Assumption-light feature discovery outperforms Cox-based selection for PM2.5 constituent analysis in an open...

Area of Science:

• Environmental Health
• Data Science
• Computational Biology

Background:

• AI misapplications are common in environmental research due to misunderstandings of machine learning assumptions.
• Complex exposure-outcome relationships require careful feature selection for accurate modeling.

Purpose of the Study:

• To evaluate the impact of different feature selection strategies on downstream AI model performance in environmental research.
• To identify robust feature selection methods for analyzing air quality and health outcomes.

Main Methods:

• Compared Cox-based significance, Feature Agglomeration (FA), Highly Variable Gene Selection (HVGS), and Spearman's rank correlation.
• Assessed feature selection methods using a COPD mortality-air quality benchmark with a fixed Random Forest model under cross-validation.
• Developed and evaluated a hybrid workflow combining unsupervised structure discovery and nonparametric screening.

Main Results:

• Spearman's rank correlation consistently yielded the highest accuracy with 5 and 8 features.
• Feature Agglomeration was competitive for smaller feature sets; HVGS showed moderate performance.
• Cox-based selection underperformed, indicating issues like nonlinearity and multicollinearity.
• The hybrid workflow produced more stable and reproducible feature sets.

Conclusions:

• Feature selection significantly influences AI model performance in environmental health studies.
• Spearman's rank correlation and a hybrid approach offer more reliable feature selection than traditional Cox-based methods.
• The findings provide a practical framework to mitigate AI misapplications and enhance causal modeling in environmental research.