Auditing Geographic Domain Shift in Genomic AI: A Memory-Efficient Pipeline for Tuberculosis Drug Resistance Prediction

The deployment of machine learning models in clinical genomics holds immense potential for rapid drug susceptibility testing (DST). However, clinical AI is highly vulnerable to geographic domain shift. This paper audits the geographic generalizability of XGBoost models predicting Isoniazid (INH) resistance in Mycobacterium tuberculosis. Utilizing the CRyPTIC dataset, we propose a memory-constrained preprocessing pipeline to extract and reshape high-volume genomic data, evaluate performance across distinct global laboratories, and deploy SHAP (SHapley Additive exPlanations) to ensure predictions are driven by biological mechanisms rather than geographic metadata artifacts. Our evaluation reveals a reverse domain shift, demonstrating that models can achieve superior performance on unseen populations with lower intrinsic genetic variance.