Maximizing Lung Cancer Screening in High-Risk Population Leveraging ML-Developed Risk-Prediction Algorithms: Danish Retrospective Validation of LungFlag.

Abstract

Background: Early detection of lung cancer (LC) is crucial for curative treatment, but current screening methods face challenges due to high costs and poor adherence. Artificial intelligence tools, such as the LungFlag model, uses routine clinical data for innovative risk stratification. This study validates LungFlag in Danish high-risk populations to assess its potential in LC screening.

Methods: This retrospective study included data from 2 populations in Southern Denmark (2013-2021): (A) LC fast-track clinic patients (∼25% LC incidence) and (B) outpatients followed with chronic obstructive pulmonary disease (COPD) (∼6% LC incidence). Data included laboratory results, comorbidities, body mass index (BMI), and smoking history from up to 3 years prior to the index date. LungFlag's performance was compared to the PLCOm2012 model. Model interpretation was conducted using Shapley additive explanation (SHAP) values, and risk stratification was analyzed by age.

Results: In Population A, 5271 LC cases were identified from 18,600 patients, with a stable LC incidence of 28%. In Population B, LC incidence varied by index-date approach: 6.6% using the diagnosis date and 2.1% using the first visit approach. LungFlag outperformed PLCOm2012 in Population A (AUC: 0.63 vs. 0.60) and showed slightly higher sensitivity in Population B, though differences were minor. Key predictors included smoking, age, and COPD. High-risk individuals identified by LungFlag were generally younger compared to using PLCOm2012.

Conclusion: LungFlag demonstrates promise as a decision-support tool in detecting LC, particularly for COPD patients, who lack systematized screening. However, prospective real-world studies are needed to confirm its effectiveness and clinical value.