Optimizing Pre-Trained Models for Medical Dataset Classification with a Fine-Tuning Approach

Abstract

Medical organizations struggle to deal with huge high-dimensional datasets that need powerful machine learning systems to produce precise healthcare outcomes. Traditional analytical techniques prove inadequate when dealing with extraction from features and performance of classifiers in this specific setting. The research introduces an algorithm which enhances Stacked Autoencoders (SAEs) by combining them with a customized Logistic Regression model intended for medical high-dimensional data analysis. This approach implements a Hybrid Imputation Method using MICE and KNN Imputation which precedes other stages and helps process missing values and outliers in medical data. We use CNNs and SAEs together for deep feature extraction before using Feature Fusion to assemble a robust feature collection. A set of the most important features is identified by executing Advanced Ensemble Feature Selection (EFS) procedures which include Few-shot Learning and Model-Agnostic Meta-Learning Algorithm (MAML) and Genetic Algorithm-Based Feature Selection (GAFS). The procedure of fine-tuning pre-trained models represents an effective enhancement for classification tasks particularly in situations of limited dataset availability. The experimental outcomes demonstrate remarkable performance gains in terms of accuracy and sensitivity and specificity as well as reduced execution time as compared to current techniques. Upcoming work for this study involves speeding up algorithm processing abilities and scalability alongside the integration of robust deep learning structures with self-supervised learning methodologies together with upgrade transfer learning approaches for medical dataset variety applications. The study will concentrate on enhancing model transparency through explainable AI and real-time validation for clinical deployment and ethical and regulatory compliance to develop this technique for practical healthcare settings.