CondiS: A conditional survival distribution-based method for censored data imputation overcoming the hurdle in...

Area of Science:

• Biomedical research
• Machine learning
• Survival analysis

Background:

• Machine learning (ML) is increasingly used in biomedical research, but standard ML methods struggle with censored time-to-event data.
• Censoring in survival data, common in clinical studies, requires specialized handling to prevent inaccurate or biased ML results.

Purpose of the Study:

• To develop a general-purpose method for imputing censored survival data to facilitate downstream machine learning analysis.
• To propose novel imputation techniques that accurately handle censored observations, enabling direct application of ML algorithms.

Main Methods:

• Introduced CondiS (Conditional Survival Distributions) using Kaplan-Meier estimators to impute censored survival times.
• Extended CondiS to CondiS-X by incorporating covariates via ML modeling for improved imputation accuracy.
• Validated methods through extensive simulations and two real-world cancer datasets.

Main Results:

• CondiS and CondiS-X achieved smaller prediction errors and higher concordance with true survival times compared to existing methods.
• Demonstrated the practical utility and advantages of the proposed methods on real-world cancer data.
• The methods allow for direct application of standard ML techniques after imputation.

Conclusions:

• CondiS and CondiS-X offer effective solutions for handling censored survival data in ML-based biomedical research.
• These imputation methods enhance the reliability and accuracy of ML analyses involving time-to-event data.
• A user-friendly R package is available, supporting broader adoption in big data-driven biomedical studies.