Utility of Machine Learning and Radiomics Based on Cavity for Predicting the Therapeutic Response of MDR-TB

Xinna Lv,1,* Ye Li,1,* Botao Cai,2 Wei He,1 Ren Wang,1 Minghui Chen,1 Junhua Pan,1 Dailun Hou1 1Department of Radiology, Beijing Chest Hospital, Capital Medical University, Beijing, 101149, People’s Republic of China; 2Department of Radiology, Harbin Chest Hospital, Harbin, 150000, People’s Republic of China*These authors contributed equally to this workCorrespondence: Junhua Pan; Dailun Hou, Department of Radiology, Beijing Chest Hospital, Capital Medical University, Beijing, 101149, People’s Republic of China, Tel +8613901097074 ; +8618001286699, Email 13901097074@139.com; hou.dl@mail.ccmu.edu.cnBackground: Sputum culture result at the sixth month is essential for predicting therapeutic response to longer multidrug-resistant tuberculosis (MDR-TB) regimens. This study aimed to construct a predictive model using cavity-based radiomics to predict sputum status at the sixth month for MDR-TB patients treated with longer regimens.Methods: This retrospective study recruited 315 MDR-TB patients treated with longer regimens from two centers (250 patients from center 1 and 65 patients from center 2), who were divided into persistently positive and conversion to negative sputum culture groups according to sputum results. Radiomics features were extracted based on the cavity, and a radiomics model was selected and established using a random forest classifier. The clinical characteristics and primary CT signs with significant differences were integrated to build a clinical model. A combined model was generated using the radiomics and clinical model. ROC curves, F1-score and DCA curves were used to assess the predictive performance of the models.Results: Twenty-eight radiomics features were selected to build a radiomics model for predicting the sputum status. The radiomics model achieved good performance, with AUCs of 0.892 and 0.839 in the training and testing cohort, respectively, which was similar to the performance of the combined model (0.913