Comparative analysis of supervised machine learning algorithms for transdermal drug delivery in brain disorders.

To improve patient compliance and provide stable and prolonged delivery of drugs to the brain, transdermal drug delivery systems (TDDS) are being explored as a viable method of treatment of various brain diseases without the invasion of any organs. Optimization of TDDS remains problematic owing to the complex nonlinear interplay between formulation elements and process factors that affect drug entrapment efficiency, stability, and drug-release properties. Therefore, this study aims to develop a supervised Machine Learning (ML) framework to simultaneously predict multiple critical formulation attributes in brain-targeted TDDS. A dataset comprising 542 formulation records collected from 48 peer-reviewed studies and 6 validated laboratory sources was utilized. Physiochemically informed preprocessing, feature engineering, and Cuckoo Catfish Optimizer (CCO)-based hyperparameter tuning were integrated with multiple supervised learning models, including Linear Regression, Decision Tree, Random Forest, Support Vector Regressor, Gradient Boosting, and Artificial Neural Networks. Seven key formulation outputs, namely entrapment efficiency, drug loading, particle size, PolyDispersity Index (PDI), zeta potential, release time, and drug release, were simultaneously predicted and interpreted using SHapley Additive exPlanations (SHAP). Ensemble models significantly outperform all linear models and the conventional standalone approaches by demonstrating an average R increase of 24.63% compared to the standard models while achieving comparable stability and performance for most of the formulation properties. In addition, SHAP revealed that lipid composition, the type of surfactant, formulation pH, and the temperature used for formulation preparation were the key determinants in formulating successful TDDS, which can ultimately lead to an accurate, interpretable, and scalable decision-support platform that significantly accelerates TDDS development by decreasing the overall experimental workload and allowing data-driven design.