Comparative Analysis of Resource-efficient Benchmarking and Observation of Neural and Machine learning models with Emission evaluation for Diagnostic datasets

This paper introduces CARBON-MED, the first unified emission-aware
benchmarking study for classical ML models across six heterogeneous medical
diagnostic datasets. We propose three novel evaluation metrics (the
Carbon-Efficiency Score (CES), Stability Index, and Emission-Aware Performance
Index (EAPI)) that jointly assess predictive performance and carbon footprint.
Our findings demonstrate that lightweight, GPU-optimized models (XGBoost, KNN)
achieve state-of-the-art diagnostic accuracy with near-negligible CO₂ emissions,
establishing a principled and reproducible benchmark for sustainable clinical AI.