Interpretable Machine Learning for Feature-Based Classification of Platelet Activation in Rotary Blood Pumps.

Area of Science:

• Biomedical Engineering
• Computational Fluid Dynamics (CFD)
• Machine Learning

Background:

• Thrombosis in rotary blood pumps is challenging to predict due to complex flow dynamics.
• Existing computational models struggle to link specific flow features to thrombus initiation and growth.
• A gap exists in understanding how flow characteristics influence thrombotic risk in blood pumps.

Purpose of the Study:

• To introduce a feature-based supervised machine learning framework for assessing thrombogenic risk.
• To spatially map activation-based thrombotic risk using CFD-derived flow features.
• To develop a computationally efficient and mechanistically transparent method for thrombogenicity screening.

Main Methods:

• A logistic regression model with a structured feature-selection pipeline was employed.
• The framework derived a compact, physically interpretable feature set, including nonlinear combinations.
• Training utilized spatial risk patterns from a validated platelet-activation-based thrombosis model.

Main Results:

• The model accurately reproduced labeled risk distributions and identified key flow features linked to thrombosis.
• When applied to a centrifugal pump, the model predicted plausible thrombosis-prone regions.
• Interpretable machine learning effectively linked local flow features to activation-based thrombotic risk.

Conclusions:

• The proposed framework complements existing physics-based models for thrombosis and platelet activation.
• It offers a methodological foundation for integrating interpretable machine learning into CFD-driven thrombogenicity analysis.
• This approach facilitates efficient thrombogenicity screening and aids in blood pump design workflows.