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

Area of Science:

• Pharmaceutical Sciences
• Chemical Engineering
• Computational Modeling

Background:

• Mathematical modeling is crucial for pharmaceutical Quality by Design (QbD).
• It aids in understanding processes and predicting drug product performance.
• Evaluating different modeling approaches for drug dissolution is essential.

Purpose of the Study:

• To critically compare knowledge-based, data-driven, and hybrid modeling for in vitro dissolution.
• To introduce a novel hybrid model combining artificial neural networks (ANN) and population balance models (PBM).
• To assess model performance using experimental data for acetylsalicylic acid (ASA) tablets.

Main Methods:

• Developed and compared ANN, PBM, and a novel ANN-PBM hybrid model.
• Utilized in vitro dissolution data from acetylsalicylic acid (ASA) tablets manufactured under varying conditions (Design of Experiments).
• Investigated different particle size distribution (PSD) representations within the ANN models.

Main Results:

• The ANN model using average particle size metrics achieved the best prediction accuracy (2.14% training, 5.49% validation RMSE).
• The novel hybrid ANN-PBM model demonstrated superior performance over the standalone PBM (6.22% training, 8.90% validation RMSE).
• All evaluated models were suitable for immediate-release predictions; ANN and hybrid models effectively described slow dissolutions.

Conclusions:

• Data-driven ANN models offer excellent predictive capabilities for drug dissolution.
• Hybrid modeling provides a reliable and interpretable alternative for predictive dissolution assessment.
• The novel hybrid model enhances process understanding and supports QbD implementation in pharmaceuticals.