Interpretable machine learning model for digital lung cancer prescreening in Chinese populations with missing data.

Area of Science:

• Oncology
• Artificial Intelligence
• Public Health

Background:

• Lung cancer remains a leading cause of cancer mortality globally.
• Early detection significantly improves patient outcomes and survival rates.
• Current lung cancer screening methods have limitations, particularly in resource-limited settings.

Purpose of the Study:

• To develop and validate an interpretable artificial intelligence model for lung cancer digital prescreening.
• To improve lung cancer detection rates using electronic health records (EHR).
• To create a cost-effective and accessible lung cancer screening tool for primary healthcare.

Main Methods:

• Development of a Bayesian network model named BOUND (Bayesian network for large-scale lung cancer Digital prescreening).
• Utilized a comprehensive EHR dataset from China, including 905,194 individuals.
• Employed Bayesian network uncertainty inference for risk prediction and identification of high-risk factors, including performance with missing data.

Main Results:

• Achieved an Area Under the Curve (AUC) of 0.866 in internal validation.
• External validation yielded AUCs of 0.848 (time-based) and 0.841 (geography-based).
• Demonstrated robust performance with AUCs ranging from 0.827 to 0.746 in datasets with 10%-70% missing data; improved detection rates up to 6.8 times.

Conclusions:

• BOUND is an interpretable, robust, and effective AI model for lung cancer digital prescreening.
• The model shows strong calibration and clinical utility, particularly in handling missing data.
• BOUND offers a non-radiative, cost-effective solution to address lung cancer screening inequities in primary healthcare.