Evaluating Imputation Techniques for Survival Data Utilizing Kaplan-Meier Curves

Area of Science:

• Epidemiology
• Biostatistics
• Data Science

Background:

• Cancer registries collect vital patient data but often suffer from missing variables.
• Incomplete data hinders accurate survival probability analyses.
• Existing imputation methods are typically evaluated only on feature-wise errors.

Purpose of the Study:

• To present a novel approach for evaluating imputation methods in cancer survival analysis.
• To assess the data distribution learned by imputation techniques.
• To improve the quality of survival analyses using imputed data.

Main Methods:

• Utilized Kaplan-Meier (KM) curves to estimate survival probabilities.
• Stratified data by Union for International Cancer Control (UICC) tumor stage.
• Compared KM curves from known vs. imputed UICC stages using log-rank test, Manhattan distance, and maximum absolute distance.

Main Results:

• The Miss Forest imputer demonstrated the best performance across all evaluation metrics for UICC stage II.
• KM curve comparisons showed alignment between imputed and known data for UICC stage II.
• The proposed evaluation metrics effectively assessed imputation quality for survival analysis.

Conclusions:

• The developed metrics aid epidemiological researchers in selecting imputation methods that preserve survival probability trends.
• Accurate imputation is crucial for reliable cancer survival analysis.
• Miss Forest shows promise for imputing cancer registry data for survival studies.