Evaluating near-infrared spectroscopy and multivariate modelling for predicting dissolution profiles of 3D printed extended-release tablets

Non-destructive quality control strategies are essential to enable production of personalised printlets (3D printed tablets). The prediction of dissolution behaviour from printlet-specific data, such as spectroscopic, remains underexplored. In this study, we demonstrate the use of multimodal data consisting of printlet near-infrared (NIR) spectra, timepoint and geometric (surface area-to-volume ratio; SA/V) information, to predict in vitro dissolution profiles of dose-personalised, extended-release (ER) printlets manufactured by fused deposition modelling. Printlets varying in theophylline, hydroxypropyl cellulose (HPC), and D-mannitol were produced in three distinct volumes and SA/Vs for training and three distinct volumes of equal SA/V for testing. Partial least squares (PLS) and Random Forest (RF) regression were investigated through prediction of single or multiple dependent release variables. RF models using a single response variable with features Min-Max scaled outperformed multi-response RF and PLS approaches. Further feature reduction based on spectral variance enhanced the predictive capability, with retraining using the 10 most informative features achieving good predictive performance (R² = 0.994, RMSE_train = 1.80%, RMSE_test = 2.75%). Time and SA/V were the most dictating features, while the retained NIR wavelengths, corresponding to characteristic vibrations from theophylline and HPC, offered model fine-tuning. The predicted dissolution profiles were equivalent to experimentally determined profiles for all test printlets, based on f1 and f2 similarity criteria. These results support the use of NIR spectroscopy with geometric parameters through multivariate modelling as a non-destructive approach for predicting drug release from 3D printed ER tablets to advance quality assurance for personalised pharmaceutical manufacturing.