A Novel Feature Selection Framework Based on Large Language Models

In the context of intelligent recommendation systems for diagnosis and treatment plans based on clinical data, the high dimensionality of the data significantly limits the system’s ability to respond rapidly and generate accurate recommendations. This directly undermines the practicality and effectiveness of decision support systems in emergency medical scenarios. This study proposes an innovative feature selection framework that integrates multiple state-of-the-art large language models to perform parallel deep correlation analysis on high-dimensional clinical data. By precisely identifying the complex intrinsic relationships and autocorrelations among features, the framework optimizes the feature selection process. Comparative experiments conducted on six medical datasets demonstrate that the framework significantly enhances the solving performance of various metaheuristic algorithms by improving the quality of initial solutions. Furthermore, the experimental results reveal significant performance variations across different datasets, indicating that the enhancement effects are closely related to the characteristics of the datasets. However, due to the limited availability of clinical data, the model’s generalizability to ultra-high-dimensional data has not been effectively validated. Benefiting from the exceptional computational capabilities of large language models, the proposed framework not only maintains a high level of analytical depth and precision for high-dimensional clinical data but also significantly improves the system’s response speed. This ensures high-performance feature selection when processing high-dimensional data, thereby providing fast and accurate support for clinical decision-making.