医疗卫生装备2026，Vol.47Issue(5)：11-20,10.DOI:10.19745/j.1003-8868.2026069
基于改进长短期记忆网络模型的女性压力性尿失禁辅助诊断研究
Auxiliary diagnosis of female stress urinary incontinence based on improved long short-term memory network model
摘要
Abstract
Objective To develop an improved long short-term memory(LSTM)network model to aid in the diagnosis of female stress urinary incontinence(SUI).Methods A CL-LSTM model was constructed with a LSTM network as the baseline model and the introduction of a cross-attention(CA)module and a logistic regression(LR)model,which was composed of a feature extraction module,a feature fusion module and a task output module.Firstly,the CA module was used to implement cross-modal interaction between pelvic floor surface electromyography(sEMG)signals and pelvic floor pressure signals.Global temporal features were extracted from the bimodal temporal signals by the LSTM network,then the SUI priority probability was determined by LR-based statistical modeling of the statistical characteristics from the two types of signals such as peak value,mean value,coefficient of variation and endurance ratio;secondly,the global temporal feature vector from the LSTM network was concatenated with the SUI probability output by the LR model,and then fed into a multi-layer perceptron for nonlinear mapping;finally,the output of the multi-layer perceptron was normalized to obtain the classification results for SUI and healthy controls.To evaluate the performance of the CL-LSTM model in assisting with the diagnosis of female SUI,it was compared with LR,support vector machine(SVM)and LSTM model,and analyses of interpretability and feature importance were carried out.Results The CL-LSTM model achieved an accuracy of 87.50%,sensitivity of 85.71%,specificity of 88.89%and an area under the curve(AUC)of 0.964,outperforming the LR,SVM and conventional LSTM models.Interpretability analysis showed that the model mainly focused on the instantaneous peak values of sEMG signals during the fast-twitch phase while low-frequency fluctuations and stable plateau characteristics of pressure signals during the endurance phase.Feature importance analysis indicated that the mean value and variability in the slow-twitch phase had the highest weights,followed by the maximum value in the fast-twitch phase,while the average value during the endurance phase also made a substantial contribution.Conclusion The proposed CL-LSTM model has high diagnostic accuracy and AUC in the assisted diagnosis of female SUI and a certain degree of interpretability,thereby showing potential application value in outpatient screening and follow-up evaluation of therapeutic efficacy.[Chinese Medical Equipment Journal,2026,47(5):11-20]关键词
压力性尿失禁/盆底表面肌电/盆底压力/长短期记忆网络/交叉注意力/逻辑回归/深度学习Key words
stress urinary incontinence/pelvic floor surface electromyography/pelvic floor pressure/long short-term memory network/cross-attention/logistic regression/deep learning分类
医药卫生引用本文复制引用
韩其成,李武森,陈文建,蒋玉梅,马掌印..基于改进长短期记忆网络模型的女性压力性尿失禁辅助诊断研究[J].医疗卫生装备,2026,47(5):11-20,10.基金项目
江苏省前沿技术研发计划项目(BF2024078) （BF2024078）
