ORCID
- Joan Mnyambo: 0000-0002-4365-7067
- Amir Aly: 0000-0001-5169-0679
- Emmanuel Ifeachor: 0000-0001-8362-6292
- Yinghui Wei: 0000-0002-7873-0009
- Shang-Ming Zhou: 0000-0002-0719-9353
- Stephen Mullin: 0000-0002-1936-394X
Document Type
Article
Abstract
Introduction: Early detection of drug-resistant tuberculosis (DR-TB) is criticalfor timely treatment and limiting transmission. Differentiating DR-TB fromdrug-sensitive TB (DS-TB) remains challenging due to overlappingmanifestations and subtle differences in chest X-rays (CXRs).Methods: We developed an explainable multimodal machine learningframework integrating CXR embeddings from a Data-Efficient ImageTransformer (DEiT) with patient-level data, including clinical, interpretedradiological, demographic, socioeconomic, and behavioral variables. Using6,279 subjects from the TB Portals program, model performance wasevaluated across three settings comprising CXR embeddings only, patientlevel features only, and multimodal. Interpretability was analyzed usingShapley Additive Explanations (SHAP).Results: A multimodal XGBoost model achieved a precision-recall area underthe curve (PR-AUC) of 0.9258 and receiver operating characteristic area underthe curve (ROC-AUC) of 0.8718. CXR-embeddings-only classifiers performedlower (PR-AUC: 0.8334; ROC-AUC: 0.7369), while patient-level XGBoostyielded highest performance. The multimodal model improved interpretabilitythrough complementary structural information. SHAP identified CXRembeddings as primary global drivers. Among patient-level variables, dailycontacts and employment status were the strongest contributors, reflectingexposure intensity and socioeconomic conditions. Comorbidity showedmoderate effects, while education and case definition had smallercontributions. Key radiological factors included non-tuberculosisabnormalities and mediastinal lymph node involvement, whereas mediumdensity stabalized fibrotic nodules, medium nodules, and small cavitiesdemonstrated minimal association. Interaction analysis revealed higher riskwith increasing comorbidity in the presence of morphological alterations; BMIand age exhibited modest influence. Independent review by two cliniciansconfirmed plausibility and consistency of SHAP explanations.Discussion: This interpretable multimodal framework supports early DR-TBdetection and targeted interventions in high-burden settings.
DOI Link
Publication Date
2026-07-30
Publication Title
Frontiers in Digital Health
Volume
8
Acceptance Date
2026-06-22
Deposit Date
2026-10-09
Funding
The author(s) declared that financial support was received for this work and/or its publication. This research was funded by the Commonwealth Scholarship Commission (CSC), United Kingdom, under the Foreign, Commonwealth and Development Office (FCDO) scholarship programme (Scholarship No. TZCS2023-543), in collaboration with the University of Plymouth, United Kingdom. It also received support from the Government of Tanzania through the Ministry of Education, Science and Technology.
Additional Links
Keywords
chest x-ray, data-efficient image transformer (DEiT), drug-resistant tuberculosis, interpretability, shapley additive explanations (SHAP)
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 International License.
Recommended Citation
Mnyambo, J., Aly, A., Ifeachor, E., Wei, Y., Zhou, S., & Mullin, S. (2026) 'Interpretable Multimodal Machine Learning for Diagnosis of Drug-Resistant Tuberculosis', Frontiers in Digital Health, 8. Available at: 10.3389/fdgth.2026.1828522
