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

Area of Science:

• Pharmaceutical Manufacturing
• Drug Delivery Systems
• Computational Chemistry

Background:

• Predicting tablet dissolution from in-process data is challenging for pharmaceutical manufacturing.
• In vitro dissolution is a critical quality attribute not measurable inline.
• Real-time release testing requires accurate dissolution prediction models.

Purpose of the Study:

• To develop a multimodal convolutional neural network (MI-CNN) for predicting tablet-level dissolution profiles.
• To compare the MI-CNN with single-input CNN (SI-CNN) and multilayer perceptron (MLP) models.
• To evaluate the impact of input selection and feature representation on dissolution prediction accuracy.

Main Methods:

• A multimodal convolutional neural network (MI-CNN) was developed, integrating UV images and compression force data.
• MI-CNN performance was compared against SI-CNN (images only) and MLP (hand-crafted features and compression force).
• Models were evaluated on a dataset from a Design of Experiments approach, varying compression force, disintegrant concentration, and particle size.

Main Results:

• The MI-CNN demonstrated superior and consistent performance with low training and validation errors (RMSE: 13.09% and 12.54%).
• MI-CNN showed robust generalization across diverse formulation conditions, including unseen particle size ranges.
• SI-CNN and MLP models exhibited lower accuracy and poorer generalization, with MLP overfitting due to limited feature representation.

Conclusions:

• Combining process variables (compression force) with image-based data via MI-CNN enables accurate and robust tablet dissolution prediction.
• This approach supports data-driven strategies for real-time release testing in pharmaceutical manufacturing.
• The study highlights the effectiveness of deep learning for in-process monitoring and quality control.