C27-15 Clinical Features Enhance the Performance of Radiomics Model in Differentiating Coccidioides From Malignant Pulmonary Nodules in Southwestern United States

Abstract Rationale In fungal endemic regions, differentiating benign from malignant pulmonary nodules on CT scans remains a major diagnostic challenge. Coccidioides nodules frequently resemble malignant lesions on CT and PET imaging, presenting significant challenges for differentiation, even among radiologists. Consequently, the benign biopsy and resection rate, especially in the Southwestern US, remains high. We previously developed DOI:10.1371/journal.pone.0196910 and externally validated DOI:10.1183/13993003.02485-2020 the Benign versus aggRessive nODule Evaluation using Radiomic Stratification (BRODERS) machine learning (ML) classifier to non-invasively distinguish between benign and malignant pulmonary nodules with an AUC of 0.87. However, its performance in certain benign nodules with malignant surface and texture characteristics like in coccidioidomycosis had a suboptimal AUC of 0.65. In this study, we redesigned our radiomics model by including numerous clinical features. Methods In this retrospective study, 113 pulmonary nodules (66 Coccidioides and 47 adenocarcinoma spectrum) detected on CTs with ground truth pathology diagnosis were identified by screening for key terms like “nodule”, “cocci”, and “adenocarcinoma” from both CT and pathology reports using a large language model trained on clinical textual data (Illuminate®). Clinical features such as demographics, smoking history, and laboratory studies were extracted by trained clinicians from the electronic medical record. All nodules were segmented and radiomic features extracted using a semi-automated seed growing algorithm previously described DOI: 10.1097/JTO.0b013e3182843721. The database was randomly split into an 80%-20% development and hold-out set prior to ML modeling. Feature selection for variables was conducted utilizing Elastic Net Regression, which provides a balanced approach between Lasso and Ridge regression. A logistic regression model was developed using the selected features and ground truth information of the training cohort. The true performance of the model was tested on the hold-out set. Results 160 radiomics and 22 clinical features were extracted. Elastic Net regression identified 44 features that contributed the most. Some of these clinical features include presence of satellite lesions, smoking history, and presence of positive IgG coccidiomycosis serology. The AUC of the ROC curve for the hold-out test set was 0.86. Conclusions The clinical features of significance that emerged during multimodal modeling are explainable. Designing explainable ML models is important, especially in the Southwestern US where Coccidioides nodules can appear to be malignant to both human experts and ML models. Although coccidioidomycosis serologies alone lack sensitivity and specificity to diagnose Coccidioides nodules, integrating them with a multimodal ML model can improve differentiation between benign Coccidioides and malignant adenocarcinoma nodules. This abstract is funded by: None