Radiomics and machine learning for predicting mycobacterial response to proteasome inhibition: A multi-center CT imaging study

Abstract

Objective

This study aimed to develop a non-invasive imaging-based model to predict the response of mycobacterial infections to proteasome inhibition in lung cancer patients.

Materials and methods

This retrospective, multi-center study included 1480 lung cancer patients with concurrent Mycobacterium tuberculosis infection, recruited from eight hospitals. All patients underwent chest CT imaging and had proteasome activity measured in tumor tissues. Tumor regions were manually segmented, and radiomic features were extracted. Features with high inter-rater reliability were selected using Intraclass Correlation Coefficients (ICCs), and Least Absolute Shrinkage and Selection Operator (LASSO) regression was applied to reduce feature redundancy. Four machine learning algorithms were trained and evaluated using cross-validation. The model with the highest performance was selected to construct a predictive nomogram. Decision Curve Analysis (DCA) was performed to quantify clinical benefit, and ComBat harmonization was applied to address inter-scanner variability.

Results

The Random Forest classifier achieved the best performance, with an area under the AUC of 0.91 and an accuracy of 86.2 %. The model demonstrated a sensitivity of 84.9 % and a specificity of 87.4 %. Radiomic features such as wavelet-transformed texture entropy and GLCM correlation showed strong associations with tumor proteasome activity (correlation coefficient r = 0.78, p < 0.001). The developed nomogram, integrating radiomic features and clinical variables, yielded robust predictive accuracy (Hosmer-Lemeshow p = 0.27). External validation on an independent cohort (n = 92) confirmed generalizability with an AUC of 0.88. DCA demonstrated superior net clinical benefit compared to conventional biomarkers.

Conclusions

Our findings support the use of radiomics and machine learning to predict immune response in lung cancer patients with tuberculosis, providing a promising, non-invasive tool for precision medicine. However, as the study population was limited to lung cancer patients with TB, further validation in broader TB cohorts without cancer is necessary to assess the generalizability of the findings.