← Advanced Bronchoscopy

Risk Stratification of Incidental Pulmonary Nodules

IPBronch Review

🎯 Background & Rationale

Incidental pulmonary nodules (IPNs) are frequently discovered in routine clinical care, creating a major diagnostic challenge. While navigational and robotic-assisted bronchoscopy have lowered the threshold for biopsying smaller, peripheral lesions, traditional clinical prediction models (like the VA model) often struggle with intermediate-risk nodules, leading to heterogeneous management, prolonged surveillance delays, and unnecessary invasive procedures. This study evaluates Bronchosolve, a fully automated, imaging-only artificial intelligence tool that analyzes full CT examinations without demographic or clinical variables, to determine its ability to improve risk stratification over standard clinical models.

👥 Study Design & Population

  • Study Type: Retrospective Cohort Study
  • Population (PICO framework):
    • Population: 187 adult patients (median age 67 years, 88.2% current/former smokers) with incidentally detected pulmonary nodules measuring 6 to 30 mm, enrolled in registries at two large health care systems between 2006 and 2019, with final clinicopathologic adjudication.
    • Intervention: Application of the fully automated AI imaging platform, Bronchosolve, to index CT scans.
    • Comparison: Performance compared against the established VA clinical prediction model (and evaluated within the 8–15 mm intermediate-sized nodule subgroup).
    • Outcomes: Discriminative performance via Area Under the Receiver Operating Curve (AUC), sensitivity, specificity, and reclassification rates for intermediate-risk nodules.

📈 Methodology & Rigor

  • Data Handling: Evaluated a consecutive dataset of 187 patients (70% malignancy prevalence, with 85% histopathologic confirmation). Excluded patients with active extrathoracic cancer.
  • Statistical Analysis: Model performance was measured using standard diagnostic metrics: AUC, sensitivity, and specificity. AUC comparisons between Bronchosolve and the VA model utilized bootstrapped CIs and the DeLong test.
  • Reclassification Analysis: Specifically targeted patients with nodules measuring 8 to 15 mm categorized as intermediate-risk by the VA model (5% to 65% malignancy risk) to quantify clinical utility in a diagnostically challenging subset. Python v3 and the sklearn library were utilized.

🔬 Key Findings

  • Overall Discrimination: Bronchosolve achieved an AUC of 0.895 in the complete dataset, compared to an AUC of 0.870 for the VA model.
  • Intermediate-Sized Nodules (8–15 mm): In this subgroup (74 of 78 initially classified as intermediate risk by the VA model; 47% malignant), Bronchosolve demonstrated wider separation with an AUC of 0.808 versus 0.701 for the VA model (P = 0.156).
  • Reclassification Performance: At a high-sensitivity operating threshold within the intermediate-risk 8–15 mm subgroup, Bronchosolve:
    • Reclassified 51 patients to high risk (capturing 33 of 37 malignancies [89%]).
    • Reclassified 23 patients to low risk (capturing 19 of 37 benign nodules [51%]).
    • Up-classified 18 benign nodules as potential high risk.
  • Diagnostic Delays: The median interval from index CT to diagnosis for malignant cases was 146 days, highlighting pre-diagnostic delays in standard care pathways.

⚖️ Critical Appraisal

  • Strengths: Evaluates a truly automated, end-to-end imaging AI tool that operates independently of manual clinical data entry; includes external validation across heterogeneous multi-manufacturer CT scanners; and focuses heavily on intermediate-sized/intermediate-risk nodules where clinicians face the highest uncertainty.
  • Limitations: Retrospective design with a modest sample size and a high prevalence of malignancy (70%), which enriches the cohort for diagnostically challenging cases but limits generalizability to routine screening settings where malignancy rates are lower. Ground-glass nodules were underrepresented due to study entry requirements. The study measures discriminative performance rather than prospective, patient-centered clinical outcomes.

💡 The Clinical Bottom Line

As advanced bronchoscopy platforms (robotic-assisted and navigational bronchoscopy) expand our reach to smaller peripheral nodules, preprocedural risk stratification remains critical. Bronchosolve demonstrates robust imaging-only discrimination that outperforms traditional clinical prediction models, particularly in challenging intermediate-sized nodules (8–15 mm). For the interventional pulmonologist, this tool offers a rapid, single-time-point automated risk assessment from the index CT scan that can help rationally guide decisions on whether to escalate to invasive tissue acquisition or safely pursue surveillance, helping bridge the gap in pre-diagnostic uncertainty.


Background: Incidental detection is the most common pathway through which pulmonary nodules are identified. With the advancement of navigational and robotic-assisted bronchoscopy, existing risk stratification models often lack sufficient discrimination power, particularly for intermediate-risk nodules. Bronchosolve is a fully automated, imaging-only risk stratification tool that was previously validated in lung cancer screening cohorts. Methods: We applied Bronchosolve to a cohort of adults with incidentally detected pulmonary nodules identified from registries at 2 large health care systems between 2006 and 2019. Nodules were classified as malignant or benign based on biopsy or extended clinical follow-up. Discrimination was assessed and compared with the VA Model. Sensitivity and specificity were reported at the Youden and high-sensitivity operating points. Additional analysis was performed in cases with nodule diameter 8 to 15 mm. Results: Among 187 patients, malignancy prevalence was 70%. Bronchosolve achieved an AUC of 0.895 versus 0.870 for the VA model. Among nodules 8 to 15 mm, discrimination differences were larger (AUC 0.808 vs. 0.701). In nodules of 8 to 15 mm in diameter with an intermediate risk for malignancy by the VA model, Bronchosolve reclassified 51 patients to high risk and 23 to low risk; 33/37 (89%) malignancies were up-classified and 19/37 (51%) benign nodules were down-classified. Conclusion: In a nonscreening setting, Bronchosolve demonstrated robust discrimination for lung nodule malignancy risk using imaging alone compared with a widely used clinical risk model, with more pronounced separation in intermediate-sized nodules. In this subgroup, Bronchosolve also yielded clinically meaningful risk reclassification, with potential implications for downstream management.
Read the full paper ↗

Keep reading

N opens the next one