三种模式识别模型诊断腰椎间盘突出症受压神经根的准确率

Accuracy of three kinds of pattern recognition models in the diagnosis of nerve root compression in lumbar disc herniation

背景:目前关于腰椎间盘突出症受压神经根的客观检查都面临着诊断能力不足的问题,将模式识别技术与表面肌电技术结合将为提高受压神经根的诊断正确率提供新思路。目的:通过3种不同的模式识别方法建立腰椎间盘突出症受压神经根的表面肌电识别模型,计算3种模型的诊断准确率并分析不同模式识别技术的应用特点。方法:采集2015年10月至2016年10月住院并接受手术治疗的24例L4/L5节段椎间盘突出合并L5神经根受压和23例L5/S1节段椎间盘突出合并S1神经根受压患者的表面肌电参数,应用逻辑回归方程、决策树和人工神经网络建立受压神经根的识别模型,计算3种模型的灵敏度、特异度和诊断正确率,通过受试者工作特征曲线比较3种模型的诊断正确率。结果与结论:(1)逻辑回归方程最终建立了一个三参数的诊断模型,其诊断率从85.7%-100%,平均为93.6%,该诊断方程的灵敏度和特异度分别为0.98和0.92;(2)卡方自交互侦测决策树诊断模型的诊断率为42.86%-85.71%,平均为66.43%,该诊断方程的灵敏度和特异度分别为0.77和0.56;(3)分类回归决策树诊断模型的诊断率为57.14%-85.71%,平均为72.14%,该诊断方程的灵敏度和特异度分别为0.71和0.73;(4)神经网络诊断模型诊断率为85.7%-100%,平均为92.14%,该诊断方程的灵敏度和特异度分别为0.93和0.92;(5)受试者工作特征曲线的曲线下面积评价3种分类模型时,神经网络为0.98,逻辑回归方程为0.97,决策树为0.90;(6)结果表明,神经网络模型与逻辑回归模型识别受压神经根的正确率均非常满意,明显高于MRI的检查结果,其中神经网络模型的诊断效能更加稳定,故其可以作为一种新的定位诊断受压神经根的辅助方法;在没有条件建立神经网络诊断模型的情况下,逻辑回归同样非常适用;决策树在重要危险因素的筛选方面性能突出,其可以和其他方法联合使用提高识别准确率。

BACKGROUND： Nerve root compression of lumbar disc herniation is difficult to diagnose. Pattern recognition technology combined with surface electromyography will provide new ideas for improving the diagnostic accuracy of compressed nerve roots. OBJECTIVE： To establish the recognition model of the nerve root compression of lumbar disc herniation through three kinds of pattern recognition methods, and to analyze the diagnostic accuracy of the models.METHODS： Twenty-four cases of disc herniation at L4/L5 segments combined with L5 nerve root compression and 23 cases of disc herniation at L5/S1 segments combined with S1 nerve root compression from October 2015 to October 2016 were enrolled. The surface electromyography parameters were collected and the Logistic regression equation, decision tree and artificial neural network were used to establish the identification model of compressed nerve roots. The sensitivity, specificity and diagnosis accuracy of the three models were calculated. The diagnosis accuracy was compared by receiver operating characteristic curve. REEULTS AND CONCLUSSION：（1） The logistic regression model had established the three models and the accuracy increased from 85.7% to 100%, with an average of 93.6%, and the sensitivity and specificity of the model was 0.98 and 0.92, respectively.（2） The Chi-squared Automatic Interaction Detector showed an accuracy of 42.86%-85.71%, with an average of 66.43%, the sensitivity and specificity of the model was 0.77 and 0.56, respectively.（3） The Classification and Regression Tree showed an accuracy of 57.14%-85.71%, with an average of 72.14%, the sensitivity and specificity of the model was 0.71 and 0.73, respectively.（4） The neural network model showed an accuracy of 85.7%-100%, with an average of 92.14%, and the sensitivity and specificity of the model were 0.93 and 0.92, respectively.（5） The area under the Receiver Operating Characteristic Curve was used to evaluate the three models, and the neural network was 0.98, the logistic regression was 0.97, and the decision tree was 0.90.（6） These results indicate that both neural network model and the logic regression model show satisfactory results in recognition of the compressed nerve roots, which are superior to MRI. The neural network model is more stable and it may be a more suitable auxiliary method for the diagnosis of nerve root compression. The Logistic regression model is suitable when no neural network diagnostic model is established. The decision tree shows a good performance in the screening of risk factors, and which can be combined with other methods to improve the recognition accuracy.