any significant statistical impact of the tested cPPs. Regarding particle size (Dv50), variable selection identified T_out and F_atom as the factors that explained most variability of this CQA, with F_atom’s coefficient being higher than for T_out, as seen in Table 1. According to the model increased atomization and temperature yield finer particles. The values of and were 0.86 and 0.82, respectively, demonstrating that the model adequately fits the data. The inclusion of T_out as a predictor was not expected as it does not influence the droplet size directly. Furthermore, no second order interactions were considered. Therefore, it can be hypothesized that T_out might have a potential impact in Dv50 by having an impact on the drying kinetics. Conversely, the omission of feed flow rate, reflected by the M_ratio, could possibly be attributed to the narrow operating range evaluated, limited by solubility challenges and the need to keep feed rates low. The FPF model encompassed F_atom and M_ratio as predictor variables, with the M_ratio having the most significant coefficient. This model follows the same trend as the Dv50 model, where higher atomization leads to a superior FPF, often linked to the smaller particle size. Regarding the M_ratio it has a negative effect on the FPF, which allows to extrapolate that high ratios are expected to negatively impact the powder aerosolization. This model resulted in a and of 0.84 and 0.79, respectively, both indicative of a robust fit of the model to the experimental data. No second order interactions were incorporated in either of the models, meaning in this situation SD optimization can rely only on the direct manipulation of process parameters. In terms of multicollinearity, the VIF for the predictors of each model were all approximately 1, confirming that there is no correlation between the variables. Furthermore, ANOVA analysis corroborated the models’ statistical relevance in explaining the variance in the output data, evident through having both the p-values <0.5 and F-Ratio> Fcrit.

To further analyze the models and ensure their accuracy, external validation was conducted. Two experiments were conducted within the established design space as intermediate random points, and one experiment was conducted outside of the previously studied conditions, with both the T_out and M_ratio being tested above their upper level of the DoE. The points inside the design space were correctly predicted by both models, with relative errors between predicted and actual results being under 10%. However, for the point outside the studied range, the Dv50 model struggled to accurately predict the particle size. This is due to the incongruence between the models expected size reduction with increased temperature and the reality of droplet drying where temperature does not significantly affects particle size distribution. On the other hand, the FPF model accurately predicted the output, maintaining a relative error below 10%. Therefore, the models are reliable, having good predictive power within the studied design space, as is usual with empirical models. Moreover, the FPF model’s omission of the T_out renders it to be more approximate to the expectation than the Dv50 model.

Overall, the results enabled the development of empirical models for two CQAs of DPIs. The models enabled the identification of the actual CPPs for the investigated CQA’s and established relationships between them. It was concluded that the T_out and F_atom are the CPPs that influence particle size, which is not in agreement with what is commonly observed, and that F_atom and M_ratio are the CPPs that affect the FPF. The models provide insight into the optimization of the SD process to finely tailor powder properties, particularly concerning API encapsulation with a 3FN. Further enhancements could encompass expanding the DoE, for instance developing a central composite design and broader factor range. This would add useful information to the model, enabling the investigation of potential second order interactions or curvatures in the output data, rendering the model more robust and applicable.