Development And Optimization Of An Electrocardiogram Intelligent Diagnosis Model For Left Ventricular Hypertrophy

Background: Left ventricular hypertrophy(LVH)is the common basis of various cardiovascular diseases,and early recognition has important clinical significance.Common methods for diagnosing LVH include cardiac magnetic resonance imaging and echocardiography.However,these two tests require high equipment and operating techniques and are not suitable for large-scale screening of subclinical symptoms.Electrocardiogram(ECG)is also commonly used to diagnose LVH,with high popularity and simple operation.However,this method has low sensitivity and great difficulty in interpretation and is generally only used for auxiliary diagnosis.In recent years,researchers have applied artificial intelligence technology to ECG examination and developed intelligent diagnostic models with good performance,which can quickly and accurately interpret the ECG changes of common diseases.Some studies have also tried to use these related techniques to improve the diagnostic ability of ECG for LVH and have made initial progress,but the overall effect could be better.Objective: This study aims to improve the recognition ability of the ECG intelligent diagnosis model for LVH and to construct an ECG intelligent diagnosis model for LVH that can achieve clinical application performance.Methods: This study retrospectively collected the relevant information of patients who underwent resting position ECG and echocardiography examination in the First Affiliated Hospital of Henan University from January 1,2021,to June 30,2022.Firstly,a residual neural network-based ECG pre-training model(ECG-Res Net)was constructed using the PTB-XL public database to distinguish normal and abnormal ECG signals.Then,transfer learning was used to construct an intelligent diagnosis model of LVH based on its dataset to evaluate the diagnostic performance of ECG in LVH.Finally,univariate and multivariate Logistic regression was used to evaluate the correlation between the patient’s clinical data and LVH,and feature screening was performed based on clinical experience.The selected clinical features and ECG scores were combined to construct a combined model to improve the diagnostic performance of the model further.Results: A total of 12800 ECGs and 21434 echocardiographic results were collected.4894 patients were matched to both ECGs and echocardiography,and 1015 patients were finally included.Firstly,the ECGRes Net model is used to pre-train the PTB-XL dataset,and the diagnostic performance of this model is compared with support vector machine,random forest,artificial neural network,and Alex Net neural network.The performance evaluation results of the pre-trained ECG intelligent diagnosis model show that: the ECGRes Net model had the most vital ability to identify abnormal ECG,with a sensitivity,specificity,accuracy,and F1 score of 0.92,0.90,0.91,and 0.91,respectively,and an AUC of 0.97.Subsequently,the migrated ECG-Res Net model was used to train the LVH electrocardiograms to be accurately labeled by echocardiography.The model’s sensitivity,specificity,accuracy,and F1 score on the test set were 0.83,0.96,0.89,and 0.93,respectively,and the AUC was 0.92.Then,univariate Logistic regression analysis,multivariate Logistic regression analysis,and clinical experience were used to screen out 8 clinical characteristics: age,gender,BMI,heart rate,dyslipidemia,history of hypertension,history of diabetes,and history of chronic kidney disease.Three joint models were constructed based on the feature selection results.The first joint model included features(Feature1),including ECG score and primary patient data(age,gender,BMI,heart rate).The features included in the second combined model(Feature2)were ECGscore and patient history(dyslipidemia,hypertension,diabetes,chronic kidney disease),and the features included in the third combined model(Feature3)were ECGscore,patient demographics,and patient history.The performance of five machine learning algorithms,including Logistic regression,gradient boosting decision tree,random forest,limit tree,and XGBoost,were compared.The results of the three combined models showed that the XGBoost algorithm had the best performance,and the diagnostic performance was significantly improved when the ECG score was combined with the clinical data of patients.Only adding the primary data of patients,such as age,gender,BMI,and heart rate,into the combined model could make the model achieve optimal performance.The sensitivity,specificity,accuracy,F1 score,and AUC of the XGBoost algorithm based on Feature 1 were 0.98,0.98,0.98,0.97,and 0.99,respectively.Finally,the feature contribution analysis found that ECGscore had the most significant impact on the prediction results of the model,and the SHAP value was 4.91.With the increased ECG score,the probability of the sample being predicted as left ventricular hypertrophy increased.Conclusions: The application of artificial intelligence in ECG can be used as a new diagnostic tool for the detection of left ventricular hypertrophy.The sensitivity,specificity,F1 score,accuracy,and AUC were0.83,0.96,0.89,0.93,and 0.92,respectively.The combined features of ECG results could further improve the diagnostic performance of the model,and the sensitivity,specificity,F1 score,accuracy,and AUC were0.98,0.98,0.98,0.97,and 0.99,respectively.The performance has met the requirements for clinical application.