Multimodal radiomics integrating deep learning and clinical features for diagnosing multidrug-resistant tuberculosis in HIV/AIDS patients

Abstract

Background: This study aimed to develop and validate a predictive model based on multimodal data, including clinical features, radiomics features, and deep learning features, to distinguish multidrug-resistant tuberculosis (MDR-TB) in HIV/AIDS patients, thereby improving diagnostic accuracy.

Methods: A retrospective cohort of HIV/AIDS patients with drug-sensitive tuberculosis (n = 164) and MDR-TB (n = 63) admitted to the Fourth People's Hospital of Nanning between January 2016 and July 2024 was included. The dataset was randomly divided into training and validation sets at a 7:3 ratio. A multimodal model was constructed by integrating a clinical model, a radiomics model, and a 2.5D multi-instance learning (MIL) approach.

Results: Key predictors-platelet count and C-reactive protein-were identified through univariate and multivariate logistic regression analysis. The integrated model achieved the highest performance in both the training and validation set (AUC=0.943 and 0.899, respectively), significantly outperforming individual models. Grad-CAM effectively localized key image regions influencing decision-making, while a nomogram quantified the contribution weights of each predictor, enhancing model transparency. The Hosmer-Lemeshow (HL) test confirmed good model calibration, and the decision curve analysis (DCA) curve demonstrated the optimal clinical net benefit of the integrated model.

Conclusion: The multimodal integrated model developed in this study significantly improved the diagnostic efficacy of MDR-TB in HIV/AIDS patients by combining clinical, radiomics, and deep learning features, providing a reliable tool for individualized precision diagnosis and treatment.

Keywords: Deep learning; Hiv/aids; Multidrug-resistant; Radiomics; Tuberculosis.