Between January 2015 and February 2019, a total of 300 ATAAD patients, diagnosed by aortic CTA and with no history of ischemic stroke or cerebrovascular disease, were retrospectively included in this study. ATAAD patients were divided into two groups according to the presence (AIS+ group) or absence (AIS- group) of pre-operative AIS. Clinical characteristics, TTE imaging information and CTA imaging information were collected. Detailed information was shown in Table 1 (clinical and TTE) and Table 2 (CTA).

ATAAD was defined as any non-traumatic dissection of the aorta proximal to the left subclavian artery presenting within 14 days of symptom onset (1). All included ATAAD patients in this study are involved with ascending aorta as well as a primary entry tear either in the ascending aorta or the aortic arch (prior to the left subclavian artery). AIS was defined as a cerebrovascular accident representing a loss of neurological function (loss or slurring of speech, altered state of consciousness) caused by an ischemic event and further confirmed by diffusion MRI. Because the MRI examination room was adjacent to CT examination room in our emergency department and only cranial DWI sequence was performed, the total examination time was <5 min.

This study complied with the Helsinki Declaration (2000) and was approved by the institutional review board of Xijing Hospital affiliated with the Fourth Military Medical University (20120216-4). Informed consent was obtained from each patient or their legal representative.

CTA examinations were performed using a second-generation dual source CT machine (Somatom Definition Flash; Siemens Healthcare, Forchheim, Germany) with a high-pitch spiral scan mode. Patients underwent combined CTA imaging of the neck and aorta in the cranio-caudal direction. For all scans, patients were in a supine position with both arms raised. Each patient received an injection of 70 ml of iopromide (Ultravist 370, 370 mgI/mL; Bayer Schering Pharma, Berlin, Germany) at a flow rate of 5 mL/s, followed by 40 mL of saline solution at a flow rate of. Bolus tracking was performed on the suprarenal descending aorta with an attenuation threshold of 100 HU. The scanning parameters were as follows: tube voltage of 100 kV, reference tube current of 300 mAs per rotation, pitch of 3.0 and slice collimation of 2 × 128 × 0.6 mm by means of a z-flying focal spot. The CTA images were transferred to an external workstation (syngo MMWP VE 36 A; Siemens Healthcare, Forchheim, Germany) for further postprocessing. Finally, CTA characteristics were identified from each patient CTA imaging by two experts with at least 10 years' experience on radiology imaging.

In this study, dataset is stratified split into training (70%) and validation (30%) cohort. A four layers deep neural network was performed with all the clinical, TTE and CTA imaging characteristics as input. The random search method (12) was used to select the best number of nodes for each hidden layer and the number of nodes for output layer was set to 1 represent AIS+ probability (Figure 1). In order to overcome the influence of data imbalance focal loss (10) were used as loss function as shown in equation (1). The Adam optimizer was used and model was trained with 300 epochs.

Whereγ is a adjust factor andγ>0

In addition, the univariate and multivariate stepwise logistic regression analysis (Uni + Multi analysis) (11), the 10-fold CV based Least absolute shrinkage and selection operator (CV LASSO) risk factor selection method (13), support vector machine (SVM) model (14), random forest (RF) model (15), and neural network (NN) were also used for comparison. According to the number of input characteristics, the number of NN's nodes for each hidden layer were set as 16, 8, and 4, respectively. The number of nodes for output layer was set to 2, the same as the number of patient's group. The deep belief network (DBF) (16) training method was used to initialize the weights of the NN. Briefly, the first four layers of NN consisted of three restricted Boltzmann machines (RBMs). When pervious RBM finished the training, the trained weights would be used as the initialized weights for the corresponding NN layer, and the output of this trained RBM would be used as the input for the current RBM training.

For Uni + Multi analysis model, characteristics with significant difference between AIS+ and AIS- group in univariate model were included in the construction of multivariate stepwise logistic regression model. For CV LASSO model, all variables were selected through CV LASSO algorithm. For SVM, RF, NN and DNN model, all variables were included in the construction of the model. The first difference between NN and DNN was that the number of nodes for each hidden layer in NN was manually set according to number of input characteristics, while, for DNN, these number was optimized through random search method, which could make the performance of model to the best. The second difference was the weight initialization (DBF training method for NN and random initialization for DNN).

The training and validate of the deep neural network and DBN were implemented using TensorFlow 2.0 with GPU support on a Python 3.8 platform (https://www.python.org/). The other conventional machine learning model was implemented with R software (Version 4.1.0, https://www.rproject.com/).

All statistical analysis was performed using R software. Summary statistics are presented as counts with percentages for categorical values, as mean ± standard deviation for normally distributed continuous variables, or as medians with quartiles for non-normally distributed continuous variables. Data distribution was checked using the Shapiro-Wilk test. Continuous variables with normal data distributions were compared using two-sample t tests. For data with skewed distributions, non-parametric Mann-Whitney tests were used. Categorical variables were compared using chi-square statistics, and the Fisher exact test was used if observed frequencies were <5. Receiver operating characteristic (ROC) curve as well as precision recall curve (PRC) analysis and the area under the curve (AUC) was calculated. The best cutoff point was obtained by the Youden-index of ROC curve, then accuracy, sensitivity, specificity, positive predictive rate, and negative predictive rate were calculated. To estimate the clinical usefulness of the different models, decision curve analysis (DCA) was conducted by calculating the net benefits at different threshold probabilities (17). For all statistical analyses, p < 0.05 was considered statistically significant.

Pre-operative AIS was detected in 86 of 300 patients with ATAAD (28.7%). Among clinical and TTE imaging characteristics, the age (52.5 ± 9.8 years vs. 48.1 ± 10.5 years, p = 0.001), the incidence of hypotension (14.0 vs. 5.1%, p = 0.019) and tamponade (11.6 vs. 4.7%, p = 0.029) were significantly higher in the AIS+ group compared to the AIS- group, the systolic blood pressure (129.3 ± 34.0 vs. 137.5 ± 31.4 ml/Hg, p = 0.040) and the incidence of the chest and back pain (24.4 vs. 39.3%, p = 0.021) and were significantly lower in the AIS+ group compared to the AIS- group. Other clinical characteristics did not differ significantly between the AIS+ group and AIS- group. Detailed clinical and TTE imaging characteristics are shown in Table 1.

Among CTA imaging characteristics, the true lumen diameter of the aAO (11.9 ± 5.9 vs. 17.8 ± 7.3 mm, p < 0.001) and the true lumen diameter ratio of the aAO (0.25 ± 0.11 vs. 0.37 ± 0.15, p < 0.001) were lower in the AIS+ group than in the AIS- group. The intimal flap plaque (34.9 vs. 39.0%, p = 0.003), the CCA dissection (77.9 vs. 29.0%, p < 0.001), the innominate artery or CCA from false lumen (14.0 vs. 3.3%, p = 0.001), the low density of the unilateral ICA (27.9 vs. 6.1%, p < 0.001), the low density of vertebral artery (9.3 vs. 2.8%, p < 0.021), and the subclavian artery dissection (37.2 vs. 17.3%, p < 0.001) were higher in the AIS+ group than in the AIS- group. Other CTA imaging characteristics did not differ significantly between the AIS+ group and AIS- group. Detailed CTA imaging characteristics are shown in Table 2.

Among those methods, the model constructed by the deep neural network had the best performance than the other methods, with an accuracy of 0.934 (95% CI: 0.891–0.963), 0.921 (95% CI: 0.845–0.968), sensitivity of 0.934, 0.960, specificity of 0.933, 0.906, and AUC of 0.983 (95% CI: 0.967–0.997), 0.964 (95% CI: 0.932–0.997) in the training and validation cohort, respectively (Table 3). The ROC curves were shown in Figure 2 and the best cutoff point of deep neural network was 0.402. The PRC as illustrated in Figure 3, which demonstrated that deep neural network has the best AUC with highest precision (positive predict rate) and recall (sensitivity). The DCA curve for the different models were presented in Figure 4. The DCA showed that if the threshold probability between 0.05 and 0.694, using the deep neural network to predict AIS status adds more net benefit than either the other models, “treat all patients” or the “treat none” strategies.