基于趋势-残差分解与跨组件注意力的低血压预测网络OA

To address the problem that existing deep learning models hard to fully utilize the complementary information between the slow-changing hemodynamic trends and fast-changing morphological features in physiological signals,we proposed a hypotension prediction network based on trend-residual decomposition and cross-component attention.Firstly,the trend-residual decomposition strategy was employed to explicitly decompose the continuous physiological signals into low-frequency trend components and high-fre-quency residual components,which correspond to the macroscopic evolution of blood pressure and the microscopic pulsation patterns,respectively.Then,a cross-component attention mechanism was designed to construct bidirectional interactions between the two sets of features,capture long-term trends while sensitively detecting early compensatory signs hidden within the high-frequency waveforms.Experiments on the VitalDB dataset demonstrated that the area under the receiver operating characteristic(AUROC)curve was 89.81%,81.30%and 80.59%for 5,10 and 15 min prediction windows,respectively,significantly outperforming the traditional meth-ods.This model can provide doctors with continuous and accurate intraoperative risk quantitative assessment.