Hannah N Marmor1, Laurel Jackson2, Susan Gawel3, Michael Kammer4, Pierre P Massion5, Eric L Grogan6, Gerard J Davis7, Stephen A Deppen8. 1. Department of Thoracic Surgery, Vanderbilt University Medical Center, 1211 Medical Center Drive, Nashville, TN 37232, USA. Electronic address: hannah.marmor@vumc.org. 2. Abbott Diagnostics Division, 100 Abbott Park Road, Abbott Park, IL 60064, USA. Electronic address: laurel.jackson@abbott.com. 3. Abbott Diagnostics Division, 100 Abbott Park Road, Abbott Park, IL 60064, USA. Electronic address: susan.gawel@abbott.com. 4. Department of Pulmonary and Critical Care Medicine, Vanderbilt University Medical Center, 1211 Medical Center Drive, Nashville, TN 37232, USA. Electronic address: michael.kammer@vumc.org. 5. Department of Pulmonary and Critical Care Medicine, Vanderbilt University Medical Center, 1211 Medical Center Drive, Nashville, TN 37232, USA. 6. Department of Thoracic Surgery, Vanderbilt University Medical Center, 1211 Medical Center Drive, Nashville, TN 37232, USA; Tennessee Valley Healthcare System, Veterans Affairs, 1310 24th Avenue South, Nashville, TN 37212, USA. 7. Abbott Diagnostics Division, 100 Abbott Park Road, Abbott Park, IL 60064, USA. Electronic address: gerard.davis@abbott.com. 8. Department of Thoracic Surgery, Vanderbilt University Medical Center, 1211 Medical Center Drive, Nashville, TN 37232, USA; Tennessee Valley Healthcare System, Veterans Affairs, 1310 24th Avenue South, Nashville, TN 37212, USA. Electronic address: steve.deppen@vumc.org.
Abstract
BACKGROUND: Non-invasive biomarkers are needed to improve management of indeterminate pulmonary nodules (IPNs) suspicious for lung cancer. METHODS: Protein biomarkers were quantified in serum samples from patients with 6-30 mm IPNs (n = 338). A previously derived and validated radiomic score based upon nodule shape, size, and texture was calculated from features derived from CT scans. Lung cancer prediction models incorporating biomarkers, radiomics, and clinical factors were developed. Diagnostic performance was compared to the current standard of risk estimation (Mayo). IPN risk reclassification was determined using bias-corrected clinical net reclassification index. RESULTS: Age, radiomic score, CYFRA 21-1, and CEA were identified as the strongest predictors of cancer. These models provided greater diagnostic accuracy compared to Mayo with AUCs of 0.76 (95 % CI 0.70-0.81) using logistic regression and 0.73 (0.67-0.79) using random forest methods. Random forest and logistic regression models demonstrated improved risk reclassification with median cNRI of 0.21 (Q1 0.20, Q3 0.23) and 0.21 (0.19, 0.23) compared to Mayo for malignancy. CONCLUSIONS: A combined biomarker, radiomic, and clinical risk factor model provided greater diagnostic accuracy of IPNs than Mayo. This model demonstrated a strong ability to reclassify malignant IPNs. Integrating a combined approach into the current diagnostic algorithm for IPNs could improve nodule management.
BACKGROUND: Non-invasive biomarkers are needed to improve management of indeterminate pulmonary nodules (IPNs) suspicious for lung cancer. METHODS: Protein biomarkers were quantified in serum samples from patients with 6-30 mm IPNs (n = 338). A previously derived and validated radiomic score based upon nodule shape, size, and texture was calculated from features derived from CT scans. Lung cancer prediction models incorporating biomarkers, radiomics, and clinical factors were developed. Diagnostic performance was compared to the current standard of risk estimation (Mayo). IPN risk reclassification was determined using bias-corrected clinical net reclassification index. RESULTS: Age, radiomic score, CYFRA 21-1, and CEA were identified as the strongest predictors of cancer. These models provided greater diagnostic accuracy compared to Mayo with AUCs of 0.76 (95 % CI 0.70-0.81) using logistic regression and 0.73 (0.67-0.79) using random forest methods. Random forest and logistic regression models demonstrated improved risk reclassification with median cNRI of 0.21 (Q1 0.20, Q3 0.23) and 0.21 (0.19, 0.23) compared to Mayo for malignancy. CONCLUSIONS: A combined biomarker, radiomic, and clinical risk factor model provided greater diagnostic accuracy of IPNs than Mayo. This model demonstrated a strong ability to reclassify malignant IPNs. Integrating a combined approach into the current diagnostic algorithm for IPNs could improve nodule management.