| Objective:Parkinson’s disease(PD)is a neurodegenerative disease with slow progression and high morbidity,which not only results in great physical and mental harms to the patients,but also huge economic costs and medical burdens to their families and the society.Although the disease cannot be completely cured at present,timely and reasonable intervention and treatment can effectively improve the symptoms and prevent the progression of the disease,especially in the early stage of the disease.It is currently known that autonomic dysfunctions in PD patients are associated with waveform changes in the ECG,and such changes often occur in its early stage,even as the first symptoms.Therefore,ECG can be used as a potential method for PD screening.However,it is less accurate for doctors to recognize PD by identifying ECG changes with the naked eye.This study aimed to make full use of artificial intelligence technology to identifythe abnormal ECG changes in patients with PD,and construst a model for PD screening and auxiliary diagnosis,which provides a new direction for the development of screening and clinical diagnosis of PD.Methods:Patients diagnosed with and without PD in the inpatient department of a grade A hospital in Nanchang from January 2017 to December 2019 were selected,and the standard 12-lead ECG recorded in the resting recumbent position and their diagnostic indicators on the ECG reports were collected.Three board-certified neurologists were invited to evaluate all clinical data and divided patients into a PD group and a healthy control group.All ECGs were screened according to the exclusion criteria,including:(1)ECGs of the died were excluded(ECGs without continuous P-QRS-T wave group in each lead).(2)ECGs of patients with heart disease were excluded.(3)ECGs of patients who had received cardiac treatment were excluded.(4)ECGs of serious noise or fuzzy or limb lead interference were excluded.(5)Excluded ECGs of patients whose diagnosis was not completely consistent in 3 experts were excluded.Finally,the ECGs included in the present study were randomly divided into training set(64%),validation set(16%)and test set(20%).On the one hand,a convolutional neural network(CNN)model was established to extract the ECG image features of the research subjects,and the ECGs were randomly selected for 10-fold cross-validation.The ECGs images of the research object were trained and verified according to the above-mentioned grouped input CNN.Then,according to the verification results,the parameters of the trained CNN were optimized to obtain the optimal CNN model.Lastly,the optimized CNN was obtained by using the test set data.The CNN was aimed to make predictions,the accuracy of the CNN was tested,and the performance of the model to screen for PD by identifying ECG image features was evaluated.In contrast,5 different traditional machine learning models(logistic regression,random forest,support vector machine,k-nearest neighbor algorithm,XGBoost)were constructed by using the ECG diagnostic index values on the ECG reports of the research subjects,and 10% off cross-validation.The process was the same as above mentioned.Finally,by analyzing and comparing the experimental results of the CNN model and five machine learning models,the optimal model was selected as the plan for early screening of PD.Results:A total of 1024 ECGs were used in the present study,including 512 ECGs in the PD group and 512 in the healthy control group.Of those,80%(818)were used as the training data set,and 20%(206)were used as the test data set.The results were shown as follows:(1)After the ANOVA test,the characteristics of ECGs(P wave,QRS,Rv5,and Rv5+Sv1)were significantly different between the patients with PD whose ECGs were diagnosed as normal ECGs,the patients with PD whose ECGs were diagnosed as abnormal ECGs,and the healthy control group whose ECGs were diagnosed as normal ECGs.(2)In univariate analysis,the diagnostic indexes of ECGs(P wave,P-R interval,QRS,Rv5,and Rv5+Sv1)were all significantly correlated with the occurrence of PD in the general population.(3)The results of traditional machine learning model were shown as follows:(1)Logistic regression model,its AUC in the test set: 0.572(95%CI,0.493-0.651),the accuracy was 0.580,the sensitivity was 0.960,the specificity was 0.221,the positive predictive value was 0.542.The negative predictive value was 0.821 and the F1 value was 0.693.(2)Random forest model,its AUC in the test set: 0.683(95%CI,0.610-0.755),the accuracy was 0.624,the sensitivity was 0.941,the specificity was0.346,the positive predictive value was 0.575,the negative predictive value was0.800,and the F1 value was 0.714.(3)Support vector machine model,its AUC in the test set: 0.563(95%CI,0.483-0.644),the accuracy was 0.605,the sensitivity was0.931,the specificity was 0.298,the positive predictive value was 0.560,the negative predictive value was 0.795,and the F1 value was 0.699.(4)K-nearest neighbor algorithm model,its AUC in the test set: 0.627(95%CI,0.553-0.700),the accuracy was 0.580,the sensitivity was 0.653,the specificity was 0.519,the positive predictive value was 0.667,the negative predictive value was 0.556,the F1 value was 0.660.(5)XGBoost model,its AUC in the test set: 0.653(95%CI,0.557-0.729),the accuracy was 0.644,the sensitivity was 0.693,the specificity was 0.606,the positive predictive value was 0.627,the negative predictive value was 0.663,and the F1 value was 0.658.(4)Parkinson’s disease-convolutional neural network(PD-CNN)model,its AUC in the test set: 0.887(95% CI,0.838-0.926).When the critical value was 0.4673,the accuracy was 0.845,the sensitivity was 0.850,the specificity was 0.833,the positive predictive value was 0.838,the negative predictive value was 0.851,and the F1 value was 0.844.Conclusions:(1)Compared with the healthy control group,the characteristics of ECGs(P wave,P-R interval,Rv5+Sv1,etc.)of PD patients were significantly changed.(2)Traditional machine learning models can aslo identify the ECG of patients with PD,but the effect is not ideal.(3)Compared with the traditional machine learning models,PD-CNN model has higher accuracy and better performance in ECG identification of PD patients.Thus,PD-CNN could quickly and accurately identify ECG and effectively realize automatic diagnosis of PD without invasions. |
PDF Full Text Request