SSA-sMLP: A venous thromboembolism risk prediction model using separable self-attention and spatial-shift multilayer...

Area of Science:

• Medical informatics
• Machine learning
• Cardiovascular research

Background:

• Accurate Venous Thromboembolism (VTE) risk assessment is crucial for clinical decisions.
• Existing methods struggle with high-order nonlinear interactions in multidimensional medical data.
• There's a need for improved multi-dimensional feature integration in VTE risk prediction.

Purpose of the Study:

• To develop a deep learning model for enhanced VTE risk assessment.
• To address limitations in modeling complex interactions within clinical VTE data.
• To improve the accuracy and robustness of VTE risk prediction models.

Main Methods:

• Constructed VTE_Data, a large dataset of 113,836 clinical records.
• Developed a deep learning model integrating separable self-attention and improved Spatial-Shift Multi-Layer Perceptron (SSA-sMLP).
• Utilized dynamic context vectors, linear decoupling, parameter-free shifts, and Split Attention for feature interaction modeling.

Main Results:

• The SSA-sMLP model achieved 87.99% accuracy, a 33.95% improvement over existing models.
• Demonstrated superior robustness with an F1-score of 65.9%.
• Maintained computational efficiency comparable to mainstream models.

Conclusions:

• The SSA-sMLP model effectively models complex feature interactions for VTE risk assessment.
• This approach offers a significant advancement in balancing computational efficiency and predictive performance.
• Provides an innovative solution for medical VTE risk assessment tasks.