Maximum Likelihood Estimation of Flexible Survival Densities with Importance Sampling

Area of Science:

• Biostatistics
• Machine Learning
• Data Science

Background:

• Survival analysis is crucial for time-to-event data with censoring.
• Recent models offer scalability and relax proportional hazards assumptions.
• These advanced models are highly sensitive to hyperparameter choices, demanding extensive tuning.

Purpose of the Study:

• To develop a survival analysis approach that eliminates the need for hyperparameter tuning.
• To reduce the burden on practitioners by simplifying model selection and optimization.
• To provide a robust and high-performing survival analysis method.

Main Methods:

• A novel survival analysis approach is proposed.
• The method avoids the need for tuning hyperparameters like mixture assignments and bin sizes.
• Empirical studies were conducted to evaluate the approach against existing baselines.

Main Results:

• The proposed survival analysis approach matches or outperforms established baseline methods.
• The method demonstrates robustness across several real-world datasets.
• The empirical studies confirmed the sensitivity of existing models to hyperparameter choices.

Conclusions:

• The new survival analysis technique simplifies the analysis of time-to-event data.
• It offers a more accessible and efficient alternative to current hyperparameter-intensive models.
• This approach is suitable for various real-world applications requiring survival analysis.