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

Abstract

Accurate risk assessment of Venous Thromboembolism (VTE) holds significant value for clinical decision-making. However, traditional scoring systems relying on linear assumptions and expert experience, along with machine learning models constrained by shallow architectures, fail to effectively model the high-order nonlinear interactions and local dynamic correlations among multidimensional medical features. To address the systematic deficiency in multi-dimensional feature integration of existing VTE data, this study constructed VTE $\_$ Data - a dataset encompassing multi-dimensional features - based on 113,836 clinical records from a hospital. For VTE risk assessment, we propose a deep learning model integrating separable self-attention and the improved Spatial-Shift Multi-Layer Perceptron (SSA-sMLP). The separable self-attention module enables dynamic cross-dimensional feature interaction modeling through dynamic context vectors and a linear decoupling strategy. The improved Spatial-Shift MLP ($S^2$-MLPv2) employs parameter-free shift operations to reorganize different feature subsets, combined with Split Attention for adaptive weight allocation, thereby precisely capturing local non-linear associations. Experimental evaluations on VTE $\_$ Data using the Caprini RAM (2010) demonstrated that the proposed model (87.99 %) achieves 33.95 % improvement in accuracy over the existing best model, along with superior robustness (F1-score: 65.9 %), while maintaining computational efficiency comparable to mainstream models. By modular integration of separable self-attention and $S^2$-MLPv2 architecture, the SSA-sMLP achieves dual enhancement in feature interaction modeling efficiency and precision, providing an innovative solution that balances computational efficiency and model performance for medical VTE risk assessment tasks.