WindowSHAP: An efficient framework for explaining time-series classifiers based on Shapley values.

Area of Science:

• Machine Learning Interpretability
• Time-Series Analysis
• Clinical Informatics

Background:

• Explaining black-box machine learning models, especially deep learning, remains a significant challenge.
• Interpreting time-series predictive models is crucial for high-stakes clinical applications.
• Existing explanation methods often fail with time-varying features in time-series data.

Purpose of the Study:

• Introduce WindowSHAP, a model-agnostic framework to explain time-series classifiers using Shapley values.
• Address computational complexity and enhance explanation quality for long time-series data.
• Provide a method applicable to clinical time-series data.

Main Methods:

• Developed WindowSHAP by partitioning time-series data into sequential windows.
• Implemented three algorithms: Stationary, Sliding, and Dynamic WindowSHAP.
• Evaluated against KernelSHAP and TimeSHAP using perturbation and sequence analyses metrics on clinical data (TBI, critical care).

Main Results:

• WindowSHAP significantly reduces computational complexity, decreasing CPU time by 80% for 120 time steps compared to KernelSHAP.
• Demonstrated superior performance in explaining clinical time-series classifiers based on quantitative metrics.
• Dynamic WindowSHAP algorithm effectively focuses on critical time steps, yielding more understandable explanations.

Conclusions:

• WindowSHAP accelerates Shapley value calculations for time-series data.
• The framework provides higher quality and more understandable explanations for clinical time-series classifiers.
• WindowSHAP offers a practical solution for interpreting complex machine learning models in healthcare.