Clinically Constrained Multi-Stage Architecture for Integrated Tuberculosis Risk, Drug-Resistance and Severity Assessment

Abstract

Tuberculosis (TB) remains a major global health burden because of high infection and mortality rates, delays in diagnosis, the emergence of drug-resistant strains, and the wide variation in how the disease presents. Diagnosis of TB, estimation of the risk of multidrug resistance (MDR), and early identification of patients with severe illness are usually carried out as separate tasks. This study proposes a clinical decision support system that detects TB, estimates MDR-TB risk, and grades disease severity through a single multi-stage pipeline, with each stage's output checked against rules set by clinicians. The system was evaluated with stratified 5-fold cross-validation on harmonized public clinical datasets. It achieved an area under the receiver operating characteristic curve (AUROC) of 0.91 for TB detection and 0.83 for MDR-TB risk estimation, and a quadratic weighted kappa (QWK) of 0.74 for severity grading, outperforming logistic regression, random forest, and multi-task neural network baselines on all three tasks. The results indicate that a unified, explainable decision support system is a feasible and practical tool for the combined assessment of patients with TB.