Missing data imputation and synthetic data simulation through modeling graphical probabilistic dependencies between...

Area of Science:

• Biostatistics
• Cancer Research
• Machine Learning

Background:

• Population-based breast cancer (BC) survival studies face challenges with missing predictive variables (e.g., Stage) and small sample sizes due to class imbalance.
• These issues necessitate advanced data modeling and simulation techniques for reliable analysis.

Purpose of the Study:

• To introduce ModGraProDep, a graphical modeling (GM) procedure designed to address missing data and small sample sizes in BC survival studies.
• To compare the performance of GM-derived models against conventional classification, machine learning, and oversampling algorithms.

Main Methods:

• Developed and applied the ModGraProDep procedure using graphical modeling on two validated BC datasets.
• Assessed performance in scenarios with data missing completely at random (MCAR) and missing not at random (MNAR).
• Compared GM models (GM.SAT, GM.K1, GM.TEST) against standard algorithms for missing data imputation and synthetic data simulation.

Main Results:

• GM models, particularly GM.K1 and GM.TEST, demonstrated superior prediction performance compared to other algorithms in both MCAR and MNAR scenarios.
• While GM.SAT showed strong performance, its direct predictions might yield unreliable conclusions; however, synthetic data from it was the poorest strategy.
• The remaining GM models offered a better alternative to oversampling for synthetic data generation.

Conclusions:

• The presented GM procedure is recommended for one-variable imputation/prediction of missing data in BC survival studies.
• GM-derived synthetic datasets can effectively augment small or imbalanced datasets, proving valuable for clinical applications like predictive risk analysis.