A predictive model for postoperative adverse outcomes following surgical treatment of acute type A aortic dissection based on machine learning

Abstract

Acute type A aortic dissection (AAAD) has a high probability of postoperative adverse outcomes (PAO) after emergency surgery, so exploring the risk factors for PAO during hospitalization is key to reducing postoperative mortality and improving prognosis. An artificial intelligence approach was used to build a predictive model of PAO by clinical data-driven machine learning to predict the incidence of PAO after total arch repair for AAAD. This study included 380 patients with AAAD. The clinical features that are associated with PAO were selected using the LASSO regression analysis. Six different machine learning algorithms were tried for modeling, and the performance of each model was analyzed comprehensively using receiver operating characteristic curves, calibration curve, precision recall curve, and decision analysis curves. Explain the optimal model through Shapley Additive Explanation (SHAP) and perform an individualized risk assessment. After comprehensive analysis, the authors believe that the extreme gradient boosting (XGBoost) model is the optimal model, with better performance than other models. The authors successfully built a prediction model for PAO in AAAD patients based on the XGBoost algorithm and interpreted the model with the SHAP method, which helps to identify high-risk AAAD patients at an early stage and to adjust individual patient-related clinical treatment plans in a timely manner.

INTRODUCTION

Aortic dissection is a rare but life-threatening cardiovascular emergency, which typically leads to poor prognosis and even death. Especially in the case of acute type A aortic dissection (AAAD), the mortality rate increases by about 1% per hour from the onset of symptoms, with a mortality rate as high as 50% within 48 h.^1–3 Hypertension is currently the most widespread public health issue globally and is closely associated with the stress on the aortic wall, often being a critical risk factor for AAAD.^2,4 As the most serious complication of hypertension, AAAD has been receiving increasing attention in the prevention and treatment of hypertension-related complications.

For AAAD, emergency surgery is the only effective treatment method. With the advancement of surgical techniques and the maturation of perioperative organ support and protection techniques, the prognostic outcome of AAAD surgery has improved significantly compared to the previous period.⁵ However, patients with AAAD are still susceptible to postoperative adverse outcomes (PAO). PAO significantly prolongs the patient's hospital stay and increases in-hospital mortality.^6,7 Although there have been studies attempting to analyze the relevant risk factors for PAO to predict and improve the prognosis of AAAD patients, this cannot comprehensively and accurately assess the individual risk of PAO for each patient.^8–10

As a special form of artificial intelligence, machine learning has the ability to select appropriate algorithms from a large amount of data, automatically induce logic or rules, and generate predictions of results.¹¹ Compared to traditional clinical prediction models based on logistic regression, prediction models built on machine learning are more accurate and practical. They are currently widely used in clinical diagnosis and treatment of various diseases.^12–14 However, there is currently no machine learning prediction model for postoperative outcomes in AAAD patients.

Therefore, this study aims to establish a machine learning prediction model to predict the risk of PAO in AAAD patients. The results of this study will be beneficial for clinicians to early identify high-risk individuals prone to PAO and provide timely targeted individualized treatment.

METHODS AND MATERIALS Study design and setting

We analyzed the clinical data of AAAD patients who underwent emergency surgical treatment at our center from January 2018 to January 2022. The inclusion criteria were as follows: (1) Stanford type A aortic dissection confirmed by computerized tomography angiography, or echocardiography; (2) age > 18 years; and (3) intraoperative repair with triple-branched stent graft implantation for total arch repair. The exclusion criteria were: (1) incomplete case data; (2) non-emergency surgical treatment; and (3) patients with malignant tumors, hematological diseases, active inflammatory diseases (such as autoimmune diseases and infections), chronic liver or kidney diseases, or long-term use of oral anticoagulants and glucocorticoids.

Definitions and grouping methods

In this study, PAO were defined as clinical events of grade III or higher complications, including death, occurring in AAAD patients after surgery, as stated in the consensus statement.⁷ The in-hospital mortality rate of patients was 5.79% (22/380). There was a total of 213 patients without complications or only experiencing grade II or lower complications (non-PAO group), while 167 patients (43.95%) experienced postoperative adverse outcomes (PAO group) (Table 3).

Surgical technique

All surgeries were performed through a mid-sternotomy incision, under general anesthesia and with the assistance of cardiopulmonary bypass. The specific surgical procedures were described in detail in a previous study by Chen and colleagues.¹⁵ The main surgical steps included aortic root reconstruction (such as sinus forming, Bentall procedure, David procedure), ascending aortic replacement, and the implantation of the triple-branched stent graft for total arch repair.

Data collection and processing

We obtained the patients' clinical data through the electronic medical record system. The demographic data included sex, age, smoking history, drinking history, height, weight, body mass index, admission systolic blood pressure, diastolic blood pressure. The preoperative clinical characteristics included hypertension, diabetes, coronary heart disease, chronic obstructive pulmonary disease, Marfan syndrome, hepatic dysfunction, renal insufficiency, left ventricular ejection fraction, pericardial effusion (medium or above), acute aortic regurgitation, symptoms at onset, and preoperative laboratory biochemical tests (white blood cell count, red blood cell count, hemoglobin, etc.). Intraoperative conditions included the aortic root procedure, total operation time, cardiopulmonary bypass time, aortic cross-clamp time, cerebral perfusion time, deep hypothermic circulatory arrest time, and transfusion of blood products (red blood cell, plasma, platelet). The least absolute shrinkage and selection operator (LASSO) regression analysis was used to select the clinical characteristic variables mentioned above.

Model construction

Using the Python software (version 3.7), a machine learning model was constructed. After inputting the final nine clinical characteristic variables, six machine learning algorithms were applied for model construction. These algorithms include extreme gradient boosting (XGBoost), logistic regression (LR), random forest (RF), gaussian naive bayes (GNB), support vector machine (SVM), and k-nearest neighbor (KNN) models. These algorithms can predict the occurrence rate of PAO.

Establishment, evaluation, and model interpretation of the predictive model

First, we use LASSO regression analysis to select potential clinical feature variables for predicting PAO from all the available features. After that, we randomly split all AAAD patients into a training set and a test set with an 8:2 ratio. Next, we analyze the training set using six different machine learning algorithms to predict the risk of PAO. We then fine-tune the model parameters using the validation set to obtain the optimal model performance (Table S3). Subsequently, we evaluate the predictive performance of all models on the training set and validation set by establishing and calculating the receiver operating characteristic (ROC) curves and the area under the curve. To prevent overfitting and improve the generalization ability of the models, the mean and standard deviation of the ROC for all models are obtained through ten-fold resampling-validation, ensuring that the evaluation performance of the models aligns with their actual performance. For a comprehensive evaluation of the models, we also established decision curve analysis to assess the clinical utility of all models. Calibration curves were used to evaluate the predictive accuracy of the models by assessing the deviation between the calibrated predicted probabilities and the actual occurrence probabilities. Furthermore, we assessed the predictive performance of the models using additional evaluation metrics such as precision-recall curves, as well as confusion matrix metrics including accuracy, sensitivity, specificity, positive predictive value, negative predictive value, and F1 score. These metrics provide a further complementary evaluation of the predictive performance of the models (Tables S1 and S2). Once the optimal model is determined, we further validate the predictive performance of the model using the test set data. We also construct calibration curves, decision curve analysis, and learning curves for the optimal model. Finally, we visualize the output process of this predictive model using the Shapley additive explanations (SHAP) analysis method to provide explanatory insights.

Statistical analysis

All Statistical analyses were performed using R version 3.6.3 and python version 3.7. Continuous variables are represented using mean ± standard deviation or interquartile range, while categorical variables are represented using frequency or percentage (%). The normality of continuous variables can be checked using the Kolmogorov-Smirnov test. Student's t-tests were used for intergroup comparisons of continuous variables that conformed to normal distribution, and Mann-Whitney U tests were used for intergroup comparisons of continuous variables that were not normally distributed. Categorical variables were tested using the chi-square test or Fisher's exact test. The potential clinical feature variables were screened out using LASSO regression analysis, and six machine learning models were established based on their feature sets. When p < .05, the differences are considered statistically significant.

RESULTS Demographic information and clinical characteristics

This study included a total of 380 patients, among whom 167 (43.95%) experienced PAO during hospitalization. The in-hospital postoperative outcomes are presented in Table 4. The demographic characteristics, preoperative, and intraoperative details of the two groups of patients are presented in Tables 1–3. All patients were randomly divided into a training set and a test set in an 8:2 ratio, and the modeling was performed using the training set. The process of model establishment and research workflow are illustrated in Figure 1.

TABLE 1 Comparison of preoperative characteristics of patients in the two groups.

Valuables	non-PAO group (n = 213)	PAO group (n = 167)	p value
Sex (Male), n (%)			.55
Male	74 (35)	63 (38)
Female	139 (65)	104 (62)
Age (year), median [IQR]	52.00 [41.00,61.00]	53.00 [45.00,63.00]	.02*
Smoking history, n (%)	103 (48)	87 (52)	.47
Drinking history, n (%)	57 (27)	54 (32)	.24
Physical examination
Height (cm)	170.00 [163.00,173.00]	170.00 [162.00,173.00]	.58
Weight (kg)	70.00 [63.00,80.00]	69.00 [60.00,78.00]	.27
BMI (Kg/Mˆ2), mean (±SD)	24.97 ± 3.43	24.74 ± 3.53	.52
SBP (mmHg), median [IQR]	142.00 [121.00,163.00]	140.00 [123.00,155.00]	.47
DBP (mmHg), mean (±SD)	75.62 ± 16.56	74.37 ± 15.29	.45
Medical history
Hypertension, n (%)	165 (77)	130 (78)	.93
Diabetes, n (%)	8 (4)	10 (6)	.31
Coronary artery disease, n (%)	2 (1)	2 (1)	.81
Chronic obstructive pulmonary disease, n (%)	2 (1)	6 (4)	.07
Marfan Syndrome, n (%)	5 (2)	2 (1)	.41
Hepatic dysfunction, n (%)	6 (3)	9 (5)	.20
Renal insufficiency, n (%)	4 (2)	4 (2)	.73
LVEF (%), median [IQR]	64.70 [60.40,68.00]	63.50 [60.60,66.80]	.19
Pericardial effusion (Medium or above), n (%)	13 (6)	15 (9)	.29
Aortic valve regurgitation (Medium or above), n (%)	76 (36)	34 (20)	<.01*
Symptom			.37
Chest pain, n (%)	190 (89)	152 (91)
Back pain, n (%)	15 (7)	6 (4)
Abdominal pain, n (%)	5 (2)	4 (2)
Other, n (%)	3 (1)	5 (3)

Abbreviations: BMI, body mass index; DBP, diastolic blood pressure; IQR, interquartile range; LVEF, left ventricular ejection fractions; SD, standard deviation; SBP, systolic blood pressure.

*

p < .05

TABLE 2 Comparison of preoperative laboratory results of patients in the two groups.

Valuables	non-PAO group (n = 213)	PAO group (n = 167)	p value
White blood cell count (×10ˆ9/L), mean (±SD)	12.28 [9.69,14.29]	13.20 [10.19,16.13]	.02*
Red blood cell count (×10ˆ12/L), mean (±SD)	4.35 [4.06,4.70]	4.28 [4.07,4.74]	.96
Heamoglobin (g/L), mean (±SD)	132.00 [121.00,144.00]	132.00 [122.00,144.00]	.64
Platelet (×109/L), median [IQR]	187.00 [149.00,223.00]	171.00 [131.00,211.00]	.03*
Albumin (g/L), mean (±SD)	38.60 [35.00,41.30]	37.80 [34.50,40.80]	.21
Total bilirubin (μmol/L), median [IQR]	15.60 [11.80,22.80]	15.60 [12.10,20.60]	.98
Direct bilirubin (μmol/L), median [IQR]	3.20 [2.00,5.20]	3.10 [1.90,5.30]	.84
Alanine aminotransferase (IU/L), median [IQR]	23.00 [15.00,38.00]	28.00 [17.00,48.00]	.04*
Aspartate aminotransferase (IU/L), median [IQR]	24.00 [19.00,42.00]	29.00 [21.000,65.00]	.01*
Alkaline phosphatase (IU/L), median [IQR]	70.62 [56.00,85.00]	72.00 [56.00,82.00]	.93
Lactate Dehydrogenase (IU/L), median [IQR]	247.00 [205.00,296.00]	266.00 [232.00,353.00]	<.01*
Serum creatinine (μmol/L), median [IQR]	80.00 [65.00,113.00]	98.00 [70.00,135.00]	<.01*
Blood urea nitrogen (mmol/L), median [IQR]	6.20 [5.10,7.70]	6.90 [5.80,8.60]	<.01*
PT (s), median [IQR]	13.90 [13.10,14.70]	13.90 [13.30,14.80]	.40
TT (s), median [IQR]	17.60 [16.40,18.90]	17.60 [16.30,19.30]	.97
APTT (s), median [IQR]	37.90 [34.80,42.50]	38.10 [34.90,40.90]	.97
Fibrinogen (g/L), median [IQR]	2.71 [2.06,3.49]	2.80 [2.06,3.66]	.84
D-dimer (ug/mL), mean (±SD)	11.22 [4.51,20.00]	12.07 [5.95,20.00]	.28
Creatine kinase (IU/L), median [IQR]	123.95 [78.00,186.51]	126.27 [86.97,201.00]	.61
Creatine kinase-MB (IU/L), median [IQR]	13.20 [8.40,18.00]	13.44 [10.00,18.00]	.29
B-type natriuretic peptide (pg/mL), median [IQR]	246.00 [110.00,661.00]	247.00 [124.00,657.00]	.68
Troponin-I (ug/L), median [IQR]	0.01 [0.002,0.08]	0.01 [0.002,0.07]	.85
C-reactive protein (mg/L), median [IQR]	10.95 [3.39,34.10]	9.23 [3.51,29.10]	.49
Procalcitonin (ug/L), median [IQR]	0.12 [0.06,0.27]	0.12 [0.06,0.36]	.44

Abbreviations: APTT, activated partial thromboplastin time; IQR, interquartile range; PT, prothrombin time; SD, standard deviation; TT, thrombin time.

*

p < .05.

TABLE 3 Comparison of intraoperative conditions between the two groups.

Valuables	non-PAO group (n = 213)	PAO group (n = 167)	p value
Intraoperative time
Total operative time (min), median [IQR]	288.00 [260.00,325.00]	303.00 [265.00,340.00]	.01*
Cardiopulmonary bypass time (min), median [IQR]	138.00 [119.00,155.00]	138.00 [128.00,163.00]	.12
Aortic cross-clamp time (min), median [IQR]	47.00 [36.00,67.00]	55.00 [44.00,71.00]	<.01*
Cerebral perfusion time (min), median [IQR]	8.00 [7.00,10.00]	8.00 [7.00,10.00]	.71
DHCA time (min), median [IQR]	3.00 [2.00,4.00]	3.00 [2.00,5.00]	<.01*
Aortic root procedure			.28
No treatment, n (%)	66 (31)	64 (38)
Sinus forming, n (%)	92 (43)	72 (43)
Bentall procedure, n (%)	54 (25)	30 (18)
Wheat procedure, n (%)	1 (0.5)	1 (1)
Intraoperative blood transfusion
Red blood cell transfusion volume (U), median [IQR]	4.00 [2.00,6.00]	4.00 [2.00,6.00]	.60
Plasma transfusion volume (mL), median [IQR]	400.00 [350.00,600.00]	400.00 [300.00,600.00]	.63
Platelet transfusion volume (U), median [IQR]	0.80 [0,1.00]	1.00 [0,4.00]	.02*

Abbreviations: DHCA, deep hypothermic cardiopulmonary arrest; IQR, interquartile range; SD, standard deviation.

*

p < .05.

TABLE 4 In-hospital postoperative outcomes in patients with AAAD.

Postoperative outcomes (n = 380)	Number	Percentage
Renal failure (need CRRT)	68	17.89
Respiratory failure	22	5.79
Gastrointestinal bleeding	16	4.21
Low cardiac output syndrome (need IABP)	12	3.16
Ventricular fibrillation	14	3.68
Permanent neurological deficits	37	9.74
Sepsis	32	8.42
Secondary thoracotomy	3	0.78
Secondary intubation	20	5.26
Tracheotomy	11	2.89
Pericardial effusion	6	1.57
Myocardial ischemia	8	2.11
Death	22	5.79

Abbreviations: AAAD, acute type A aortic dissection; CRRT, continuous renal replacement therapy; IABP, intra-aortic balloon pump.

FIGURE 1. Model construction methodology and study flowchart.

Feature variable selection

In the LASSO regression model, a vertical line was drawn at the selected value using 10-fold cross-validation. The lambda value corresponding to the minimizes the standard error of the distance was 0.04. This value resulted in nine non-zero coefficient feature variables (Figure 2A,B), which are: age, left ventricular ejection fraction, acute aortic regurgitation, white blood cell, creatinine, total operation time, deep hypothermic circulatory arrest time, aortic root procedure, and platelet transfusion volume.

FIGURE 2. Selection of potential clinical characteristics factors associated with PAO by LASSO regression analysis. (A) By performing ten-fold cross-validation and plotting a vertical line at the selected value, the optimal lambda value produces nine non-zero feature coefficients. (B) In the LASSO regression model, the lambda value that minimizes the mean square error is 0.031, while the lambda value that minimizes the standard error of the distance is 0.04.

Classified multi-model comparison

To perform a comprehensive analysis on the training set, the XGBoost, LR, RF, GNB, SVM, and KNN algorithms were employed. The performance of the classification models was evaluated by calculating the area under the receiver operating characteristic. The mean and standard deviation of the area under the receiver operating characteristic for all models were computed using ten-fold resampling. The final results indicated that RF exhibited the best performance on the training set and XGBoost demonstrated the best performance on the validation set (Figure 3A,B). The forest plot in Figure S2 shows the area under the curve scores for all models. Due to the area under the curve's focus on evaluating the prediction accuracy of the model, it is necessary to employ other methods to further evaluate the model's strengths, weaknesses, and clinical applicability. So, all models were comprehensively evaluated using calibration curves, decision curve analysis (Figure 3C,D), and precision recall curves (Figure S1). The results of the decision curve analysis indicate that XGBoost model can achieve good clinical applicability. The calibration curve shows that the XGBoost model has lower Brier scores, indicating higher predictive accuracy. Based on the above analysis, we believe that the RF model may suffer from overfitting, while XGBoost model appears to have relatively good stability. Therefore, we can consider the XGBoost model as the optimal model.

FIGURE 3. The comprehensive analysis of six machine learning models. (A) The ROC curve and area under the curve of the training set. (B) The ROC curve and area under the curve of the validation set. (C) The calibration curve plot of the models. (In the calibration curve, the x-axis represents the average predicted probability, and the y-axis represents the actual probability of a positive event occurring. The diagonal line represents the perfectly calibrated reference line. The solid lines of different colors correspond to the fitting lines of the respective models. The smaller the Brier score value in parentheses, the higher the predictive accuracy of the model). (D) The decision curve analysis of the validation set.

Model development and evaluation

The XGBoost algorithm was applied for classification on the training set using 10-fold cross-validation. The results showed that the mean train ROC (95% confidence interval) was 0.997 (0.993–1.000) (Figure S3A), and the mean validation ROC was 0.716 (0.508–0.922) (Figure S3B). The final model achieved an ROC of 0.761 (0.668–0.854) on the test set, with an accuracy of 0.693 (Figure 4A). The results from the calibration curve and decision curve analysis indicate that the model has good predictive accuracy and high clinical utility (Figure 4C,D). The area under the curve results of the training set, validation set, and test set remained stable above 0.72 throughout the analysis. The learning curve results indicate that this model fits well on both the training and validation sets (Figure S3C).

FIGURE 4. The test, and clinical applicability of the XGBoost model. (A) The ROC curve and area under the curve of the test set. (B) The calibration curve plot of the XGBoost model (Brier score = 0.192). (C) The decision curve analysis of the XGBoost model.

Model interpretation

To further explain the predictive model, we used SHAP to illustrate how the feature variables contribute to the prediction of PAO occurrence. Figure 5A is the SHAP summary plot of our predictive model in this study. It displays a total of nine feature variables sorted according to their impact on PAO. In each feature variable, red indicates high feature values, blue indicates low feature values, and purple indicates feature values close to the average level. Figure 5B displays the importance of each feature variable in the development of the final predictive model, as evaluated by the mean absolute SHAP values.

FIGURE 5. Explaining the model using the SHAP analysis. (A) The scatter plot of feature distributions using the SHAP analysis. (B) Ranking feature importance based on the absolute mean values of SHAP values. (SHAP values represent the predictive features of individual patients and the contribution of each feature to predicting PAO). (C) Force plot for patients in the testing set with PAO. (D) Force plot for patients in the testing set without PAO. (Starting from the baseline, which is a constant for interpreting the model, each attribution value is represented by an arrow indicating an increase [positive value] or decrease [negative value] in the prediction. Red indicates positive contributions, while blue indicates negative contributions).

In addition, we provided an explanation of the model's interpretability by describing two independent samples from the test set. One sample is an AAAD patient without PAO, with a lower SHAP predicted score (f(x) = 0.106) (Figure 5C). The other sample is an AAAD patient with PAO, with a higher SHAP predicted score (f(x) = 0.768) (Figure 5D).

DISCUSSION

Hypertension has become a common chronic disease worldwide, and its prevalence is further increasing due to population aging and lifestyle changes.¹⁶ Therefore, prevention and control of hypertension and its related complications are crucial for public health. As one of the most severe complications of hypertension, aortic dissection often leads to catastrophic complications or even death.¹⁷ Although surgical intervention can save the lives of patients with AAAD, the short-term prognosis remains poor mainly due to severe postoperative complications such as organ dysfunction, sepsis, which lead to PAO.^1,18 Once patients experience PAO, their hospitalization time, hospitalization costs, and even in-hospital mortality rates significantly increase, which hampers the recovery process. Although there are many risk factors for PAO, there is currently a lack of effective predictive models to individualize the prognosis for each patient.¹⁰

So far, our study is the first to apply machine learning algorithms to predict AAAD patients' PAO. Based on different machine learning algorithms, we successfully trained and validated six machine learning models for assessing the risk of PAO during hospitalization in AAAD patients. Among these models, the RF model and XGBoost model demonstrated superior performance, but the RF model may have a possibility of overfitting, while the XGBoost model generated more accurate and reliable predictions. Therefore, we chose the model based on the XGBoost algorithm as the final model.

The XGBoost algorithm is an improved version based on the gradient boosting framework. Compared to traditional logistic regression methods, it performs better in complex classification problems, with advantages such as high accuracy, automatic handling of complex features, strong interpretability, ability to handle imbalanced data, and wide applicability.^19–21 Previous research results have shown that predictive models built based on the XGBoost algorithm can accurately assess the risk of acute kidney injury and guide early treatment.²² In a study on machine learning-based prediction of in-hospital mortality after sepsis, the XGBoost model has also been validated for its ability to handle large datasets and analyze complex relationships between variables, showing superior prognostic prediction capability.²³ In this study, although we attempted to build multiple machine learning predictive models, the XGBoost model still demonstrated the best performance in predicting postoperative pulmonary artery occlusion in AAAD patients, consistent with previous research results. Therefore, clinicians can make targeted individualized treatment based on the predictive results of the XGBoost model, which can help further improve in-hospital outcomes and prognosis for patients.

To further visualize the predictions of the XGBoost model, we used the SHAP method to explain the model. By plotting the feature density scatter plot of SHAP values and the permutation feature importance, we can improve the interpretability of the model and demonstrate the actual predictions for several samples in the dataset. The results of SHAP feature importance ranking indicate that preoperative acute aortic regurgitation, total operation time, and preoperative white blood cell count have been identified as the most important variables associated with adverse postoperative outcomes in patients with AAAD. When the dissection involves the aortic root or causes severe dilation of the aortic annulus, it often leads to acute aortic regurgitation. The research results of Iarussi and colleagues²⁴ indicate that preoperative concomitant acute aortic regurgitation in AAAD patients is more prone to postoperative serious complications such as heart failure or arrhythmia. This may be due to the hemodynamic instability caused by acute aortic regurgitation, resulting in a certain degree of cardiac dysfunction and myocardial damage. Due to the significant trauma inflicted on the body by aortic dissection, the systemic inflammatory response continues to be activated and intensified, greatly affecting the function of vital organs such as the heart, lungs, and brain, leading to adverse prognostic outcomes. As a reliable marker of the body's inflammatory level, WBC levels are often closely associated with prognostic outcomes.²⁵ Research by Zhang and colleagues²⁶ indicates a positive correlation between perioperative elevated white blood cell counts due to inflammation and increased in-hospital mortality rates. Although the direct impact of preoperative white blood cell count on PAO is not yet clear, the conclusion that patients with high white blood cell counts have a significantly higher incidence of postoperative complications and mortality rates than those with low white blood cell counts has been confirmed.^27,28 The surgical procedure for AAAD is extremely complex, involving the use of cardiopulmonary bypass and deep hypothermic circulatory arrest to perform the repair of the aortic arch. Therefore, the longer the operation time, the greater the trauma and stress caused by the surgery. Prolonged cardiopulmonary bypass or deep hypothermic circulatory arrest further participants the body to a secondary insult, resulting in adverse postoperative prognostic outcomes.

In summary, we have successfully established and validated a model based on the XGBoost algorithm that can be used to predict the risk of PAO in AAAD patients. The clinical utility of this model was comprehensively evaluated through calibration curves, precision recall curves, and decision curve analysis. In the process of preventing and treating complications related to hypertension, AAAD is often the most risky and daunting condition primarily due to its poor prognosis and high mortality rate. This predictive model can effectively assist non-surgical physicians in assessing the postoperative risks and prognostic outcomes of AAAD patients, providing a more intuitive understanding of potential risk factors that may impact the postoperative prognosis of AAAD. Consequently, this could further improve the prognostic outcomes of this severe hypertension complication.

However, our study also has some limitations. First, the data comes from a single center, and the sample size is relatively small. Due to the retrospective nature of the study, there may be some degree of selection bias, and further expansion of the sample size is needed to validate the model. Second, since all the data are from AAAD patients who underwent total arch repair with the triple-branched stent graft at our center, the accuracy of the model when used for prediction in other medical institutions needs to be verified. Therefore, when the model is used in other centers, it is necessary to recalibrate the model's parameters, which may affect the weights of certain clinical variables. Last, as the model lacks an independent dataset for external validation, its generalizability and applicability need to be tested. We will further collect external validation data to test the clinical utility of the model.

CONCLUSIONS

This study attempted to develop six different prediction models for PAO in AAAD patients and evaluated the models comprehensively using the area under the receiver operating characteristic, calibration curve, decision curve analysis, and precision recall curve. Finally, the best-comprehensive performing algorithm with the highest clinical utility was selected for modeling. In the end, we established a predictive model for PAO based on the XGBoost algorithm. This model aims to identify high-risk AAAD patients early on and develop individualized diagnosis and treatment plans to reduce the risk of PAO and improve the prognosis of AAAD patients. It serves as an important tool to assist clinicians in making informed decisions and developing effective strategies.

AUTHOR CONTRIBUTIONS

Lin-feng Xie was responsible for writing the manuscript; Yu-ling Xie was responsible for reviewing & editing the manuscript; Jian He was responsible for data curation; Xin-fan Lin was responsible for software processing; Qing-song Wu was responsible for funding acquisition; Zhi-huang Qiu was responsible for resources; and Liang-wan Chen was responsible for methodization and methodology.

ACKNOWLEDGMENTS

The authors thank the patients who participated in the study and the research assistants and study coordinators who assisted with data collection and management of the study, including Zhao-feng Zhang and Xing-hui Zhuang. This research was sponsored by the Startup Fund for Scientific Research, Fujian Medical University (2022QH2019), Fujian Provincial Center for Cardiovascular Medicine Construction Project (NO.2021-76), and Key Laboratory of Cardio-Thoracic Surgery (Fujian Medical University), Fujian Province University Construction Project (No.2019-67).

CONFLICT OF INTEREST STATEMENT

The authors declare that they have no conflicts of interest to report regarding the present study.

DATA AVAILABILITY STATEMENT

The data underlying this article cannot be shared publicly due to the privacy of individuals that participated in the study. The data will be shared on reasonable request to the corresponding author.