C2BM: Causal Concept Disentanglement for Fair Multimodal COVID-19 Detection

Abstract

Algorithmic bias in COVID-19 detection systems poses a serious threat to equitable pandemic response, as demographic disparities in model performance risk worsening health outcomes across vulnerable populations. We present an adopted Causal Concept Bottleneck Model (C2BM) framework that systematically addresses fairness in multimodal COVID-19 detection by learning interpretable concepts from chest CT scans and patient metadata. Our approach targets the Country → Institution → COVID causal pathway through principled interventions, achieving substantial bias reduction: age and gender demographic parity differences decrease from 51.15% to 18.50% (64% reduction), gender disparate impact improves from 0.6475 to 0.9812 (51% improvement), while preserving 98.45% diagnostic F1-score. Through comprehensive evaluation across four model variants, we demonstrate that causal interventions enable stable and reproducible fairness improvements without compromising clinical utility. Our work establishes that principled causal reasoning can achieve practical fairness-accuracy trade-offs in COVID-19 detection systems, providing actionable guidance for equitable healthcare AI deployment.