基于机器学习的帕金森病诊断研究

Diagnosis of Parkinson’s disease with machine learning

目的帕金森病是最常见的神经退行性疾病之一,然而目前诊断仍面临一定的挑战。机器学习可以应用于神经系统疾病的诊断与预测,然而基于大脑皮层厚度通过机器学习是否可以准确诊断帕金森病尚不清楚。方法纳入77名帕金森病患者和37名正常被试,所有被试均进行3Dmagnetization-prepared rapid acquisition gradient echo(MPRAGE)序列扫描,并提取大脑皮层所有的皮层厚度数据。将所有被试分为训练集和测试集,并进行标准化。采用Support Vector Machine(SVM)(拟合函数:Radial Basis Function)方法进行帕金森病诊断,采用Genetic Algorithm(GA)优化SVM中最佳参数。最后采用前面所确定的最优化参数对测试集进行最终预测。结果通过voxel-wise方法进行对比,并经过False discovery rate(FDR)校正后,帕金森病患者的左侧前扣带回、右侧额叶区域和左侧颞叶区域皮层厚度较正常对照组减少。进行Cross Validation最佳的平均正确率为65.4%。应用此参数创建基于SVM的机器学习网络,并对测试集进行预测,最终得到的正确率为75%,表明该模型对于帕金森病具有良好的预测性能。结论与正常对照组相比,帕金森病患者的左侧前扣带回、右侧额叶区域和左侧颞叶区域皮层厚度较正常对照组减少。通过基于全脑皮层厚度的机器学习,诊断的准确性可以达到75%。

Objective Parkinson's disease is one of the most common neurodegenerative diseases,but the current diagnosis still faces certain challenges.Machine learning can be applied to the diagnosis and prediction of neurological diseases.However,it is unclear whether it is possible to accurately diagnose Parkinson's disease by machine learning based on cerebral cortex thickness.Methods In this study,77patients with Parkinson's disease and 37normal subjects were included.All subjects underwent 3D magnetization-prepared rapid acquisition gradient echo(MPRAGE)sequence scan.And the whole brain cortical thickness was measured.All subjects were divided into training set and test set,and then the data was normalized.In this study,the Support Vector Machine(SVM)was used to diagnose Parkinson's disease,and the Genetic Algorithm(GA)was used to optimize the parameters in SVM.Finally,the prediction of the test set was performed using the optimized parameters.Results After using voxel-wise method and corrected by False discovery rate(FDR),the cortical thickness of the left anterior cingulate gyrus,right frontal lobe and left temporal lobe of Parkinson's disease patients was lower than that of the control group.The optimized average accuracy of Cross Validation was 65.4%.Applying this parameter to create SVM-based machine learning network and predicting the test set,the accuracy was 75%,indicating that the model has good diagnostic capacity for Parkinson's disease.Conclusion Compared with the control group,cortical thickness is lower in the left anterior cingulate gyrus,right frontal lobe,and left temporal lobe of Parkinson's disease patients.With machine learning based on total cortical thickness,diagnostic accuracy can reach to 75%.