MS-CPFI: A model-agnostic Counterfactual Perturbation Feature Importance algorithm for interpreting black-box...

Area of Science:

• Computational biology
• Machine learning in healthcare
• Biostatistics

Background:

• Multi-state processes are crucial for modeling disease progression, but interpreting complex machine learning models used for these processes remains a challenge.
• Existing interpretability methods are often limited to simpler models or specific disease states.
• Interpretability is essential for clinical adoption, regulatory approval, and patient trust in AI-driven healthcare predictions.

Purpose of the Study:

• To introduce a novel, model-agnostic interpretability algorithm, Multi-State Counterfactual Perturbation Feature Importance (MS-CPFI), for general multi-state models.
• To enable the interpretation of feature importance for each transition within complex disease progression pathways.
• To address the need for interpretable machine learning in healthcare, particularly for multi-state processes.

Main Methods:

• Developed MS-CPFI, a model-agnostic algorithm designed for multi-state models (including survival, competing-risks, and illness-death).
• Employed a novel counterfactual perturbation technique to capture non-linear and potentially time-dependent feature effects.
• Validated the algorithm using both simulated data and a real-world breast cancer patient dataset.

Main Results:

• MS-CPFI successfully increased model interpretability, especially in scenarios with non-linear effects, as demonstrated in simulations.
• Application to a breast cancer dataset revealed clinically significant features, differentiating protective factors from risk factors at various disease stages.
• The method effectively provides insights into disease progression by analyzing feature importance across different transitions.

Conclusions:

• MS-CPFI offers a significant advancement in the interpretability of machine learning and deep learning models for multi-state processes.
• The algorithm enhances trust and facilitates the clinical and regulatory acceptance of predictive models in healthcare.
• MS-CPFI provides valuable, stage-specific insights into disease progression, aiding clinical decision-making and research.