Comparison of Modeling Approaches in Design of Experiment: Statistical Methods versus Machine Learning Techniques

Abstract

Objective: To compare the modeling performance between Linear Regression (LR) and four Machine Learning (ML) algorithms, namely Decision Tree (DT), Support Vector Machine (SVM), K-Nearest Neighbor (K-NN), and Neural Network (NN). Methods: Data were obtained from 88 pharmaceutical Design of Experiment (DOE) studies published between October 2021 and October 2022, retrieved from PubMed and Scopus databases. The raw data were used to construct models using RapidMiner Studio 10, and model performance was evaluated by the coefficient of determination (R²). Statistical significance was tested using Kruskal–Wallis test. Results: Across all DOE types, K-NN and DT models achieved significantly higher mean R² values than LR, while NN and SVM performed comparably to K-NN and DT but without statistical differences. In the Central Composite Design, K-NN, NN, and SVM models showed significantly higher R² than LR, whereas DT performed similarly to LR. For the Full Factorial Design, K-NN and DT yielded significantly higher R² than LR, and K-NN also outperformed NN and SVM, while NN and SVM were comparable to LR. In the Mixture Design, no significant difference in mean R² was observed between LR and any ML algorithms, indicating similar overall model performance. Conclusion: The study demonstrated that ML techniques—particularly K-NN, NN, and DT—can effectively enhance modeling performance in pharmaceutical DOE, yielding higher R² values than LR in several DOE types, including All DOE types, Central Composite Design, and Full Factorial Design. Although no clear difference was found in Mixture Design, the findings highlight the potential of ML methods to improve experimental data analysis and serve as promising tools for formulation design in future pharmaceutical research.