This study is a retrospective study, and all relevant data were anonymized and extracted solely from the patients’ existing medical records. Since the study did not involve direct patient contact or additional medical procedures, the study was exempted by the Ethics Committee of the First Affiliated Hospital of Wannan Medical College, and patient informed consent was not required. The entire study strictly adhered to the relevant regulations of the Declaration of Helsinki (2013 revision). All participants were from Institution 1 and Institution 2, both of which participated in the study. The study period spanned from June 2020 to August 2024, during which 2036 patients’ TTE and clinical data were collected. A total of 971 patients were selected for the final analysis, with 631 patients from Institution 1 (HCM = 210, HHD = 216, UCM = 205) and 340 patients from Institution 2 (HCM = 93, HHD = 147, UCM = 100). Patients were divided into the HCM group and the non-HCM group based on the underlying cause. Patients from Institution 1 were randomly assigned to a training set (n = 503) and an internal validation set (n = 128) in an 8:2 ratio; all patients from Institution 2 formed the external validation set (n = 340). Inclusion criteria for HCM were: (1) end-diastolic left ventricular wall thickness (LVWT) $\ge$15 mm, not due to load conditions; (2) LVWT of 13–14 mm with a family history of HCM, not due to load conditions; (3) presence of pathogenic mutations in genetic Validation (e.g., MYH7 and MYBPC3). Inclusion criteria for HHD were: (1) LVWT $>$12 mm, with a confirmed diagnosis of hypertension; (2) patients diagnosed with HHD must not have other heart diseases that could lead to myocardial hypertrophy (e.g., valvular heart disease, aortic valve stenosis, athletic heart syndrome, and other types of cardiomyopathies). Inclusion criteria for UCM were a glomerular filtration rate (eGFR) 15 mL/min/1.73 m². Exclusion criteria included: (1) myocardial amyloidosis; (2) unknown cause of myocardial hypertrophy; (3) myocardial disarray; (4) myocardial infarction; (5) unclear echocardiographic images or missing clinical data. The process of inclusion and exclusion of experimental subjects is shown in Fig. 2.

Fig. 2.

The flowchart illustrates the participant selection process and exclusion criteria. HHD, hypertensive heart disease; UCM, uremic cardiomyopathy.

This study followed the relevant guidelines published by the American Society of Echocardiography (ASE) and was conducted by two experienced echocardiologists (Echocardiologist 1 with 8 years of clinical experience and Echocardiologist 2 with 20 years of clinical experience) responsible for TTE image acquisition. Imaging was performed using Philips EPIQ 7C (Philips, Amsterdam, Netherlands) and GE Vivid E95 (GE, Boston, MA, USA) ultrasound machines, and the apical 4-chamber (A4C) view was selected for imaging. During the image acquisition process, the image quality was ensured to meet the required standards, and the relevant ultrasound parameters were accurately measured. The TTE parameters collected included left ventricular dimension at end-diastole (LVDd), left atrial dimension (LAD), left ventricular ejection fraction (LVEF), interventricular septal thickness (IVS), and left ventricular posterior wall thickness (LVPWT).

This study first standardizes the TTE images and then imports the processed A4C view images into the ITK-SNAP software (version 3.8, Philadelphia, PA, USA) for further analysis. Subsequently, Ultrasound Physician 1 uses software to delineate the left ventricular myocardium as the region of interest (ROI). Ultrasound Physician 2 reviews the ROI delineated by Ultrasound Physician 1, and in case of any discrepancies, the judgment of Ultrasound Physician 2 is considered final.

This study utilizes the open-source tool Pyradiomics package of Python Software (version 3.12, Python Software Foundation, DE, USA) for the extraction of radiomic features. A total of 1561 radiomic features were extracted through the analysis of the ROI regions, which can be categorized into three main types: geometric features, intensity features, and texture features. Geometric features describe the morphology of the delineated region, intensity features reflect the distribution of voxel intensities within the region, and texture features reveal the patterns of intensity or higher-order spatial distributions. The methods used for texture feature extraction include the Gray Level Co-occurrence Matrix (GLCM), Gray Level Run Length Matrix (GLRLM), Gray Level Size Zone Matrix (GLSZM), and Neighborhood Gray Tone Difference Matrix (NGTDM). These methods allow for the extraction of detailed texture information from the ROI regions, thereby contributing to more accurate regional descriptions and diagnoses. To evaluate the reproducibility of the radiomic feature extraction, two ultrasound physicians randomly selected ultrasound images from 50 patients, re-delineated the ROIs, and calculated the intra-class correlation coefficient (ICC). Features were considered to have high reproducibility when the ICC value was greater than 0.75.

Before performing deep learning feature extraction, a minimal bounding rectangle that encompasses the ROI region must first be cropped. For model construction, this study selected DenseNet121 as the base network for transfer learning due to its outstanding performance on the ImageNet dataset. The DenseNet121 model was first loaded, retaining the weights of the initial convolutional layers, while freezing the parameters of these layers. Next, the final fully connected layer of the model was replaced with a new layer to ensure that its output matched the number of classes in the target dataset. To further enhance the model’s diagnostic performance, cross-entropy loss was used for fine-tuning. During the optimization process, the Adam optimizer was chosen, with a batch size of 64, a learning rate of 0.001, and 50 epochs of training. After fine-tuning, the parameters of the network were frozen to ensure the stability and consistency of the feature extraction process. Finally, by extracting the output from the second-to-last layer of the model, 50,177 DTL features were obtained for each patient.

This study first performed dimensionality reduction on DTL features using Principal Component Analysis (PCA), ultimately retaining 128 features. Subsequently, the 128 DTL features were fused with 1561 radiomics features to create a combined feature dataset. Next, an independent samples t-test was conducted on all features, and significant features with p-values less than 0.05 were selected. Following this, a LASSO regression model was employed for feature selection, using the regularization parameter $\lambda$ to shrink the regression coefficients to zero, thereby eliminating many irrelevant features. To determine the optimal $\lambda$ value, 10-fold cross-validation was used to select the $\lambda$ value that minimized the validation error. Ultimately, the features with non-zero regression coefficients were retained, and LASSO modeling was performed using the scikit-learn library in Python. To address high correlation between features, Spearman’s rank correlation coefficient was used to assess relationships between features, and features with correlation coefficients exceeding 0.9 with other features were removed. To further enhance the accuracy of feature selection, a greedy recursive elimination method was employed, which removed the most redundant features in each iteration. After feature selection, 30 radiomics features and 13 DTL features were obtained. The entire feature selection process and the final selected features are shown in Fig. 3.

Fig. 3.

Feature selection process for the integrated feature dataset. (a) Feature distribution plot. (b) Coefficients from 10-fold cross-validation. (c) Mean squared error (MSE) from 10-fold cross-validation. (d) Feature weights of non-zero coefficients.

Model 1 was constructed using radiomics features extracted from the ROI region. Model 2 was built based on DTL features extracted from the ROI region. Model 3 (fusion model) combines radiomics features and DTL features. Fig. 4 illustrates the structural schematic of these three models.

Fig. 4.

Structural schematic of these three models. Model 1 (RAD): 1561 radiomics features. Model 2 (DTL+PCA): 128 deep transfer learning features. Model 3 (RAD+DTL+PCA): 1561 radiomics features and 128 deep transfer learning features. PCA, principal component analysis.

This study employs multiple evaluation metrics to assess the performance of each model, including area under the curve (AUC), receiver operating characteristic (ROC) curve, accuracy, sensitivity, specificity, decision curve analysis (DCA), predictive scores, and 95% confidence intervals (95% CI). To compare the differences in ROC curves between different models, the DeLong test was applied in this study. To select the best machine learning algorithm, the radiomics features from Model 1 were input into ten machine learning algorithms, and the optimal algorithm was determined by comparing their performance. After selecting the best algorithm, the features from Models 2 and 3 were also modeled using the same algorithm. To validate the effectiveness of Model 3, two experienced ultrasound doctors were invited to perform the evaluation (Doctor 1 with 8 years of experience and Doctor 2 with 20 years of experience). During the evaluation process, both doctors classified the patients into the HCM or non-HCM group based on the TTE images, TTE parameters, and clinical information, without knowledge of the patient’s diagnosis. Finally, the diagnostic results of Model 3 were compared and analyzed with the diagnostic results of the two ultrasound doctors.

This study performed data analysis using R software (version 4.0.2, R Foundation for Statistical Computing, Vienna, Austria) and Python. Continuous variables are expressed as mean $\pm$ standard deviation (SD). Feature selection was carried out using the $\chi$² test, and independent samples t-test, with a p-value less than 0.05 considered statistically significant. The reproducibility of the experiments was assessed using the ICC, and results were considered statistically significant only when the ICC value was greater than 0.75. ROC curves and AUC values were plotted and calculated using the scikit-learn package in Python to evaluate the diagnostic accuracy of the internal validation set and external validation set. DCA was employed to assess the clinical utility of the predictive model.

This study collected clinical data and echocardiographic parameters from 971 patients, with detailed information provided in Table 1. The patients were divided into two groups: 303 cases in the HCM group and 668 cases in the non-HCM group. Independent samples t-test, Mann-Whitney U test, and chi-square test were used to compare the clinical characteristics between the two groups. The results indicated that there were significant differences in IVS thickness, age and LVEF (p 0.05). However, no significant differences were found in LVDd, LAD, LVPWT, or gender (p $>$ 0.05).

This study utilized ten machine learning algorithms, including logistic regression (LR), SVM, Naive Bayes, K-nearest neighbors (KNN), RandomForest, LightGBM, ExtraTrees, Gradient Boosting, XGBoost, and AdaBoost, which were evaluated as shown in Table 2. Due to overfitting issues, RandomForest, ExtraTrees, and XGBoost were excluded. In the training set of Model 1, LightGBM achieved the highest AUC of 0.944, outperforming Gradient Boosting (AUC = 0.895), KNN (AUC = 0.849), and SVM (AUC = 0.836). In the internal validation set, SVM achieved the highest AUC of 0.853. In the external validation set, SVM again achieved the highest AUC of 0.772. Based on these results, to enhance the stability and reliability of the model, SVM was ultimately selected as the optimal machine learning algorithm for building Model 1. SVM was also chosen as the algorithm for constructing Models 2 and 3.

This study evaluated the performance of Models 1, 2, and 3 using AUC (95% CI), accuracy, sensitivity, and specificity (see Table 3). In the training set, the AUC values for Models 1, 2, and 3 were 0.836, 0.938, and 0.966, respectively. In the internal validation set, the AUC values for Models 1, 2, and 3 were 0.853, 0.885, and 0.945, respectively. In the external validation set, the AUC values for Models 1, 2, and 3 were 0.772, 0.856, and 0.934, respectively. In the training set, Model 3 (AUC = 0.966) performed the best, outperforming Model 1 (AUC = 0.938) and Model 2 (AUC = 0.836); in the internal validation set, Model 3 (AUC = 0.945) performed the best, outperforming Model 1 (AUC = 0.885) and Model 2 (AUC = 0.853); in the external validation set, Model 3 (AUC = 0.934) also outperformed Model 1 (AUC = 0.856) and Model 2 (AUC = 0.772). These results suggest that Model 3, which integrates radiomic features and DTL features, significantly outperforms Models 1 and 2, which rely solely on radiomic features and DTL features, respectively. The ROC curves for Models 1, 2, and 3, along with the DCA curve for Model 3, are shown in Fig. 5.

Fig. 5.

ROC curves of three models and DCA curves of Model 3. (a) The ROC curves of Model 1. (b) The ROC curves of Model 2. (c) The ROC curves of Model 3. (d) The DCA curve of Model 3 in the training set. (e) The DCA curve of Model 3 in the internal validation set. (f) The DCA curve of Model 3 in the external validation set.

We conducted SHAP analysis on Models3 to enhance the interpretability of the models. Fig. 6 presents the SHAP analysis results for Model 3.

Fig. 6.

The SHAP analysis plot of Model 3.

We compared the diagnostic performance of Model 3 with that of two ultrasound physicians (see Fig. 7). On the internal validation set, Model 3 achieved an AUC of 0.945, significantly outperforming Ultrasound Physician 1 (AUC = 0.859) and Ultrasound Physician 2 (AUC = 0.911). On the external validation set, Model 3 achieved an AUC of 0.934, also surpassing Ultrasound Physician 1 (AUC = 0.875) and Ultrasound Physician 2 (AUC = 0.908). To further compare the diagnostic capabilities of Model 3 and the ultrasound physicians, we performed the Delong test (results shown in Table 4). The DeLong test can help determine whether there is a significant difference between the diagnostic performance of the fusion model and that of ultrasound physicians when comparing the two. In this study, the results of the DeLong test indicated that the fusion model significantly outperformed the diagnostic performance of ultrasound physicians.

Fig. 7.

Comparison With Ultrasound Physicians. (a) Diagnostic performance of Model 3 and the two ultrasound physicians on the internal validation set. (b) Diagnostic performance of Model 3 and the two ultrasound physicians on the external validation set.