A Bayesian network approach for modeling local failure in lung cancer

Area of Science:

• Oncology
• Radiotherapy
• Biostatistics

Background:

• Locally advanced non-small cell lung cancer (NSCLC) has a high rate of local failure after radiotherapy.
• Current dose-volume models have not significantly improved prospective prediction of tumor local failure.
• Biomarker proteins involved in hypoxia and inflammation may predict radiotherapy response.

Purpose of the Study:

• To develop and test a graphical Bayesian network framework for predicting local failure in NSCLC patients treated with radiotherapy.
• To investigate the combined predictive power of physical (dosimetric) and biological (biomarker) factors.
• To interpret the relationships among variables in predicting radiotherapy outcomes.

Main Methods:

• A graphical Bayesian network framework was proposed.
• The framework was tested on two datasets: one retrospective (clinical, dosimetric) and one prospective (clinical, dosimetric, biomarkers).
• Candidate biomarkers were extracted from patient blood samples at various time points.

Main Results:

• The proposed Bayesian network efficiently predicts local failure and interprets variable relationships.
• Integrating heterogeneous physical and biological variables improved model prediction.
• The combined model showed slightly higher performance than individual models, with biological variables contributing most significantly.
• Performance surpassed competing Bayesian-based classifiers on the retrospective dataset.

Conclusions:

• The integrated Bayesian network approach shows potential for predicting post-radiotherapy local failure in NSCLC.
• Combining physical and biological data offers a promising strategy to enhance predictive accuracy.
• Biomarker analysis is crucial for improving radiotherapy outcome predictions in NSCLC.