Mechanistic, data-driven, and hybrid models: A critical comparison in surrogate drug dissolution modeling

Mathematical modeling is becoming increasingly important in the pharmaceutical industry. It supports the Quality by Design framework by aiding process understanding and examining the impact of critical material and process parameters on the critical product properties. In this work, a critical comparison of knowledge-based, data-driven, and hybrid modeling approaches was conducted through surrogate in vitro dissolution modeling of acetylsalicylic acid (ASA) tablets for the first time. Furthermore, a novel serial-type hybrid model that integrates a neural network (ANN) and a population balance model (PBM) is presented. All models were evaluated on a dataset covering multiple compression forces, disintegrant contents, and ASA particle size values following a Design of Experiments plan. Three ANNs were developed with different representations of the ASA's PSD (using average values, number-, or volume-based PSD). The model, using PSD average metrics, resulted in the best prediction with 2.14 % and 5.49 % training and validation root mean squared errors (RMSE). Explainable ANN was developed for a more reliable comparison of models by investigating the influential input parameters. The standalone PBM achieved the least accurate predictions with 9.20 % and 12.08 % training and validation RMSEs. The hybrid model outperformed the PBM model with 6.22 % and 8.90 % training and validation RMSEs. The comparative assessment showed that all models are suitable for immediate-release predictions, and ANN and hybrid models can accurately describe slow dissolutions. Although the data-driven model obtained the best prediction, the results demonstrated that hybrid modeling can serve as a reliable and interpretable predictive tool.