Lung Cancer Detection Using Bayesian Networks: A Retrospective Development and Validation Study on a Danish...

Area of Science:

• Oncology
• Biostatistics
• Medical Informatics

Background:

• Lung cancer (LC) is a leading global cause of cancer mortality.
• Current LC screening relies on age and smoking history, but more advanced risk models are needed.
• Bayesian networks (BNs) offer a probabilistic approach for disease detection.

Purpose of the Study:

• To develop and evaluate a Bayesian network (BN) model for lung cancer (LC) detection.
• To assess the resilience of BN models to missing data.
• To compare BN model performance against a traditional machine learning (ML) model.

Main Methods:

• Analysis of 9940 patient records from Southern Denmark (2009-2018).
• Inclusion of variables: age, sex, smoking status, and laboratory results.
• Experiments involved varying missing data percentages (0-30%) and BN configurations.

Main Results:

• Bayesian network models maintained stable performance (AUC 0.737-0.757) across missing data levels, comparable to the ML model (AUC 0.77).
• BN structure and discretization methods had minimal impact on model efficacy.
• BNs demonstrated good calibration and clinical utility, particularly for predicted risks above 5%.

Conclusions:

• BN models are robust and effective even with substantial missing data (up to 30%).
• BNs offer comparable performance, calibration, and clinical utility to ML models for LC detection.
• BNs represent a promising methodology for developing future lung cancer risk prediction models.