Multimodal convolutional neural network for tablet-level dissolution prediction using compression force and UV...

The prediction of tablet dissolution from in-process data remains a key challenge in pharmaceutical manufacturing, as in vitro dissolution is a critical quality attribute that cannot be measured inline. In this study, a multimodal convolutional neural network (MI-CNN) was developed to predict tablet-level dissolution profiles for immediate-release tablets, combining images taken under UV illumination and compression force. To evaluate the contribution of input selection and feature representation, the MI-CNN was compared with a single-input CNN (SI-CNN) using images only, and a multilayer perceptron (MLP) based on hand-crafted image descriptors and compression force. All models were evaluated on a dataset generated using a Design of Experiments approach, covering multiple compression forces, disintegrant concentrations, and acetylsalicylic acid particle size fractions. The MI-CNN achieved the most consistent performance, with comparable training and validation errors (RMSE: 13.09% and 12.54%, respectively), and demonstrated robust generalization across formulation conditions, including an unseen particle size range. The SI-CNN showed reduced accuracy (RMSE_val: 25.41%), particularly in cases where dissolution differences were governed by tablet compaction. The MLP model exhibited excellent training performance (RMSE_train: 2.94%) but poor generalization (RMSE_val: 27.15%), indicating overfitting due to the limited ability of histogram-based features to adequately represent the complexity of the dataset. Model explanation using SHapley Additive exPlanations (SHAP) revealed that both compression force and image-derived features contributed to the predictions. Overall, the results demonstrate that combining process variables with image-based information enables accurate and robust dissolution prediction at the tablet level, supporting data-driven approaches for real-time release testing.