What question did this study set out to answer?

This study evaluates the effectiveness of Brock, Mayo, and VA algorithms in predicting lung cancer risk linked to pulmonary nodules.

May 20, 2026

B80-4-11 Pragmatic Application of Brock, Mayo and VA Nodule Prediction Algorithms on Lung Cancer Diagnoses

Key Points

This study evaluates the effectiveness of Brock, Mayo, and VA algorithms in predicting lung cancer risk linked to pulmonary nodules.
Retrospective analysis of 481 patients with abnormal chest imaging referred for evaluation.
Patients underwent PET-CT scans and biopsies, with follow-up classification of diagnoses.
Risk scores calculated using gold standard and real-world methods to evaluate algorithm performance.
Brock and VA models showed significant differences identifying malignant nodules compared to others.
Identified thresholds for cancer detection ranged from 25% to 65%, with notable malignancy rates below these thresholds.
44% to 55% of cases below thresholds were still malignant, highlighting poor clinical discrimination.

Abstract

Abstract Rationale Lung cancer risk prediction algorithms help clinicians decide when to biopsy pulmonary nodules. They aim to expedite cancer diagnoses while reducing unnecessary biopsies in benign cases. This study assessed the Brock, Mayo, and VA lung cancer risk prediction algorithms to provide real-world data supporting biopsy decisions in patients with nodules suspicious for malignancy. Methods We retrospectively analyzed patients referred to the Pulmonary Division at the Miami Veterans Affairs Medical Center for abnormal chest imaging between 2013 and 2018. Of 481 patients, 329 had findings concerning for lung cancer and underwent positron emission tomography-computed tomography (PET-CT) and biopsy. After ≥7 years of follow-up, patients were classified as (1) Malignant (NSCLC, SCLC, metastasis), (2) Other (granulomatous disease/sarcoidosis or infection), or (3) Unchanged/Resolved. Each patient’s risk was calculated using the Brock University, Mayo Clinic, and VA algorithms. Scores were generated using a “gold standard (GS)” method limited to allowed parameters and a “real-world (RW)” method, which truncated values exceeding limits (e.g., nodules 30 mm entered as 30 mm). Violin plots visualized score distributions and comparisons used the Kruskal-Wallis and repeated-measures tests. Results The VA and Brock models showed significant differences between Malignant vs. Other and Malignant vs. Unchanged/Resolved in both GS and RW approaches. The Mayo model was significant for Malignant vs. Unchanged/Resolved only. Violin plots identified thresholds optimizing cancer detection while minimizing benign findings: Brock RW and GS: 25%, Mayo RW: 50%, Mayo GS: 40%, VA RW and GS: 65%. Cancer cases identified below thresholds: Brock RW 47.7%, Brock GS 44.2%, Mayo RW 54.6%, Mayo GS 48.8%, VA RW 50.4%, VA GS 55.1%. Mean cancer nodule’s maximum diameters below thresholds: Brock RW 19.0 mm (6-30), Brock GS 16.8 mm (6-30), VA RW 19.3 mm (6-30), VA GS 19.3 mm (6-29), Mayo RW 15.3 mm (6-25), Mayo GS 13.2 mm (6-24). Percent total cancers below threshold by model; Brock RW 17.4%, Brock GS 24.6%, Mayo RW 25.1%, Mayo GS 17.4%, VA RW 24.2%, VA GS 39.1%. Despite statistical differences between overall groups, 44% to 55% of nodules below the threshold were malignant. Conclusions Risk prediction models accurately differentiate group risk between Malignant versus Unchanged/Resolved. However, current lung cancer risk algorithms do not identify thresholds that balance timely diagnosis against unnecessary biopsies. Despite statistical significance, clinical discrimination remains poor, underscoring the need for improved, real-world calibrated prediction models. This abstract is funded by: None

Ask AI

Helpful

Bookmark