An Interpretable Two-Phase Modeling Approach for Lung Cancer Survivability Prediction

Area of Science:

• Oncology
• Biostatistics
• Data Science

Background:

• Lung cancer survival prediction often focuses on status or duration separately.
• Interpretable models combining both aspects are needed for clinical decision-making.

Purpose of the Study:

• To develop and validate a two-phase framework for predicting lung cancer survival status and duration.
• To compare the performance of interpretable models (GLM) against complex models (XGBoost, ANN).
• To identify and quantify key factors influencing short-term lung cancer survival.

Main Methods:

• Utilized Surveillance, Epidemiology, and End Results (SEER) data (2010-2017).
• Employed general linear models (GLM), extreme gradient boosting (XGBoost), and artificial neural networks (ANN).
• Applied data balancing (SMOTE variants, OS), feature selection (LASSO, Random Forest), and one-hot encoding.

Main Results:

• A computationally efficient general linear model (GLM) performed comparably to complex models.
• Quantified the impact of individual features on survival status and duration using GLM coefficients.
• Visualized top factors affecting survival odds via odds ratio changes.

Conclusions:

• The proposed two-phase framework provides a performant, efficient, and interpretable approach to lung cancer survival prediction.
• GLM offers a viable alternative to black-box models for clinical interpretability.
• This study comprehensively explores short-term lung cancer survival prediction using a novel two-phase methodology.