Survival Stacking Ensemble Model for Lung Cancer Risk Prediction

Area of Science:

• Oncology
• Biostatistics
• Machine Learning in Healthcare

Background:

• Smoking is the primary risk factor for lung cancer (LC), causing approximately 85% of cases.
• Existing tools like the Lung Cancer Risk Assessment Tool (LCRAT) predict LC risk using multiple factors.
• There is a need for more accessible and easily implementable risk assessment models in clinical practice.

Purpose of the Study:

• To develop and validate a simplified, feature-reduced model for lung cancer risk prediction.
• To improve upon the performance and accessibility of current lung cancer risk assessment tools.
• To enhance the robustness and generalizability of lung cancer risk prediction models through ensemble methods.

Main Methods:

• A simplified stacking ensemble model was developed using a reduced feature set.
• The model was trained and tested on data from two large US cohorts: the National Lung Screening Trial (NLST) and the Prostate, Lung, Colorectal, and Ovarian (PLCO) Cancer Screening Trial.
• Model performance was evaluated using Area Under the Curve (AUC) and percentage of positives detected.

Main Results:

• The proposed simplified model achieved an AUC of 0.799, comparable to the established LCRAT (AUC 0.782).
• In the top 50% of the population, both models detected a similar proportion of cases (0.766 for the new model vs. 0.754 for LCRAT).
• The ensemble approach enhanced model robustness and efficiency.

Conclusions:

• A simplified, feature-reduced lung cancer risk model demonstrates competitive performance and improved accessibility.
• The ensemble method enhances the reliability and generalizability of lung cancer risk prediction.
• This model offers a promising, more implementable alternative for routine healthcare settings.