Abstract Background Accurate malignancy prediction for indeterminate pulmonary nodules (IPNs) referred for biopsy remains challenging in high-risk cohorts, where commonly used prediction models underperform relative to clinician assessment. Research Question Accurately predicting malignancy in biopsy-referred indeterminate pulmonary nodules (IPNs) is challenging. We therefore developed and internally validated two logistic regression models using VERITAS data: (1) a model combining physician gestalt and Brock to test whether integrating clinician judgment with a validated model improves malignancy prediction, and (2) a clinical model using age, smoking history, and nodule size. We evaluated discrimination, calibration, and pathway-specific performance (CT-guided transthoracic needle biopsy TTNB vs non-TTNB). Methods We conducted a secondary analysis of the multicenter VERITAS randomized trial (seven U.S. academic centers), including 210 biopsy-referred IPNs. We trained and internally validated two prespecified logistic regression models: Model 1, which combined physician gestalt (PG) with the Brock Model, and Model 2, which used age, smoking history, and mean nodule diameter. Model performance was assessed by ROC curves, AUC, SN, SP and Brier scores. We also evaluated subgroup performance by biopsy approach, CT-guided transthoracic needle biopsy TTNB vs non-TTNB. Threshold analyses across predicted-probability cutoffs 0.10-0.90 estimated sensitivity, specificity, PPV, and NPV. Internal validation used 200 bootstrap resamples (R, Version 2024.04). Results Malignancy prevalence was 135/210, 64%. Model 1 AUC 0.675 (95% CI 0.589-0.761) exceeded Model 2 AUC 0.629 (95% CI 0.540-0.718). Within Model 1, PG was a significant predictor (χ² = 14.0, p = 0.0001), whereas Brock was not (p = 0.203). Calibration was modest for both (Brier 0.206 for Model 1; 0.220 for Model 2); the bootstrap-corrected calibration slope for Model 1 was 0.977, with a corrected Brier of 0.211. Model 2 showed a similar pattern (0.684 vs 0.530), with comparable Brier scores (∼0.205-0.220). Threshold analysis: At a 0.50 cutoff, Model 1 sensitivity 89.5%, specificity 38.2%, PPV 70.7%, NPV 66.7%, subgroups: Model 1 discrimination was higher in non-TTNB cases with AUC 0.725 as compared to TTNB AUC 0.597; overall accuracy ∼70.5%; Model 2 sensitivity 91.1%, specificity 10.3%, accuracy 62.7%. Exploratory nonlinearity (RCS) in Model 2 yielded modest overall fit (χ² = 10.66, p = 0.0726; C-index 0.643), with no significant nonlinear effects of age or size. Interpretation In a high-risk, biopsy-referred cohort, logistic regression models demonstrated moderate predictive accuracy. Physician gestalt contributed more signal than the Brock score, and performance was heterogeneous by biopsy pathway, favoring non-TTNB cases. These data support integrating structured tools with clinician judgment while motivating next-generation models that incorporate multimodal features. This abstract is funded by: none
Pervaiz et al. (Fri,) studied this question.