Interpretable Machine Learning for Survival Analysis

Area of Science:

• Machine Learning
• Artificial Intelligence
• Biostatistics

Background:

• The proliferation of complex
• black box
• machine learning (ML) models necessitates the development of interpretable machine learning (IML) or explainable artificial intelligence (XAI) techniques.
• IML is vital for survival analysis in healthcare, ensuring transparency, accountability, and fairness in clinical decision-making, treatment development, and risk prediction.
• Lack of accessible IML methods hinders the adoption of ML for time-to-event data analysis.

Purpose of the Study:

• To provide a comprehensive review of existing IML methods applicable to survival analysis.
• To adapt and detail the application of common IML techniques (ICE, PDP, ALE, feature importance, Friedman's H-interaction) for survival outcomes.
• To offer a practical guide for researchers using IML in survival analysis.

Main Methods:

• Systematic review of IML literature within the general IML taxonomy.
• Formal adaptation of established IML methods for survival data.
• Empirical application of selected IML methods to breast cancer recurrence data (GBSG2).

Main Results:

• A structured overview of IML techniques suitable for survival analysis is presented.
• Demonstration of how standard IML methods can be effectively modified for time-to-event predictions.
• Practical insights gained from applying IML to real-world breast cancer data.

Conclusions:

• This work bridges the gap between IML methodologies and their practical implementation in survival analysis.
• The adapted IML methods enhance the understanding of survival models, facilitating bias detection and feature influence identification.
• The tutorial application empowers researchers to leverage IML for more trustworthy and interpretable survival predictions in medical contexts.