262MO - Real-world validation of artificial intelligence-defined lung nodule malignancy score (qXR-LNMS) in predicting risk of lung cancer

Background

Low-dose computed tomography (LDCT), though recommended for lung cancer screening has limited uptake due to cost and accessibility in most parts of the world. We validated an AI-based LNMS to detect high risk incidental pulmonary nodule/s (IPN) across 5 countries (Egypt, India, Indonesia, Mexico, Turkey).

Methods

CREATE (NCT05817110), is a prospective, observational study for individuals >35 yrs with IPN (size ≥8 to ≤30 mm) on chest xray (CXR). A total of 713 individuals; high LNMS (499; 70%) and low LNMS (214; 30%) were enrolled between Apr-23 and Dec-24 to cross the threshold of at least 20% positive predictive value (PPV) and 70% negative predictive value (NPV). The primary and secondary outcomes included PPV and NPV of LNMS against the risk of malignancy assessed by radiologists using LDCT and binarized risk categories based on Lung-RADS score, and Mayo clinic model with 95% CI using Wilson’s score method. The PPVs and NPVs of LNMS by clinicodemographic characteristics were assessed. In the ongoing phase II of this study, participants will be followed up for 2 yrs from the date of first CT scan.

Results

The median ages of participants with high and low LNMS were 58 and 60 yrs respectively, with 71.7% non-smokers and 2.8% with family history of lung cancer. Overall, the PPV and the NPV of LNMS against radiologists assessment were 54.1% (95% CI: 49.7–58.4) and 93.5% (95% CI: 89.3–96.1). The agreement between Mayo clinic model and LNMS was observed in 70.7% (n = 504) individuals and Spearman’s correlation was 0.246. Results across key subgroups were consistent with all PPV and NPV point estimates exceeding the prespecified threshold for success.

Table 262MO

Conclusions

The observed PPV and NPV demonstrate the utility of qXR-LNMS to predict the likely risk of benign and malignant IPN on CXR against radiologist assessment of LDCT. These results support the use of AI-enabled triaging of incidental CXR to optimize lung cancer screening workflows across diverse healthcare settings.

Clinical trial identification

NCT05817110.

Editorial acknowledgement

The authors would like to thank Prajakta Nachane and Neelam Joglekar of Fortrea Scientific Pvt Ltd for medical writing support that was funded by AstraZeneca International.

Legal entity responsible for the study

AstraZeneca International.

Funding

AstraZeneca International.

Disclosure

H.K. Gonuguntla: Financial Interests, Personal, Invited Speaker: Fujifilm, Erbe, Pfizer. R. Cordova: Financial Interests, Institutional, Other, Received fee from AstraZeneca for each patient enrolled in the study through institution: AstraZeneca. S. McCutcheon, A. Saha, P. Kantharaju: Financial Interests, Personal, Full or part-time Employment: AstraZeneca. S. Sen, R. Agrawal: Financial Interests, Personal, Full or part-time Employment: Qure.ai. All other authors have declared no conflicts of interest.