Understanding the complexity of near-infrared quantification of highly porous patient-tailored drug products by utilizing chemometrics and stimulated Raman imaging.

Additively manufactured drug products, typically produced using small-scale, on-demand batch mode, require rapid and non-destructive quantification methods. A tunable modular design (TMD) approach combining porous polymeric freeze-dried modules and an additive manufacturing method, inkjet printing, was proposed in an earlier study to fabricate accurate and patient-tailored doses of an antidepressant citalopram hydrobromide. This approach addresses the unmet medical needs associated with antidepressant tapering. Non-destructive quantification of printed porous structures is challenging due to the presence of residual solvents and frequent fluctuation of the material density. These shortcomings were mitigated by utilizing a spinning near-infrared spectroscopy (NIRS) measurement setup and a post-print drying step. A machine learning algorithm (ML), specifically support vector regression, was implemented to lessen potential non-linearities caused by the complex structure of TMD drug products. The non-linear support vector regression models performed better than linear partial least squares (PLS) models when modeling the entire sample set (prediction error improved by 19 %). By dividing the TMD samples into subtypes and creating individual models for each subtype improved model performance: linear PLS models performed better or equally to non-linear models. It was hypothesized that this outcome was due to the structural differences between different TMD sample subtypes that was later confirmed by stimulated Raman scattering (SRS) microscopy. It was demonstrated that for complex porous drug products ML algorithms can improve NIRS model performance when a single universal robust model is preferred, and SRS is a powerful tool to explain the challenges that printing onto porous drug products can introduce to the NIRS quantification.