A case-control study of 138 subjects, including 74 verified CA cases and 64 patients verified HCM cases were referred to the First Affiliated Hospital of Zhejiang University School of Medicine from June 2015 to November 2018. All 74 subjects with CA who had strain images were pathologically confirmed with a positive non-cardiac biopsy for amyloidosis. The characteristics of echocardiography of CA are consistent with the Expert Consensus Recommendations for Multimodality Imaging. We further incorporated 64 patients with HCM as comparator groups whose diagnose were created according to recently published guidelines from the American College of Cardiology /the American Heart Association. Echocardiographic examination was performed in all patients with HCM who presented unexplained left ventricular asymmetrical hypertrophy with septal wall thickness ≥15 mm. In the case of positive family history (such as sudden death, cardiac hypertrophy, etc.), interventricular septal thickness ≥13 mm was also enrolled. Subjects with left ventricular ejection fraction < 45%, secondary cardiac hypertrophy caused by severe aortic valve disease, long-term uncontrolled hypertension, or thyroid disease were excluded from the study. Patients were also excluded if the relevant data were not available. Local institutional ethics committee approved the study.
All echocardiographic studies were conducted on GE Vivid E9 Color Doppler Ultrasound system (GE Medical, Milwaukee, Wisconsin, USA) equipped with a 2-dimension probe M5S with a frequency of 2.0 ~ 4.5 MHz and a frame rate of 50 ~ 70 frames per second. The grayscale dynamic images of the 4-chamber views, the long axis view of left ventricle, the 2-chamber views and the short axis section with 3 consecutive cardiac cycles were obtained and stored on the hard disk. M-mode and tissue Doppler ultrasound were used to collect ultrasonic parameters, which included: the left atrial volume index using an ellipse formula, end-diastolic left ventricular diameter, end-systolic left ventricular diameter, end-diastolic left ventricular volume, end-systolic left ventricular volume, ejection fraction using the biplane Simpson’s method in 4-chamber views and 2-chamber views, septal wall thickness, posterior wall thickness. Eccentricity index was calculated as septal wall thickness divided by posterior wall thickness. Relative wall thickness was calculated as the ratio of 2septal wall thickness divided by end-diastolic left ventricular diameter. Left ventricular mass index was calculated based on Cube formula. Concentric hypertrophy was diagnosed in patients with relative wall thickness >0.42 and left ventricular mass index >115 g/m2. Diastolic parameters, including peak early (E) and late (A) diastolic mitral inflow velocity and the ratio of E/A, e’ and the ratio of E/ e' ratio were also measured.
2D-STE acquisition and analysis
Offline analysis of the video clips was based on Echo PAC Version 201 software (GE Company, Fairfield, Connecticut, USA), running on Windows 10 Version 1709 (Microsoft Corporation, Washington State, USA). Selecting clear dynamic images and using the 4-chamber views, the long axis view of left ventricle and the 2-chamber views, the left ventricular endocardial and epicardial myocardium were automatically tracked combined with manually adjusted frame by frame throughout the cardiac cycle, and divided into 16 segments to generate a ‘bull’s-eye’ plot.
The strain data are gathered by time and space parameters. Each cardiac cycle was divided into 17 equal segments (T1-T17), and Tj represented the corresponding time points (j = 1,2...17); Strain measurements were included as follows: longitudinal strain(LS)，global longitudinal strain, longitudinal strain velocity, longitudinal strain rate, longitudinal displacement, circumferential strain, global circumferential strain, circumferential strain rate, radial strain, global radial strain, radial strain rate , rotational rate, left ventricular twist, left ventricular twist rate. According to the above methods, 3791 (223×17) variables were systematically extracted for each patient (223 are strain-derived variables and 17 are time points). Average time strain-derived variables (223 variables) were used to train the models. Relative apical sparing was calculated as average apical LS divided by sum of the average basal and mid LS, septal apical to base ratio as apical septal LS divided by basal septal LS, and ejection fraction strain ratio as ejection fraction divided by global longitudinal strain.
Establishment and assessment of prediction models
Two ML-based prediction models were established respectively: one model was built using clinical characteristics, conventional echocardiography and 2D-STE data; another was constructing model using 2D-STE data only.
Prediction models of using clinical characteristics, conventional echocardiography and 2D-STE data
We developed prediction models using four approaches: logistic regression, support vector machine, random forest, and XGBoost. These represent the comprehensive analysis from traditional logistic regression to classic ML algorithms (support vector machine, random forest), and then to advanced gradient boosting (XGBoost). To assess the validity of models, we performed 10-fold (or 5-fold) cross-validation by randomly dividing the entire data into 10(or 5) parts for 10(or 5) iterations. In each iteration, we selected 7 parts as training data and 3 parts as test sets. We reported average results for each model on 30% of unseen test sets.
Logistic regression is the most commonly used risk prediction model. First, univariate logistic regression was used to screen out the variables that were meaningful to predict CA. Then, variables with P < 0.1 were enrolled in the multivariate regression analysis for modeling according to the previous research, clinical experience, and the multiple requirements between variables and outcome. In addition, Spearman correlation was used to exclude the influence of collinearity among variables.
Support vector machine
Support vector machine converts data into complex high-dimensional space to looks for the largest difference margin to realize the differentiation of diseases. We applied linear basis kernel and cost function to build model and tuning parameters to minimize the error classification.
Random forest is a tree-based method, the essence of which is to continuously split variables at discrete cutting points, usually presenting in the form of tree graph. A separate tree is built from bootstrapped of the data and variables, and the final model is a collection of many trees.
The core idea of gradient boosting is to set up a series of initial models based on decision tree, which is called base classifiers[15, 16]. Subsequently, weaker base classifiers are iterated and adjusted the weights to create a single stronger classifier. Information gain (IG), a technique of feature selection, is defined as a metric of effective classification. It is measured in terms of the entropy reduction of the class, which reflects additional information about the class provided by the variables.
Prediction models of using 2D-STE data
Boosting-based algorithms are increasingly used because they involve sequential creation of models, with each iteration attempting to correct errors in the previous models. LightGBM and XGBoost are two widely used algorithms. We developed predictive classifiers using 2D-STE data: (1) LightGBM; (2) XGBoost; (3) voting model based on LightGBM and XGBoost. To evaluate the validity of models, 5-fold cross-validation was performed. We split the dataset into training set and test set in a 4:1 ratio and reported the performance on the test data.
Categorical variables were expressed as number of cases and percentage, and were compared using the chi-square test or Fisher's test. Continuous variables were expressed as mean ± SD. Kolmogorov-Smirnov test was used to determine whether the data were normally distributed. If the data conform to the normal distribution, the independent sample T test was used for comparison; otherwise, the Mann-Whitney U test was suitable. P＜0.05 was considered statistically significant. Sensitivity, specificity, positive predictive value, negative predictive value, accuracy, receiver operating characteristic curve and area under the curve (AUC) were used to evaluate the performance of models. DeLong test was used to evaluate whether the AUC in different models was statistically significant. The data analysis was implemented on SPSS 23.0 (Version 23.0), R (Version 4.0.3) and Python (Version 3.7).