Hierarchical taxonomy-aware modeling for granular drug resistance and disease outcome prediction in multimodal tuberculosis cohorts.

Tuberculosis is a disease with many phenotypes, complicated by the hierarchical clinical taxonomies of drug resistance that are not typically leveraged in existing predictive modeling. We developed a taxonomy-driven, hierarchical classification framework trained on a large and diverse multimodal cohort of subjects from the NIAID TB Portals to build node-specific classifiers covering multi-modal features including clinical, imaging and genomic attributes. The process systematically predicts progressively detailed categories of disease phenotype, drug resistance status and response to therapy by aligning the model structure with actual diagnostic pathways. The evaluation consists of various hierarchically related loss functions and supervisory metrics that include accuracy, hierarchical loss, node-wise precision/recall and F1-scores, and error propagation to evaluate both predictive power as well the interpretability under a suit of analyses with different hierarchical dependencies. Our results show that the hierarchical taxonomy-aware approach to SPLIT models can deliver better performance at fine-grained subcategory levels than flat classifiers without sacrificing explainable predictions and reduced error propagation. The work presents a novel method that relates to core medical ontologies and capable of placing an AI-driven clinical decision support in context that can be transparent to users, which is a milestone for holistic tuberculosis care as much as it depicts risk propagation at different layers of conceptual hierarchies within the model.