Search for: All records

Algorithmic bias in COVID-19 detection systems poses aserious threat to equitable pandemic response, asdemographicdisparities in model performance risk worsening healthoutcomes across vulnerable populations. We present anadoptedCausal Concept Bottleneck Model (C2BM) framework thatsystematically addresses fairness in multimodal COVID-19detection by learning interpretable concepts from chest CTscans and patient metadata. Our approach targets theCountry → Institution → COVID causal pathway throughprincipledinterventions, achieving substantial bias reduction: age andgender demographic parity differences decrease from 51.15%to 18.50% (64% reduction), gender disparate impact improvesfrom 0.6475 to 0.9812 (51% improvement), whilepreserving 98.45% diagnostic F1-score. Throughcomprehensive evaluation across four model variants, wedemonstrate that causal interventions enable stable andreproduciblefairness improvements without compromising clinicalutility. Our work establishes that principled causalreasoning canachieve practical fairness-accuracy trade-offs in COVID-19detection systems, providing actionable guidance forequitable healthcare AI deployment. 

Accurate prediction of the transmission fitness of emerging SARS-CoV-2 variants is vital for timely public health responses. In this study, we present a deep learning framework that predicts variant fitness from raw genomic sequences using a convolutional neural network (CNN) trained to regress Differential Population Growth Rate (DPGR) values. Our approach achieves high predictive accuracy R-square value of 0.92 on genomic sequences sampled from the USA and Europe. To interpret the model’s predictions, we apply SHapley Additive exPlanations (SHAP) to identify nucleotide-level contributions to predicted fitness. Our analysis highlights key mutations in ORF9 (nucleocapsid), ORF2 (spike), ORF5 (membrane), and ORF8 that either enhance or reduce predicted DPGR. Notably, we identify amino acid–altering mutations such as D3L, E484K, N501Y, and V97I as strong positive contributors to fitness, while synonymous or non-coding mutations had more subtle or regulatory effects. These findings validate the potential of sequence-based modeling and interpretable AI to support early detection and prioritization of high-risk variants.