Optimizing super-feature selection for machine learning-enhanced spectroscopic analysis in biomedical research

Purpose

Machine-learning-powered label-free infrared spectroscopic methods offer significant potential for diagnostic and biomedical applications. However, their applications have been limited by spectral noise, where critical features are often obscured by overlapping bands and data redundancy. Although various feature selection methods have been proposed, many suffer from limitations in consistency and interpretability. To address these challenges, we introduce a novel multi-model machine learning approach that integrates five distinct algorithms to identify a set of “super-features”—spectral features consistently deemed significant across all models.

Principal results

This novel workflow outperforms traditional algorithms, achieving superior classification accuracy (>99%) in distinguishing infected from healthy cells, despite using fewer spectral features. To ensure robustness and generalizability, we developed a comprehensive validation strategy that includes independent classifier evaluations, label randomization, and unsupervised analyses. Importantly, the identified super-features accurately differentiated infection states across multiple time points and enhanced the biological interpretability of infection-associated biochemical changes.

Conclusions

These findings highlight the potential of advanced multi-model feature selection techniques to enhance the diagnostic power of spectroscopic data in biomedical research, offering high accuracy and valuable biological insights into infection progression.