基于机器学习模型评估关节摩擦音诊断膝关节半月板损伤的应用价值

Application value of joint friction sounds in diagnosing meniscus injury of the knee based on machine learning models

目的探讨基于机器学习模型评估关节摩擦音诊断膝关节半月板损伤的应用价值。方法采用病例对照研究分析2020年8月至2022年10月南京医科大学附属逸夫医院收治的17例半月板损伤(半月板损伤组)和招募的75例无膝关节疾病健康受试者(健康组)的临床资料。在相对安静(峰值不超过40 dB)的环境下采集受试者膝关节摩擦音,将屈-伸-屈运动方式下所得关节摩擦音分割后得到的音频按4∶1的比例进行随机分层抽样,得到训练集(半月板损伤组125段,健康组187段)和测试集(半月板损伤组33段,健康组47段);将坐-站-坐运动方式下所得关节摩擦音分割后得到的音频按4∶1的比例进行随机分层抽样,得到训练集(半月板损伤组81段,健康组164段)和测试集(半月板损伤组20段,健康组40段)。建立线性核支持向量机、径向基函数支持向量机、随机森林和极端随机树4种机器学习模型。分别对训练集数据进行模型的学习训练,通过测试集数据对模型性能进行验证。记录受试者单次采集所需时间和数据分析判读所需时间。分别在试验前及试验后1 d采用Lysholm评分评估受试者膝关节功能。根据测试集结果比较两种运动方式下关节摩擦音诊断半月板损伤的准确率得出更优运动方式。基于更优运动方式下4种模型下所得关节摩擦音诊断半月板损伤的准确率、灵敏度、特异度、F1分数、受试者工作特征曲线下面积(AUC)比较模型的有效性,找出适合本研究数据框架的最佳机器学习模型。观察更优运动方式的最佳机器学习模型下关节摩擦音对半月板损伤的诊断准确率、误诊率及漏诊率。结果单次关节摩擦音采集所需时间为5~10 min[(7.1±1.3)min],数据分析判读所需时间约为1 min。试验前后膝关节Lysholm评分半月板损伤组分别为(75.6±4.0)分和(77.7±3.7)分(P>0.05),健康组分别为(99.6±0.9)分和(99.5±1.0)分(P>0.05)。在线性核支持向量机的模型下,屈-伸-屈和坐-站-坐运动方式的诊断准确率分别为0.775和0.817,径向基函数支持向量机模型下的诊断准确率分别为0.813和0.900,随机森林模型下的诊断准确率分别为0.800和0.867,极端随机树模型下的诊断准确率分别为0.800和0.900,坐-站-坐的诊断准确率均高于屈-伸-屈。在坐-站-坐运动方式下,极端随机树模型的诊断准确率为0.900、灵敏度为0.900、特异度为0.950、F1分数为0.900、AUC为0.942,均高于其余3种模型,极端随机树表现出更好的机器学习效能。在坐-站-坐运动方式下的极端随机树模型中,60段测试集音频(半月板损伤组20段,健康组40段)诊断为半月板损伤组22段(真阳性18段,假阳性4段),健康组38段(真阴性36段,假阴性2段)。关节摩擦音诊断半月板损伤的准确率为0.900,误诊率为0.100,漏诊率为0.100。结论关节摩擦音诊断半月板损伤可缩短诊断时间,增强检查过程的安全性,其基于机器学习的人工智能的诊断模式更加快速、稳定,可以作为膝关节半月板损伤的诊断标志物。

Objective To investigate the application value of joint friction sounds in diagnosing meniscus injury of the knee based on machine learning models.Methods A case-control study was conducted to analyze the clinical data of 17 patients with meniscus injury of the knee(meniscus injury group)admitted to Sir Run Run Shaw Hospital Affiliated to Nanjing Medical University from August 2020 to October 2022,as well as 75 recruited healthy subjects without knee joint diseases(healthy group).The knee joint friction sounds of the subjects were collected in a relatively quiet environment(peak value below 40 dB).The sounds collected in a flexion-extension-flexion mode of exercise were split and divided randomly with a ratio of 4∶1 into the training set(125 segments from the meniscal injury group and 187 segments from the healthy group)and the test set(33 segments from the meniscal injury group and 47 segments from the healthy group).The sounds obtained in a sit-stand-sit mode of exercise were split and divided randomly with a ratio of 4∶1 into the training set(81 segments from the meniscal injury group and 164 segments from the healthy group)and the test set(20 segments from the meniscal injury group and 40 segments from the healthy group).Four machine learning models were built,including support vector machine with linear kernels,radial basis function support vector machine,random forest,and extremely randomized trees.The learning training of the model was performed on the training set,and its model performance was verified with the test set.The time required in a single collection of joint friction sound from the subjects and the interpretation of data analysis was recorded.Knee function of the subjects were scored according to the Lysholm Score before and at 1 day after the test.The accuracy rates of diagnosis of meniscus injury with friction sounds under the two modes of exercise were compared based on the test results to yield an optimal one.The effectiveness of the four models was compared to find the best machine learning model fitting the data frame of this study according to the test results such as accuracy,sensitivity,specificity,F1 score,and area under the receiver operating characteristic curve(AUC)obtained with the optimal mode of exercise.The diagnostic accuracy,misdiagnosis rate and missed diagnosis rate of joint friction sound for meniscal injury under the optimal machine learning model with the optimal mode of exercise were observed.Results The time required in a single collection of joint friction sound ranged from 5 to 10 minutes[(7.1±1.3)minutes],when the time required for interpretation of data analysis was approximately 1 minute.The Lysholm Score before and after the test was(75.6±4.0)points and(77.7±3.7)points respectively in the meniscal injury group(P>0.05),and(99.6±0.9)points and(99.5±1.0)points respectively in the healthy group(P>0.05).The diagnosing accuracy rates for flexion-extension-flexion of exercise and sit-stand-sit modes of exercise were 0.775 and 0.817 under the support vector machine model with linear kernels;0.813 and 0.900 under the radial basis function support vector machine model;0.800 and 0.867 under the random forest model;0.800 and 0.900 under the extremely randomized tree model.The accuracy rates for sit-stand-sit mode of exercise were all higher than those for flexion-extension-flexion mode of exercise.In the sit-stand-sit mode of exercise,the extremely randomized tree model had an accuracy rate of 0.900,sensitivity of 0.900,specificity of 0.950,F1 score of 0.900,and AUC of 0.942,which were higher than those under the remaining 3 models,showing better machine learning efficacy.Under the extremely randomized tree model in the sit-stand-sit mode of exercise,22(18 true positive and 4 false positive)were diagnosed as meniscal injury and 38(36 true negative and 2 false negative)as healthy out of 60 segments in the test set(20 from the meniscal injury group and 40 from the healthy group).The diagnostic accuracy of joint friction sounds in diagnosing meniscus injury of the knee was 0.900,with the misdiagnosis rate of 0.100 and the missed diagnosis rate of 0.100.Conclusion Diagnosis of meniscus injury of the knee with joint friction sounds can shorten time and enhance safety during the examination process.The diagnostic model using machine learning-based artificial intelligence is faster and more stable,which can be used as a diagnostic marker for such injury.