Prognosis prediction in type B aortic intramural hematoma patients using a combined model based on aortic computed tomography angiography radiomics

Introduction

Aortic intramural hematoma (IMH) is a life-threatening aortic disease caused by the rupture of vasa vasorum and the formation of a hematoma in the middle layer of the aortic wall, or small intimal tears that cannot be visualized on standard imaging examinations (1). Due to its expedient scanning time, simple accessibility, and superb sensitivity, specificity, and negative predictive value, computed tomography angiography (CTA) is usually the preferred method for identifying IMH (2). The main manifestation of IMH is an eccentric or concentric thickening of the aortic wall >5 mm, extending longitudinally. Along with aortic dissection (AD), aortic IMH affecting the descending aorta distal to the left subclavian artery is classified as Stanford type B, and type B aortic IMH accounts for 70% of all cases of aortic IMH (3).

The course of type B aortic IMH is unstable, and different studies have shown that the evolution is highly heterogeneous. Some studies have reported that 4–37% of type B aortic IMH can progress to AD, and 20% can develop progressive aortic dilation, while other studies have reported that the regression rate of type B aortic IMH is 10–78%, and the complete absorption rate of hematoma is up to 55% (4-6). Possible risk factors for the progression of type B aortic IMH include recurrent or persistent chest/back pain, the presence of a penetrating ulcer (PAU) or ulcer-like protrusion (ULP) at the affected segment, the involvement of the ascending aorta, the ischemia of related organs, an aortic diameter >50 mm, progressive aortic diameter enlargement, maximum aortic wall thickness >11 mm, recurrent pleural effusion, and difficulty in blood pressure (BP) control (7-10). However, reports of the above indicators have varied greatly in different studies, and it is challenging to make precise prognostications regarding the individual outcomes of patients with type B aortic IMH. Therefore, there is a critical need for accurate and personalized prognostic tools that can assist in making treatment decisions.

Radiomics combines image information, statistics, artificial intelligence, machine learning, and deep learning to transform traditional visual image data into deep features for quantitative studies to create diagnostic, prognostic, and predictive models (11). By leveraging these advanced methodologies, radiomics enables the extraction of high-dimensional data that can be pivotal in identifying subtle patterns and correlations not apparent through conventional analysis. Cardiovascular diseases, such as high-risk coronary atherosclerosis plaque, aortic valve stenosis, aortic aneurysm, and AD, have widely utilized radiomics for diagnosis and prognosis analysis (12-16). However, radiomics models for predicting the risk of type B aortic IMH progression have not yet been developed.

Therefore, this study aimed to explore the value of radiomics in predicting the prognosis of type B aortic IMH, and to construct a prediction model of IMH progression by analyzing the relationship between high-dimensional data and quantitative information, and by integrating clinical and CTA imaging risk factors. This study utilized a retrospective review of patient data, along with sophisticated radiomic feature extraction and machine learning approaches, to create a reliable predictive model. We present this article in accordance with the TRIPOD reporting checklist (available at https://qims.amegroups.com/article/view/10.21037/qims-24-1914/rc).

Methods

Patients

This study was approved by the Institutional Review Board of Tongji Hospital Affiliated to Huazhong University of Science and Technology. The requirement for written informed consent was waived due to the retrospective nature of this study. The study was conducted in accordance with the Declaration of Helsinki (as revised in 2013). We retrospectively included consecutive patients diagnosed with type B aortic IMH who underwent aortic CTA examination, without surgical or thoracic endovascular aortic repair (TEVAR) treatment, and aortic CTA re-examination between January 2014 and December 2023 at Tongji Hospital Affiliated to Huazhong University of Science and Technology. The patients were randomly assigned to the training and validation cohorts at a ratio of 4:1. Figure 1 shows the patient recruitment pathway.

Figure 1 Flowchart depicting the enrollment of the patients and model building. IMH, intramural hematoma; CTA, computed tomography angiography.

To be eligible for inclusion in this study, the patients had to meet the following inclusion criteria: have had their data collected less than 7 days from the onset; have undergone aortic CTA during the acute stage; and have not undergone aortic surgery or TEVAR before the aortic CTA re-examination. Patients were excluded from the study if they met any of the following exclusion criteria: were aged <18 or >90 years; were a pregnant female; were poor at breath holding; were unable to lift both upper limbs; were unconscious; had severe renal impairment, had an estimated glomerular filtration rate <30 mmol/L; had severe liver damage; had severe heart failure or an ejection fraction <30%; had IMH involving the ascending aorta; had a history of aortic stent implantation or aortic artificial vessel replacement; did not wish to participate in the study or did not wish to undergo follow up; had image artifacts affecting an aortic diagnosis; and/or had incomplete baseline clinical data.

The diagnostic criteria for type B aortic IMH included aortic wall thickening >5 mm, a round or crescent shape, no contrast-enhanced false lumen, and the hematoma’s confinement to the distal left subclavian artery. Prognosis was classified as follows: (I) progression, referring to deterioration in the aortic condition, including significant increases in the thickness of the IMH, progression of the IMH to a PAU/ULP, progression to AD, aortic dilation, or rupture; or (II) stabilization, referring to an unchanged appearance or decrease in the size or disappearance of the IMH.

The demographic factors and possible risk factors for the progression of type B aortic IMH described in previous studies (2,4,6) were recorded in the present study. All the clinical endpoints can be found in Table 1, which were all obtained from the medical records.

Computed tomography (CT) image acquisition

An electrocardiography-gated contrast-enhanced multidetector CT scan was performed with a GE light speed VCT 64 scanner (Revolution; GE Healthcare, Milwaukee, WI, USA), a 320-slice CT scanner (Aquilion One; Toshiba Medical Systems, Tokyo, Japan), and an ultrasonic CT 780 scanner (United Imaging, United Imaging Medical Technology Co., LTD., Shanghai, China). The patient was positioned supine with feet advanced and both upper limbs raised. The scan range was from the thoracic entrance to the lower margin plane of the ischial tubercle. First, the localization image was scanned, followed by the whole neck, chest, and abdominal arteries. Next, 60–80 mL of iodobitol was injected into the right elbow vein at a speed of 3–5 mL/s with a high-pressure syringe, and 30 mL of sodium chloride was injected at the same speed. The scanning system was triggered by mass injection. When the density of the area of interest (located at the aortic root) reached the preset value (120–150 HU), the scan automatically began. The scanning parameters were as follows: tube voltage: 120 kVp; tube current: intelligent mA technology (150–750 mA); tube speed: 0.6 s/r; pitch: 0.984; and scanning layer thickness: 1.25 mm. After the scanning was completed, the automatic reconstruction data were transmitted to the workstation. The diagnosis was based on the axial image, which was reconstructed using multi-plane reconstruction, maximum density projection, and volume reconstruction. The data were stored in DICOM 3.0 standard format.

The mean CT value of the hematoma, maximum CT value of the hematoma, maximum thickness of the hematoma (MHT), involvement range of the hematoma, maximum aortic diameter (MAD), involvement of branch vessels, ULP location, diameter, ULP depth, accompanying arterial dilation, accompanying dissection, aortic atherosclerosis, pleural effusion, and pericardial effusion were respectively recorded by two cardiovascular radiologists with more than 5 years of experience each. When the two radiologists disagreed, a third diagnostician (with 25 years of experience) was consulted to resolve the disagreement between the two radiologists.

Radiomics analysis

The region of interest (ROI), which was defined as the hematoma area, underwent manual segmentation on the CTA image. This process involved carefully outlining and selecting the specific blood clot region by hand, ensuring the accurate delineation for the next-step analysis. The image segmentation of the ROI was performed slice by slice using the open-source ITK-SNAP 4.0 software by the same two radiologists. If the two annotators could not reach a consensus in a case, the same third diagnostician was consulted to resolve the issue. This ensured the segmentation process employed a standardized and expert-driven approach, minimizing inter-observer variability and enhancing the reliability of the delineated ROIs.

Radiomic features were extracted from the ROIs using the uAI Research Portal (URP). In total, 1004 features were extracted. The radiomic features extracted included intensity features (e.g., first-order features, such as energy, entropy, kurtosis, skewness, maximum, mean, median, minimum, standard deviation, and uniformity), shape features (e.g., volume, surface area, sphericity asymmetry, sphericity, and surface volume ratio), and texture and gradient features derived from the following four categories: 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). Additionally, the features included those extracted after the filtering of the images, such as wavelet transform features. This comprehensive feature set captured a wide range of information from the images, enabling the detailed characterization of the tissue properties in the ROI.

To mitigate the effects of the dimensional disparities among indices and ensure comparability across different sequences, the extracted radiomic features were normalized using z-score standardization. This process transformed the feature distributions into a common scale, approximating a standard normal distribution, thereby facilitating a consistent and unbiased comparative analysis. The least absolute shrinkage and selection operator (LASSO) logistic regression (LR) methodology was used for the feature selection. This approach aimed to identify and select radiomic features with the highest degree of relevance to the primary objective (IMH stabilization versus progression) from the previously extracted feature set. By applying LASSO, the algorithm performed feature screening by effectively penalizing and reducing the coefficients of less contributory features to zero, thereby retaining only those features that were most strongly associated with the differentiation between stable and progressive IMH states. This resulted in a streamlined model with enhanced predictive accuracy and interpretability.

We constructed five predictive models using the following distinct combinations of feature sets: clinical features, CTA imaging features, and radiomic features. Specifically, the models were built using radiomic features alone, clinical features alone, CTA imaging features alone, a combination of clinical and CTA imaging features, and finally, a comprehensive model incorporating all three types of features simultaneously. In each model, LASSO was employed for the feature selection, ensuring that only the most relevant features were retained for the model construction and prediction.

Various machine-learning algorithms, including random forest (RF), support vector machine (SVM), LR, K-nearest neighbor (KNN), decision tree, and stochastic gradient descent (SGD) algorithms, were employed for the model training. To evaluate the performance of the models and to comprehensively assess their predictive capabilities, metrics, such as the area under the curve (AUC), sensitivity, specificity, accuracy, and F1 score, were used. This approach was used to develop robust models capable of accurately distinguishing between stable and progressing cases of type B aortic IMH, leveraging the synergistic power of multi-type features.

Statistical analysis

The statistical analyses were carried out using IBM SPSS version 25 (SPSS Inc., USA). The Shapiro-Wilk test was used to assess the normality of the quantitative data. The quantitative variables are expressed as the mean ± standard deviation if normally distributed, and the median and inter-quartile range if non-normally distributed. The categorical data are presented as the frequency and percentage. The variables were compared using the Student’s t-test, an analysis of variance, the Mann-Whitney U test, the Kruskal-Wallis H test, or the Chi-squared test as appropriate. The receiver operating characteristic (ROC) curves were generated, and the corresponding AUCs were obtained. All the statistical tests were two-tailed, and a P value <0.05 was considered statistically significant.

Results

Clinical characteristics

A total of 119 patients (92 males, 27 females; mean age: 58.78±11.54 years; range, 33–86 years) were enrolled in the study, of whom 61 had unstable and 58 had stable type B aortic IMH. Figures 2,3 show the progression and regression of type B intramural aortic hematoma, respectively. All the patients were randomly assigned to the training cohort (n=95) or validation cohort (n=24). Table 1 shows the clinical characteristics of the patients in the training and validation cohorts. There were no significant differences in the clinical features between the training and validation groups, except for smoking and alcohol history.

Figure 2 An IMH progressing to AD was observed on CTA in a patient with type B intramural hematoma. (A,D) Axial and sagittal views of a CTA demonstrating a type B IMH on admission. (B,E) Axial and sagittal views demonstrating progression of a type B IMH 10 days post initial presentation despite optimal pulse-pressure control therapy. (C,F) 3D reconstruction of the CTA depicturing the progression of a type B IMH. IMH, intramural hematoma; AD, aortic dissection; 3D, three-dimensional; CTA, computed tomography angiography.

Figure 3 Absorption of the IMH was observed on CTA in a patient with type B IMH. (A,B) Axial and sagittal views of a CTA demonstrating a type B IMH on admission. (C,D) Axial and sagittal views demonstrating the complete resolution of the type B IMH 5 months later, after receiving optimal pulse-pressure control therapy. IMH, intramural hematoma; CTA, computed tomography angiography.

Feature selection and prediction model construction

From an initial pool of 1,004 radiomic features, we performed feature selection and selected 12 radiomic features (Figure 4). These 12 features were subsequently used to establish a machine-learning classification task based on radiomic characteristics, and to construct a classification model in conjunction with clinical features and imaging features. Similarly, LASSO regression was performed, and the following 12 clinical features were selected from 21 clinical features: age, sex, chest distress, alcohol consumption, history of hypertension, heart rate (HR), BP, white blood cell (WBC) count, low density lipoprotein (LDL), aspartate aminotransferase (AST), history of smoking, and hemoglobin (Hb) (Figure 5). Next, the following 11 features were selected from the 15 CTA imaging features using the same method: involvement range of hematoma, mean CT value of hematoma, MAD, MHT, ULP depth, ULP location, branch vessel involvement, dissection, pleural effusion, aortic dilation, and atherosclerosis (Figure 6).

Figure 4 Coefficients and ROC of radiomic model in training and test cohort. (A) Diagram of selected radiomic features, indicating the coefficients of each feature in the LASSO regression. (B) Radiomic feature selection using the LASSO regression model: coefficient changes of the features under different α parameters. (C) ROC curves of the radiomics model with different machine-learning algorithms in the training cohort. (D) ROC curves of the radiomics model with different machine-learning algorithms in the test cohort. ROC, receiver operating characteristic; LR, logistic regression; AUC, area under the curve; SVM, support vector machine; KNN, K-nearest neighbor; SGD, stochastic gradient descent; LASSO, least absolute shrinkage and selection operator.

Figure 5 Coefficients and ROC of clinical model in training and test cohort. (A) Diagram of selected clinical features, indicating the coefficients of each feature in the LASSO regression. (B) Clinical feature selection using the LASSO regression model: coefficient changes of the features under different α parameters. (C) ROC curves of the clinical model with different machine-learning algorithms in the training cohort. (D) ROC curves of the clinical model with different machine-learning algorithms in the test cohort. ROC, receiver operating characteristic; WBC, white blood cell; RBC, red blood cell; Hb, hemoglobin; LDL, low density lipoprotein; AST, alanine transaminase; LR, logistic regression; AUC, area under the curve; SVM, support vector machine; KNN, K-nearest neighbor; SGD, stochastic gradient descent; LASSO, least absolute shrinkage and selection operator.

Figure 6 Coefficients and ROC of CTA model in training and test cohort. (A) Diagram of selected CTA imaging features, indicating the coefficients of each feature in the LASSO regression. (B) CTA imaging feature selection using the LASSO regression model: coefficient changes of the features under different α parameters. (C) ROC curves of the CTA imaging model with different machine-learning algorithms in the training cohort. (D) ROC curves of the CTA imaging model with different machine-learning algorithms in the test cohort. ROC, receiver operating characteristic; PAU, penetrating atherosclerotic ulcer; CT, computed tomography; CTA, computed tomography angiography; ULP, ulcer-like protrusion; LR, logistic regression; AUC, area under the curve; SVM, support vector machine; KNN, K-nearest neighbor; SGD, stochastic gradient descent; LASSO, least absolute shrinkage and selection operator.

We constructed five predictive models using distinct combinations of the clinical features, CTA imaging features, and radiomic features. Specifically, models were built using radiomic features alone, clinical features alone, CTA imaging features alone, a combination of clinical and CTA imaging features, and finally, a comprehensive model incorporating all three types of features. When constructing the combination model of clinical and CTA imaging features (i.e., the clinical + CTA imaging model), the following 12 features were selected by LASSO regression: WBC count, Hb, LDL, AST, alcohol intake, chest distress, mean CT value of hematoma, MAD, involvement range of hematoma, ULP location, associated dissection, and pleural effusion (Figure 7). The following 16 features were selected by LASSO regression to construct the comprehensive model that incorporated all three types of features (i.e., the radiomics + clinical + CTA model): Wavelet_glszm_waavelet HLL–SmallAreaEmphasis, original_glrlm_GrayLevelNon UniformityNormalized, original_shape_LeastAxisLen GTH, laplaciansharpening_glszm_GrayLevelNonUniformity, wavelet_glszm_wavelet HLL–SmallAreaHighGrayLevelEmphasis, original_glszm_SizeZoneNonUniformityNormalized, normalize_glszm_SizeZoneNonUniformityNormalized, wavelet_firstorder_wavelet-LHL–Kurt Osis, wavelet_gldm_wavelet HHL-SmallDependenceLowGrayLevel Emphasis, wavelet_ngtdm_wavelet-LLL–Contrast, wavelet_firstorder_wavelet-HHL-median, chest distress, WBC count, LDL, PAU location and associated dissection (Figure 8).

Figure 7 Coefficients and ROC of clinical + CTA model in training and test cohort. (A) Diagram of selected clinical and CTA imaging features, indicating the coefficients of each feature in the LASSO regression. (B) Clinical and CTA imaging feature selection using the LASSO regression model: coefficient changes of the features under different α parameters. (C) ROC curves of the clinical and CTA imaging model with different machine-learning algorithms in the training cohort. (D) ROC curves of the clinical and CTA imaging model with different machine-learning algorithms in the test cohort. ROC, receiver operating characteristic; CTA, computed tomography angiography; WBC, white blood cell; RBC, red blood cell; Hb, hemoglobin; LDL, low density lipoprotein; CT, computed tomography; AST, aspartate aminotransferase; ULP, ulcer-like protrusion; AUC, area under the curve; LR, logistic regression; SVM, support vector machine; KNN, K-nearest neighbor; SGD, stochastic gradient descent; LASSO, least absolute shrinkage and selection operator.

Figure 8 Coefficients and ROC of radiomic+ CTA + clinical model in training and test cohort. (A) Diagram of selected clinical features, CTA imaging features, and radiomic features, indicating the coefficients of each feature in the LASSO regression. (B) Clinical, CTA imaging, and radiomic features selection using the LASSO regression model: coefficient changes of the features under different α parameters. (C) ROC curves of the comprehensive model incorporating all three types of features with different machine-learning algorithms in the training cohort. (D) ROC curves of the comprehensive model with different machine-learning algorithms in the test cohort. ROC, receiver operating characteristic; CTA, computed tomography angiography; LASSO, least absolute shrinkage and selection operator; Hb, hemoglobin; PLT, platelet; TC, total cholesterol; TG, total triglycerides; LDL, low density lipoprotein; ALT, alanine transaminase; AST, aspartate aminotransferase; Cr, creatinine; WBC, white blood cell; LDL, low density lipoprotein; LR, logistic regression; AUC, area under the curve; SVM, support vector machine; KNN, K-nearest neighbor; SGD, stochastic gradient descent.

As Figures 4-8 show, among the five models, the KNN algorithm outperformed the others, achieving the highest AUC (1.00) in the training set. For the test set, the AUC of the SVM algorithm in the radiomics model was the highest (0.833), while the AUC of the KNN algorithm in the clinical model was the highest (0.701), the AUC of the RF algorithm in the CTA imaging model was the highest (0.806), and the AUC of the LR and SGD algorithms in the clinical + CTA imaging model were the highest (both 0.792). In the clinical + CTA + radiomics model, the KNN algorithm had the highest AUC (0.917).

Table 2 sets out the algorithm with the highest test set AUC values among the five models, and provides the corresponding sensitivity, specificity, accuracy, precision and F1 score. Figure 9 shows the ROC curves and AUC values of the aforementioned algorithms for the five models. The combined radiomics + clinical + CTA model had the highest AUC value (0.917), which was higher than that of the single radiomics model (0.833), CTA model (0.806), clinical + CTA model (0.792), and clinical model (0.701). The sensitivity, specificity, accuracy, accuracy and F1 scores of the combined radiomics + clinical + CTA model were all ≥0.75.

Figure 9 ROC curves and corresponding AUCs of the five models, among which the combined model incorporating radiomics, clinical, and CTA features had the highest AUC value. AUC, area under the curve; ROC, receiver operating characteristic; CTA, computed tomography angiography.

Discussion

In this study, a prognostic model for type B aortic IMH was established by incorporating clinical, CTA imaging, and radiomic features. We built five models using radiomic features alone, clinical features alone, CTA imaging features alone, clinical + CTA image features, and radiomic + clinical + CTA image features. We found that the radiomics + clinical + CTA imaging model outperformed all other models, achieving an AUC value of 0.917 in the test set. This study is the first to construct a combined model using radiomic features, clinical factors, and CTA features to predict the progression of type B aortic IMH.

Type B aortic IMH’s natural course is unpredictable. It can increase or decrease in size, and in some cases complete resolution is observed (17). In some patients, type B aortic IMH may progress to aortic dilation and aneurysm formation (18,19). It is also not rare that type B aortic IMH develops into AD during the follow-up period (20). Therefore, the treatment of type B aortic IMH varies according to different guidelines; for example, guidelines in Asia and the Americas recommend medical therapy, while the guidelines of the European Society of Cardiology recommend TEVAR (IIaC) (1,21).

Numerous clinical and imaging features have been used to predict the prognosis of type B aortic IMH. Clinical predictors of complications in patients with type B IMH include persistent and recurrent pain, and difficult-to-control hypertension (22). Mussa et al. (23) reported that uncontrollable pain was a significant indicator of IMH progression. Meng et al. (10) showed that an abnormal D-dimer level was a powerful independent risk factor for predicting aortic-related adverse events in type B aortic IMH, and an observation endpoint BP of 100–120 mmHg was a protective factor for predicting aortic-related adverse events in type B IMH.

Current clinical practice guidelines list a series of biomarkers that should be measured in all patients admitted with chest pain for differential diagnosis or the detection of complications (21,24). Among them, the WBC count can be used to help identify the presence of a concomitant infection and an inflammatory response. High levels of alanine transaminase (ALT) and AST may indicate the presence of liver ischemia, and an elevated creatinine level may indicate renal insufficiency. These biomarkers can serve as useful prognostic markers and identify potential complications in patients with IMH.

In the present study, clinical features, including the WBC count, AST level, LDL level, Hb level, age, gender, HR, smoking, drinking, hypertension, and chest distress, were screened by LASSO regression and used to construct a pure clinical feature model. Our findings suggest that these clinical features may be related to the prognosis of IMH, but the predictive power of the pure clinical model was lower than that of the other models (AUC =0.701).

CT scans can be used to identify some features of the type B aortic IMH that are associated with a worse prognosis. A MHT >10–15 mm, a MAD >40–55 mm, and an increasing aortic diameter (>5 mm/year) in repeated imaging tests, pleural or pericardial effusion, and organ ischemic signs have all been linked to a higher risk of complications and a worse prognosis; however, the precise cut-off values for these features remain in dispute (25,26). Additionally, the identification of focal intimal disruptions or a ULP is associated with aortic rupture and often indicates the need for aggressive endovascular therapy, but the magnitude of increased risk for each lesion remains controversial (22,27). A ULP >10 mm in depth is associated with a greater risk of complications (28). A systematic review showed that MAD cut-off values ranging from 38–44.75 mm and a ULP of the aortic wall were associated with aortic-related adverse events in IMH (22). In addition, the dissection of certain aortic segments often indicates a greater risk of aortic rupture, and invasive treatments must be considered (29).

Similar to previous studies (2,4,6), we screened imaging features such as the MAD, MHT, range of hematoma involvement, mean CT value of hematoma, pleural effusion, accompanying aortic dilation/dissection with atherosclerosis, ULP location and depth, and branch vessel involvement, and used these features to construct a CTA only imaging model with an AUC of 0.806. When the clinical + CTA imaging model was later constructed, the clinical and imaging features were further screened, and the final clinical features included the WBC count, LDL, AST, Hb, alcohol consumption, and chest distress symptoms, while the imaging features included the MAD, the mean CT value of hematoma, the range of hematoma involvement, pleural effusion, ULP location and pleural effusion. However, the combined clinical and CTA model did not have better diagnostic efficiency than the CTA only model.

Radiomics is an emerging area in quantitative image analysis that aims to relate the large-scale data mining of images to clinical endpoints. Radiomic features can provide rich information about intensity, shape, size or volume, and texture using different imaging modalities, including CT. It has been suggested that the rupture of vasa vasorum is closely related to atherosclerosis, oxidative damage, and inflammation, and radiomics can be used to identify atherosclerotic plaques and perivascular inflammation (30,31). In the present study, the radiomic features extracted included intensity features, shape features, texture features, and gradient features. Of the 12 selected radiomic features, seven were wavelet transform features, four were texture features, and one was a histogram feature. The use of a variety of different types of features enables the deep characteristics of the lesions to be more comprehensively and accurately described.

Machine learning, a subfield of artificial intelligence, has undergone rapid development in the last two decades. Classification algorithms include LR, SVM, RF, KNN, and neural network algorithms. SVM aims to find an optimal hyperplane that separates the classes as much as possible with an acceptable number of misclassified cases (32). RF is one of many ensemble methods that constructs a group of classifiers and then sorts previously unseen data by voting on predictions made by the set of weak learners (33). KNN is the most straightforward machine-learning algorithm used for both classification and regression tasks (34).

RF, SVM, LR, KNN, decision tree, and SGD algorithms were employed for model training in the current study. For different models, different algorithms can obtain the highest AUC value. The SVM algorithm was used in the radiomics model, the RF algorithm was used in the CTA imaging model, the SGD algorithm was used in the clinical + CTA imaging model, and the KNN algorithm was used in the clinical model and clinical + CTA image + radiomics model to obtain the maximum AUC value for each model, respectively. Compared with the radiomics only model, the AUC value of the combined clinical + CTA imaging + radiomics model was significantly increased (0.833 vs. 0.917), and was significantly higher than that of the other three models. Moreover, the sensitivity, specificity, accuracy, precision and F1 score of the combined clinical + CTA imaging + radiomics model were all ≥0.75.

However, there are a few caveats when conducting radiomic studies of aortic CTA, including the need for high-quality, artifact-free images of aortic CTA, good reconstruction techniques, correction methods that eliminate the influence of different CT scanning machines and parameters, and independent training and validation sets (35).

This study had several limitations. First, it was a single-center retrospective study with a relatively small sample size, which might have led to selection bias. Second, there was no external verification in this study. Third, because this study examined the patients over an extended period, the CT equipment used varied greatly. Future research should conduct a large-sample multi-center study to establish a more stable and reliable model for verification.

In conclusion, this study constructed a variety of models to quantitatively predict the prognosis of patients with type B aortic IMH. The results showed that the comprehensive model incorporating clinical, CTA imaging, and radiomic features performed best and could accurately predict the progression of type B aortic IMH. This predictive model could assist clinicians to accurately predict the evolution of type B aortic IMH patients, thereby selecting the most appropriate treatment strategy.

Acknowledgments

None.

Footnote

Reporting Checklist: The authors have completed the TRIPOD reporting checklist. Available at https://qims.amegroups.com/article/view/10.21037/qims-24-1914/rc

Funding: This study was supported by the Huazhong University of Science and Technology Affiliated Tongji Hospital Research Fund (No. 2023B23).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://qims.amegroups.com/article/view/10.21037/qims-24-1914/coif). T.L. is from United Imaging Medical Technology Co. LTD., Shanghai, China. The other authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. The study was conducted in accordance with the Declaration of Helsinki (as revised in 2013). This study was approved by the Institutional Review Board of Tongji Hospital Affiliated to Huazhong University of Science and Technology. The requirement for written informed consent was waived due to the retrospective nature of this study.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Isselbacher EM, Preventza O, Hamilton Black J 3rd, Augoustides JG, Beck AW, et al. 2022 ACC/AHA guideline for the diagnosis and management of aortic disease: A report of the American Heart Association/American College of Cardiology Joint Committee on Clinical Practice Guidelines. J Thorac Cardiovasc Surg 2023;166:e182-331. [Crossref] [PubMed]
2. Kanagami T, Saito Y, Hashimoto O, Nakayama T, Okino S, Sakai Y, Nakamura Y, Fukuzawa S, Himi T, Kobayashi Y. Clinical characteristics and outcomes in patients with acute type A aortic intramural hematoma. Int J Cardiol 2023;391:131355. [Crossref] [PubMed]
3. Sá MP, Jacquemyn X, Tasoudis P, Dufendach K, Singh MJ, de la Cruz KI, Serna-Gallegos D, Sultan I. Five Year Results of Endovascular versus Medical Therapy in Acute Type B Aortic Intramural Haematoma: Meta-Analysis of Reconstructed Time to Event Data. Eur J Vasc Endovasc Surg 2024;67:584-92. [Crossref] [PubMed]
4. Raptis CA, Braverman AC. Intramural Hematoma and Focal Intimal Disruption: The Importance of Communication. Radiology 2021;301:320-1. [Crossref] [PubMed]
5. Murana G, Gliozzi G, Rucci P, Votano D, Orioli V, Rosa S, Folesani G, Buia F, Lovato L, Pacini D. Survival and reoperation in acute aortic syndromes-a single-centre experience of 912 patients. Eur J Cardiothorac Surg 2023;64:ezad350. [Crossref] [PubMed]
6. Vacirca A, Dias Neto M, Baghbani-Oskouei A, Huang Y, Tenorio ER, Estrera A, Oderich GS. Timing of Intervention for Aortic Intramural Hematoma. Ann Vasc Surg 2023;94:14-21. [Crossref] [PubMed]
7. Jiang X, Pan T, Liu Y, Chen B, Jiang J, Guo D, Xu X, Hou K, Lin J, Ju S, Fu W, Dong Z. Prognostic Implications of Initial Focal Contrast Enhancement in Acute Type B Intramural Hematoma. J Am Coll Cardiol 2024;83:503-13. [Crossref] [PubMed]
8. Miyoshi Y, Kaji S, Masumoto A, Kim K, Kitai T, Kinoshita M, Furukawa Y. Aortic enlargement in two weeks is associated with subsequent aortic events in patients with type B acute aortic syndrome. J Thorac Cardiovasc Surg 2023;166:410-418.e1. [Crossref] [PubMed]
9. Piazza M, Squizzato F, Porcellato L, Casali E, Grego F, Antonello M. Predictors of Intervention in Acute Type B Aortic Penetrating Ulcer and Intramural Hematoma. Semin Thorac Cardiovasc Surg 2024;36:1-10. [Crossref] [PubMed]
10. Meng D, Gu R, Wang Y, Zhang Z, Xu T, Wang X. Analysis of Short-Term Clinical Outcomes and Influencing Factors in Patients with Acute Type B Aortic Intramural Hematoma Treated with Optimal Medical Therapy. Int Heart J 2023;64:708-16. [Crossref] [PubMed]
11. Chen M, Copley SJ, Viola P, Lu H, Aboagye EO. Radiomics and artificial intelligence for precision medicine in lung cancer treatment. Semin Cancer Biol 2023;93:97-113. [Crossref] [PubMed]
12. Wang Y, Xiong F, Leach J, Kao E, Tian B, Zhu C, Zhang Y, Hope M, Saloner D, Mitsouras D. Contrast-enhanced CT radiomics improves the prediction of abdominal aortic aneurysm progression. Eur Radiol 2023;33:3444-54. [Crossref] [PubMed]
13. Polidori T, De Santis D, Rucci C, Tremamunno G, Piccinni G, Pugliese L, Zerunian M, Guido G, Pucciarelli F, Bracci B, Polici M, Laghi A, Caruso D. Radiomics applications in cardiac imaging: a comprehensive review. Radiol Med 2023;128:922-33. [Crossref] [PubMed]
14. Chen Q, Pan T, Wang YN, Schoepf UJ, Bidwell SL, Qiao H, Feng Y, Xu C, Xu H, Xie G, Gao X, Tao XW, Lu M, Xu PP, Zhong J, Wei Y, Yin X, Zhang J, Zhang LJ. A Coronary CT Angiography Radiomics Model to Identify Vulnerable Plaque and Predict Cardiovascular Events. Radiology 2023;307:e221693. [Crossref] [PubMed]
15. Lo Iacono F, Maragna R, Guglielmo M, Chiesa M, Fusini L, Annoni A, et al. Identification of subclinical cardiac amyloidosis in aortic stenosis patients undergoing transaortic valve replacement using radiomic analysis of computed tomography myocardial texture. J Cardiovasc Comput Tomogr 2023;17:286-8. [Crossref] [PubMed]
16. Lu X, Gong W, Yang W, Peng Z, Zheng C, Zha Y. Deep learning-based radiomics of computed tomography angiography to predict adverse events after initial endovascular repair for acute uncomplicated Stanford type B aortic dissection. Eur J Radiol 2024;175:111468. [Crossref] [PubMed]
17. Crepy D'Orleans C, Duwayri YM, Zellner AB, Binongo J, Farrington WJ, Keeling WB, Jordan WD Jr, Leshnower BG. Acute Type B Intramural Hematoma: Novel Insights in the Endovascular Era. Ann Vasc Surg 2024;101:195-203. [Crossref] [PubMed]
18. Chen M, Yang F, Chen L, Liu J, Luo S, Li J, Huang W, Liu Y, Fan R, Geng Q, Chen J, Luo J. Volumetric assessment is superior to axial measurements in estimating the outcome of patients with acute uncomplicated type B aortic intramural haematoma. Eur J Cardiothorac Surg 2022;62:ezac160. [Crossref] [PubMed]
19. Vaddavalli VV, Wilkins PB, Schaller MS, Morrison JJ, Shuja F, Colglazier JJ, Mendes BC, Rasmussen TE, Kalra M, DeMartino RR. Natural history and long-term outcomes of medically managed Type B intramural hematoma. J Vasc Surg 2024;80:648-655.e2. [Crossref] [PubMed]
20. Wei C, Li J, Du E, Miao Y, Li P, Guan W. Clinical and imaging differences between Stanford Type B intramural hematoma-like lesions and classic aortic dissection. BMC Cardiovasc Disord 2023;23:378. [Crossref] [PubMed]
21. Erbel R, Aboyans V, Boileau C, Bossone E, Bartolomeo RD, Eggebrecht H, et al. 2014 ESC Guidelines on the diagnosis and treatment of aortic diseases: Document covering acute and chronic aortic diseases of the thoracic and abdominal aorta of the adult. The Task Force for the Diagnosis and Treatment of Aortic Diseases of the European Society of Cardiology (ESC). Eur Heart J 2014;35:2873-926. [Crossref] [PubMed]
22. Colacchio EC, Squizzato F, Piazza M, Menegolo M, Grego F, Antonello M. Clinical and Imaging Predictors of Disease Progression in Type B Aortic Intramural Hematomas and Penetrating Aortic Ulcers: A Systematic Review. Diagnostics (Basel) 2022.
23. Mussa FF, Horton JD, Moridzadeh R, Nicholson J, Trimarchi S, Eagle KA. Acute Aortic Dissection and Intramural Hematoma: A Systematic Review. JAMA 2016;316:754-63. [Crossref] [PubMed]
24. Hu X, Yang F, Liu J, Liu Y, Fan R, Luo J. A single-center experience of type B aortic intramural hematoma. J Vasc Surg 2024;79:514-25. [Crossref] [PubMed]
25. von Kodolitsch Y, Csösz SK, Koschyk DH, Schalwat I, Loose R, Karck M, Dieckmann C, Fattori R, Haverich A, Berger J, Meinertz T, Nienaber CA. Intramural hematoma of the aorta: predictors of progression to dissection and rupture. Circulation 2003;107:1158-63. [Crossref] [PubMed]
26. Ferrera C, Vilacosta I, Gómez-Polo JC, Villanueva-Medina S, Cabeza B, Ortega L, Cañadas V, Carnero-Alcázar M, Martínez-López I, Maroto-Castellanos L, Serrano-Hernando FJ. Evolution and prognosis of intramural aortic hematoma. Insights from a midterm cohort study. Int J Cardiol 2017;249:410-3. [Crossref] [PubMed]
27. Spanos K, Kölbel T. Role of Initial Focal Contrast Enhancement in Type B Intramural Hematoma. J Am Coll Cardiol 2024;83:514-5. [Crossref] [PubMed]
28. Chen L, Yang F, Liu J, Luo S, Yuan H, Fan R, Zeng H, Geng Q, Luo J. Risk stratification of ulcer-like projection in uncomplicated acute type B aortic intramural haematoma. Eur J Cardiothorac Surg 2021;60:1032-40. [Crossref] [PubMed]
29. Hagan PG, Nienaber CA, Isselbacher EM, Bruckman D, Karavite DJ, Russman PL, et al. The International Registry of Acute Aortic Dissection (IRAD): new insights into an old disease. JAMA 2000;283:897-903. [Crossref] [PubMed]
30. Cundari G, Marchitelli L, Pambianchi G, Catapano F, Conia L, Stancanelli G, Catalano C, Galea N. Imaging biomarkers in cardiac CT: moving beyond simple coronary anatomical assessment. Radiol Med 2024;129:380-400. [Crossref] [PubMed]
31. Lee SE, Hong Y, Hong J, Jung J, Sung JM, Andreini D, et al. Prediction of the development of new coronary atherosclerotic plaques with radiomics. J Cardiovasc Comput Tomogr 2024;18:274-80. [Crossref] [PubMed]
32. Luts J, Ojeda F, Van de Plas R, De Moor B, Van Huffel S, Suykens JA. A tutorial on support vector machine-based methods for classification problems in chemometrics. Anal Chim Acta 2010;665:129-45. [Crossref] [PubMed]
33. Tang J, Henderson A, Gardner P. Exploring AdaBoost and Random Forests machine learning approaches for infrared pathology on unbalanced data sets. Analyst 2021;146:5880-91. [Crossref] [PubMed]
34. Siddalingappa R, Kanagaraj S. K-nearest-neighbor algorithm to predict the survival time and classification of various stages of oral cancer: a machine learning approach. F1000Res 2022;11:70. [Crossref] [PubMed]
35. Avanzo M, Wei L, Stancanello J, Vallières M, Rao A, Morin O, Mattonen SA, El Naqa I. Machine and deep learning methods for radiomics. Med Phys 2020;47:e185-202. [Crossref] [PubMed]