Abstract
In this research, advanced regression techniques are investigated for modeling intricate release patterns utilizing a high-dimensional dataset comprising more than 1500 spectrum-based variables and categorical inputs. The spectral data are collected from Raman spectroscopy for analysis of drug release from a solid dosage formulation coated with Polysaccharides (a high-dimensional dataset of 155 samples, with drug release measured at 2, 8, and 24 h). The considered drug is 5-aminosalicylic acid for colonic drug delivery, and its release was estimated using Raman data as inputs along with other categorical parameters. The models, including Kernel Ridge Regression (KRR), Kernel-based Extreme Learning Machine (K-ELM), and Quantile Regression (QR) incorporate sophisticated approaches like the Sailfish Optimizer (SFO) for hyperparameter optimization and K-fold cross-validation to enhance predictive accuracy. Notably, KRR exhibited exceptional performance, achieving an R² of 0.997 on the training set and 0.992 on the test set, with a mean squared error (MSE) of 0.0004. In comparison, K-ELM and QR achieved lower R² values of 0.923 and 0.817 on the test set, respectively. The key innovation lies in integrating these non-linear regression models with robust data preprocessing steps, including dimensionality reduction via Principal Component Analysis (PCA), categorical feature encoding through Leave-One-Out (LOO), and outlier detection using Isolation Forest. This study significantly contributes by offering a comprehensive framework for managing high-dimensional and heterogeneous datasets, while emphasizing the effectiveness of optimization strategies in predictive modeling. By accurately predicting the release of 5-ASA from polysaccharide-coated formulations, these models can aid in the design of targeted colonic delivery formulations with optimized release kinetics, ultimately enhancing the efficacy of treatments for colonic diseases.