Distance-Metric Learning for Personalized Survival Analysis.

Area of Science:

• Biostatistics
• Machine Learning
• Medical Informatics

Background:

• Accurate time-to-event prediction is crucial in healthcare, especially with right-censored data.
• Current machine learning models (e.g., random survival forests, neural networks) lack interpretability for survival predictions.
• There is a need for interpretable methods in personalized survival analysis.

Purpose of the Study:

• To propose a novel, interpretable method for personalized survival prediction using weighted nearest neighbors.
• To develop a model that provides explanations for individual predictions by identifying influential data points.
• To evaluate the proposed method's performance against established survival prediction techniques.

Main Methods:

• A weighted nearest neighbors approach is proposed for survival prediction.
• Model fitting involves optimizing weights by learning an appropriate metric.
• Interpretability is achieved by presenting influential data points and their associated weights for each prediction.

Main Results:

• The weighted nearest neighbors method demonstrates competitive predictive performance.
• Strengths and weaknesses were analyzed using simulated data.
• The method was successfully applied to two real-world breast cancer patient datasets.

Conclusions:

• The proposed weighted nearest neighbors method offers an interpretable alternative for survival prediction.
• This approach enhances transparency in personalized risk assessment for time-to-event data.
• The method shows promise for clinical applications, particularly in oncology.