Predicting tablet dissolution from in-process data is a major challenge in pharmaceutical manufacturing because in vitro dissolution is a critical quality attribute that cannot be measured inline. This study addresses that challenge by developing a multimodal deep learning approach that combines imaging and process variables to predict dissolution at the tablet level. The authors aimed to evaluate whether adding a process variable—compression force—to image-based inputs would improve predictive accuracy and generalization compared with image-only or hand-crafted-feature models.
Model development and evaluation used a dataset generated following a Design of Experiments. The experimental space included multiple values of compression force, varying disintegrant concentrations, and different acetylsalicylic acid particle size fractions. The dataset intentionally included an unseen particle size range during validation to test model generalization across formulation conditions. The abstract does not report the exact number of tablets, images, or preprocessing details; those specifics are available in the full text.
Three modeling approaches were compared:
The comparison was designed to probe both input selection (images vs images+process variable) and feature representation (learned features via CNNs vs histogram descriptors).
Model performance was evaluated using root mean square error (RMSE) on training and validation sets. The evaluation emphasized not only within-sample performance but also generalization to unseen formulation conditions, including an unseen particle size range. The authors also applied explainable artificial intelligence methods—specifically SHapley Additive exPlanations (SHAP)—to quantify the contribution of compression force and image-derived features to the model outputs.
The MI-CNN delivered the most consistent performance across training and validation, indicating robust generalization. Reported error metrics for the MI-CNN were RMSEtrain: 13.09% and RMSEval: 12.54%.
The SI-CNN, relying on images alone, showed reduced validation accuracy with RMSEval: 25.41%. The abstract highlights that the SI-CNN was particularly less accurate in cases where dissolution differences were governed by tablet compaction, underscoring the value of including compression force as an input.
The MLP based on histogram descriptors achieved excellent training performance (RMSEtrain: 2.94%) but generalised poorly (RMSEval: 27.15%). This disparity suggests overfitting and indicates that histogram-based features lacked sufficient representational capacity to capture the dataset's complexity.
Taken together, these results show that combining image-based information with process variables improves predictive accuracy and robustness when compared with image-only or simple descriptor-based models.
The authors used SHapley Additive exPlanations (SHAP) to interpret model predictions. SHAP analysis revealed that both compression force and image-derived features contributed meaningfully to the MI-CNN's dissolution predictions. This finding supports the multimodal design by demonstrating that the process variable and learned image features each provide unique information relevant to dissolution behavior.
The study demonstrates that a multimodal deep learning surrogate model can predict tablet-level dissolution with improved accuracy and robustness relative to single-input or histogram-feature approaches. These outcomes support data-driven strategies for surrogate dissolution modeling and suggest a pathway toward enabling real-time release testing for immediate-release tablets by combining non-destructive imaging and accessible process measurements such as compression force.
The abstract reports key comparative error metrics and the experimental design factors but does not provide comprehensive details on several technical aspects. Missing or unreported items in the abstract include exact dataset size, image acquisition parameters, preprocessing workflows, full model architectures and hyperparameters, training epochs, and cross-validation strategy. The abstract also does not provide detailed quantitative performance broken down by formulation condition beyond the aggregate RMSE values. Those technical specifics would need to be obtained from the full text.
Overall, the study provides evidence that multimodal deep learning combining UV fluorescent imaging and compression force can produce accurate and generalizable predictions of tablet dissolution, and that explainable AI methods can identify the relative contributions of image and process inputs to model output.