EUVIMEDEUVIMED entdecken
Crossref

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

Yoshiyasu Takefuji

Environmental Pollution · 2026-04 · 394 · 127738

DOI: 10.1016/j.envpol.2026.127738 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Environmental Pollution, 2026-04. Verfasst von Yoshiyasu Takefuji. Beteiligte Einrichtung: Faculty of Data Science, Musashino University, 3-3-3 Ariake Koto-ku, Tokyo, 135-8181, Japan. Electronic address: takefuji@keio.jp. Hauptthemen laut MEDLINE: Particulate Matter, Air Pollutants, Air Pollution, Environmental Monitoring. Die Arbeit wurde bisher 1-mal zitiert (Crossref) und nennt 16 Literaturangaben.

AbstractSchlagwörterBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

AI misapplications are widespread in environmental research, often arising from limited understanding of machine learning assumptions and their alignment with complex exposure-outcome relationships. Motivated by Chen et al., who used marginal structural Cox models to study PM2.5 constituents and highlighted sensitivity to modeling choices, we use a public COPD mortality-air quality benchmark as a proxy to examine how feature selection strategies affect downstream performance. We compare Cox-based significance, Feature Agglomeration (FA), Highly Variable Gene Selection (HVGS), and Spearman's rank correlation, assessing each with a fixed Random Forest under cross-validation. Spearman consistently delivered the highest accuracy with 5 and 8 features, FA was competitive for compact sets, HVGS was moderate, and Cox-based selection underperformed-patterns consistent with nonlinearity, multicollinearity, and potential violations of proportional hazards. A hybrid workflow that combines unsupervised structure discovery with nonparametric screening produced more stable and reproducible feature sets, offering a pragmatic guardrail against common misapplications and a stronger foundation for subsequent flexible causal modeling. Public Python code supports reproducibility.

Schlagwörter

MeSH-Schlagwörter (MEDLINE)

Particulate MatterAir PollutantsAir PollutionEnvironmental MonitoringProportional Hazards ModelsMachine LearningRandom Forest

Stichwörter der Autor:innen

Causal inferenceFeature selectionNonparametric screeningPM2.5 constituentsProportional hazards

Substanzen

Particulate MatterAir Pollutants

Autor:innen und Institutionen

  • Faculty of Data Science, Musashino University, 3-3-3 Ariake Koto-ku, Tokyo, 135-8181, Japan. Electronic address: takefuji@keio.jp.

Zitieren

APA

Takefuji, Y. (2026). Assumption-light feature discovery outperforms Cox-based selection for PM2.5 constituent analysis in an open benchmark. Environmental Pollution, 394, 127738. https://doi.org/10.1016/j.envpol.2026.127738

Vancouver

Takefuji Y. Assumption-light feature discovery outperforms Cox-based selection for PM2.5 constituent analysis in an open benchmark. Environmental Pollution. 2026;394:127738. doi:10.1016/j.envpol.2026.127738

BibTeX
@article{Takefuji2026,
  author = {Takefuji, Yoshiyasu},
  title = {Assumption-light feature discovery outperforms Cox-based selection for PM2.5 constituent analysis in an open benchmark},
  journal = {Environmental Pollution},
  year = {2026},
  volume = {394},
  pages = {127738},
  doi = {10.1016/j.envpol.2026.127738}
}
Kostenlos starten →