Machine learning models for predicting postoperative paraplegia in acute type A aortic dissection patients

Abstract

Background

Acute type A aortic dissection (ATAAD) is an extremely life-threatening cardiovascular emergency characterized by a critical clinical course and high mortality rate. Once the diagnosis is confirmed, emergency surgical intervention is required to save the patient’s life. Postoperative paraplegia is a relatively rare but severe surgical complication, typically manifesting as transient or permanent loss of sensory and motor functions in both lower extremities after surgery. It severely impairs patients’ survival and prognosis, imposing a heavy burden on both patients’ families and the healthcare system. This study aimed to explore the risk factors for postoperative paraplegia in patients with ATAAD and to develop and validate a risk prediction model.

Methods

A retrospective analysis was performed on the clinical data of ATAAD patients who were admitted to Beijing Anzhen Hospital and underwent surgical treatment between August 2018 and July 2022. Postoperative paraplegia was defined as the primary outcome endpoint. Patients were divided into a training set (70%) and a validation set (30%) using stratified sampling. Least absolute shrinkage and selection operator (Lasso) regression was applied to the training set to screen for key variables influencing the primary outcome. Seven machine learning algorithms were used to construct risk prediction models. The predictive performance of these models was validated using confusion matrix metrics, including the area under the receiver operating characteristic curve (AUC). The optimal prediction model was selected, and Shapley Additive exPlanations (SHAP) analysis was conducted to interpret the model.

Results

Among the 572 patients included in this study, 22 (3.84%) developed paraplegia. Comprehensive evaluation of confusion matrix metrics showed that the Neural Networks model had the best AUC, lower Brier score, and higher F1 score and Kappa value. According to the SHAP analysis results, the risk factors most strongly associated with postoperative paraplegia were: pancreatic lipase level, left subclavian artery involvement, Sun’s procedure, age, pancreatic amylase level, hemoglobin level, and secondary surgery.

Conclusions

Among the seven machine learning models for predicting postoperative paraplegia in ATAAD patients, the Neural Networks model demonstrated the best predictive performance.

Highlight box.

Key findings

• This study investigated the risk factors for postoperative paraplegia in patients with acute type A aortic dissection (ATAAD) and successfully established a corresponding risk prediction model.

What is known and what is new?

• At present, there is a paucity of research focusing on postoperative paraplegia in patients with ATAAD, and no relevant studies on this specific topic have been reported to date.

• This study effectively identified a variety of factors—including preoperative laboratory test results and surgical approaches—that contribute to postoperative paraplegia in patients with ATAAD, and established multiple machine learning-based risk prediction models.

What is the implication, and what should change now?

• This study aims to guide clinicians in effectively identifying ATAAD patients who are at risk of developing postoperative paraplegia, facilitate the assessment of surgical risks, support the individualized design of surgical plans based on patients’ specific conditions, enable earlier identification of the onset of paraplegia in such patients, and thereby facilitate more timely and effective therapeutic interventions. Constrained by the low incidence of this disease and the relatively small number of positive outcomes in the current study, future research may further expand the sample size to identify the occurrence of postoperative paraplegia with higher precision.

Introduction

Acute type A aortic dissection (ATAAD) is an extremely life-threatening cardiovascular disease that requires urgent surgical intervention. According to relevant research statistics, within 48 hours of a confirmed ATAAD diagnosis, approximately 23.7% of patients die from aortic rupture, with the mortality risk increasing at a rate of 0.5% per hour (1). Surgical treatment is the primary therapeutic measure to prevent aortic rupture in patients (2). However, approximately 2% to 4% of ATAAD patients develop paraplegic symptoms after surgery. Such patients typically present with flaccid paralysis of the lower limbs, loss of pain and temperature sensation, urinary retention, and intestinal paralysis—conditions that significantly prolong hospital stay and increase in-hospital mortality (3). Early identification of patients at risk of postoperative paraplegia following surgery is crucial for improving patient prognosis.

Machine learning (ML) is an implementation approach of artificial intelligence (AI), whose core concept is to enable machines to learn from data and automatically identify feature points in data through algorithms, thereby assisting physicians in clinical decision-making (4). Compared with traditional logistic regression, ML has strong nonlinear processing capabilities: it can handle complex nonlinear relationships in medical data, automatically learn features and patterns, and achieve higher prediction accuracy in some complex medical tasks (5,6). Specifically, in predicting cardiovascular risks, ML may outperform traditional logistic regression (7,8).

Due to the relative rarity of such patients, there are currently few models available to predict postoperative paraplegia in this population. We developed and validated an ML-based risk prediction model to assess relevant risk factors and predict the likelihood of postoperative paraplegia after surgery. We present this article in accordance with the TRIPOD reporting checklist (available at https://jtd.amegroups.com/article/view/10.21037/jtd-2025-2000/rc).

Methods

Study population

This study retrospectively collected and analyzed the clinical data of 1,257 patients with ATAAD who were admitted to the Cardiac Surgery Center of Beijing Anzhen Hospital, Capital Medical University, and underwent surgical treatment between August 2018 and July 2022. A total of 572 cases that met the requirements of this study were selected based on the inclusion and exclusion criteria (Figure 1). This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Ethics Committee of Beijing Anzhen Hospital Affiliated to Capital Medical University (No. 2026013x). In view of the retrospective case-control design of this study, the requirement for individual informed consent from patients was waived.

Figure 1.

Inclusion and exclusion criteria. CTA, computed tomography angiography.

Inclusion criteria

1. Patients with Stanford type A aortic dissection confirmed by aortic computed tomography angiography (CTA) at Beijing Anzhen Hospital before surgery;

2. Age >18 years;

3. Patients who underwent emergency or elective surgical treatment;

4. Complete medical record data.

Exclusion criteria

1. Patients with Stanford type A aortic dissection who did not undergo surgical treatment;

2. Patients who did not undergo postoperative aortic CTA examination;

3. Incomplete medical record data;

4. Patients who had developed paraplegic symptoms before surgery.

Definitions and grouping methods

In accordance with international guidelines, paraplegia is defined as the impairment or loss of motor and/or sensory function at the corresponding segments following spinal cord injury (SCI) involving the thoracic, lumbar, or sacral segments of the spinal cord (9). The primary outcome endpoint selected in this study was the development of paraplegic symptoms in patients after ATAAD surgery. Patients were divided into two groups: the paraplegia group and the non-paraplegia group, based on whether the primary outcome occurred.

Data collection and processing

This study collected patients’ basic information [age, gender, smoking history, drinking history, body mass index (BMI), coronary artery disease, cerebrovascular disease, hypertension, diabetes, hyperlipidemia, etc.]; preoperative imaging examination indicators [ascending aortic diameter, branch vessel involvement, bicuspid aortic valve, aortic regurgitation, left ventricular ejection fraction (LVEF), left ventricular diameter (LVD), etc.]; preoperative medication and blood test indicators [use of anticoagulants, pancreatic lipase, pancreatic amylase, lactic acid, hemoglobin (Hb), myoglobin, alanine transaminase (ALT), aspartate transaminase (AST), etc.]; intraoperative factors [cardiopulmonary bypass (CPB) time, aortic cross-clamp time, deep hypothermic circulatory arrest (DHCA) time, nasopharyngeal temperature, rectal temperature, surgical method, etc.]; and main outcome indicator (paraplegia). Variables with a missing rate of ≥30% were excluded, and the multivariate imputation by chained equation (MICE) function was used for multiple imputation of missing values. Least absolute shrinkage and selection operator (Lasso) regression was used to identify key variables and exclude independent variables irrelevant to the main outcome, so as to improve the prediction accuracy of ML.

Statistical analysis

SPSS 27 and R V4.5.1 were used for statistical analysis in this study. Normality tests were performed on continuous variables: continuous variables that conformed to a normal distribution were compared using the t-test and expressed as mean ± standard deviation (SD). Continuous variables with a non-normal distribution were compared using the Mann-Whitney U test and expressed as median and interquartile range (IQR). Categorical variables were expressed as counts (percentages) and compared using the chi-squared test and Fisher’s exact test. A difference was considered statistically significant when P<0.05.

Model construction

Lasso regression was employed, which introduces an L1 regularization term to drive some regression coefficients towards zero, thereby selecting feature variables. Based on the occurrence of the primary outcome endpoint, the data were divided into a training set and a test set in a 7:3 ratio using stratified sampling. Resampling methods were used to adjust the data distribution of the training set and optimize the model training conditions. Seven popular ML algorithms were applied to construct risk prediction models: logistic regression, Support Vector Machine (SVM), Decision Tree, Neural Network, Random Forest, XGBoost, and K-Nearest Neighbor (KNN). Five-fold cross-validation was used to adjust the hyperparameters of the training set and improve the predictive performance of the models.

Model evaluation and interpretability analysis

The seven constructed ML models were validated using the validation set data, and confusion matrix metrics were obtained. These metrics included the area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, specificity, F1 score, Brier score, and Kappa value, which were used to evaluate the predictive ability of the models.

AUC can well clarify the predictive efficacy of each risk prediction model and is a commonly used method for evaluating the efficacy of ML models at present. The Brier score is an indicator for assessing the accuracy of probabilistic predictions; a lower Brier score indicates better predictive accuracy. The F1 score is a quantitative indicator for evaluating the correctness of model classification. As a harmonic mean that integrates Precision and Recall, a higher F1 score indicates a more perfect model classification. The Kappa value is a key indicator for measuring the consistency between the model’s predicted results and the actual results. It excludes the consistency caused by random guessing and reflects the true performance of the model more objectively.

Furthermore, the calibration curves and clinical decision curves of the validation set were analyzed to evaluate the model calibration and clinical efficacy. The optimal model was selected by comprehensively evaluating indicators such as the AUC value, Brier value, and Kappa value of each model.

The Shapley Additive exPlanations (SHAP) algorithm was used to calculate the SHAP value of each variable in the optimal model. The magnitude of the SHAP value reflects the degree of contribution of the variable to the model, and the importance of each feature variable to the model output was presented in the form of a bar chart.

Results

Patient characteristics

A total of 572 patients with ATAAD were enrolled in this study, among whom 22 patients developed paraplegia after surgery. Data on the demographic characteristics, preoperative imaging examination results, and laboratory indicators of all patients are presented in Table 1. Compared with the non-paraplegia group, patients in the paraplegia group had more severe involvement of the innominate artery, left common carotid artery, and left subclavian artery, were older, had a lower nasopharyngeal temperature during surgery, and a higher proportion of patients underwent Sun’s procedure and secondary surgery. In terms of preoperative laboratory tests, the paraplegia group showed higher levels of pancreatic enzymes, C-reactive protein, urea, creatine kinase-MB (CK-MB), lactate dehydrogenase (LDH), and myoglobin, while lower levels of platelets (PLT), Hb, and lymphocytes (LYM) were observed. These risk factors may lead to poor prognosis in patients.

Table 1. Baseline characteristics of ATAAD patients.

Characteristics	Total (n=572)	Paraplegia (n=22)	Non-paraplegia (n=550)	P
Primary branch				0.53
Aortic sinus	121 (21.2)	6 (27.3)	115 (20.9)
Ascending aorta	302 (52.8)	11 (50.0)	291 (52.9)
Aortic arch	109 (19.1)	5 (22.7)	104 (18.9)
Descending aorta	40 (7.0)	0 (0)	40 (7.3)
Ascending aortic diameter, mm	46.5 (41.025, 53.475)	45.1 (42.3, 52.275)	46.6 (41, 53.5)	0.94
Brachiocephalic artery	266 (46.5)	15 (68.2)	251 (45.6)	0.03
Left common carotid artery	203 (35.5)	12 (54.5)	191 (34.7)	0.057
Left subclavian artery	216 (37.8)	14 (63.6)	202 (36.7)	0.01
Anastomotic leak	78 (13.6)	4 (18.2)	74 (13.5)	0.58
Gender				0.91
Male	410 (71.7)	16 (72.7)	394 (71.6)
Female	162 (28.3)	6 (27.3)	156 (28.4)
Age, years	51 (41, 59)	55 (50, 63)	50 (40, 59)	0.04
Height, cm	172 (165, 176)	167.5 (164, 175)	172 (165, 176)	0.17
Weight, kg	75 (65, 86)	75.5 (64.5, 85)	75 (65, 86)	0.64
BMI, kg/m2	25.9 (23.725, 28.4)	25.8 (23.5, 29.575)	25.9 (23.7, 28.4)	0.87
Hypertension	361 (63.1)	16 (72.7)	345 (62.7)	0.34
Diabetes	12 (2.1)	0 (0)	12 (2.2)	0.48
Hyperlipidemia	26 (4.5)	2 (9.1)	24 (4.4)	0.29
Cerebrovascular disease	29 (5.1)	1 (4.5)	28 (5.1)	0.91
Coronary heart disease	58 (10.1)	2 (9.1)	56 (10.2)	0.87
Aspirin	8 (1.4)	1 (4.5)	7 (1.3)	0.20
Smoking history	255 (44.6)	10 (45.5)	245 (44.5)	0.93
Drinking history	184 (32.2)	9 (40.9)	175 (31.8)	0.37
Bicuspid aortic valve	16 (2.8)	0 (0)	16 (2.9)	0.42
Moderate-severe aortic regurgitation	220 (38.5)	11 (50.0)	209 (38.0)	0.26
Ejection fraction, %	60 (58, 65)	60 (58, 62.75)	60 (58, 65)	0.53
Left ventricular diameter, mm	49 (44, 54)	50 (46.25, 55.25)	49 (44, 54)	0.41
CPB pump time, min	184 (157.25, 211)	205 (167, 223.5)	184 (156, 211)	0.09
Aortic cross-clamp time, min	103 (87, 123.75)	104 (89, 121.5)	103 (87, 124)	0.73
Deep hypothermic circulatory arrest time, min	21 (15, 29.75)	21.5 (15.75, 31.25)	21 (15, 29)	0.60
Minimum nasopharyngeal temperature, °C	24.7 (24, 27)	24.25 (23.775, 24.83)	24.75 (24, 27.13)	0.051
Minimum rectal temperature, °C	26.1 (24.8, 28)	25.5 (24.775, 26.33)	26.1 (24.775, 28)	0.18
Surgical method
Ascending aorta replacement	223 (39.0)	12 (54.5)	211 (38.4)	0.13
Bentall	347 (60.7)	10 (45.5)	337 (61.3)	0.14
Wheat	1 (0.2)	0 (0)	1 (0.2)	0.84
David	3 (0.5)	0 (0)	3 (0.5)	0.73
Sun’s surgery	451 (78.8)	21 (95.5)	430 (78.2)	0.05
Partial aortic arch replacement	78 (13.6)	1 (4.5)	77 (14.0)	0.21
CABG	50 (8.7)	1 (4.5)	49 (8.9)	0.48
MVR	14 (2.4)	0 (0)	14 (2.5)	0.45
Femoral-femoral arterial bypass grafting	3 (0.5)	0 (0)	3 (0.5)	0.73
Carotid-axillary arterial bypass grafting	3 (0.5)	0 (0)	3 (0.5)	0.84
Axillary-axillary arterial bypass grafting	12 (2.1)	0 (0)	12 (2.2)	0.73
Ascending aorta-femoral arterial bypass grafting	13 (2.3)	1 (4.5)	12 (2.2)	0.47
Secondary surgery	21 (3.7)	3 (13.6)	18 (3.3)	0.01
Pancreatic amylase, U/L	48.2 (35.7, 65.8)	64.4 (37.925, 113.25)	47.8 (35.65, 64.575)	0.046
Pancreatic lipase, U/L	15.3 (9.5, 30.03)	35.85 (14.4, 53.4)	15.1 (9.2725, 27.7)	<0.001
WBC, ×109/L	9.925 (7.22, 13.5)	11.6 (7.6075, 14.48)	9.845 (7.21, 13.43)	0.27
PLT, ×109/L	194 (149, 238)	139.5 (91.25, 181)	195 (151, 238.25)	0.003
Hb, g/L	135 (116.25, 147)	118.5 (87, 142.25)	135 (117, 148)	0.044
Lymphocyte count, ×109/L	1.245 (0.83, 1.78)	0.88 (0.58, 1.1)	1.27 (0.85, 1.8)	0.004
Monocyte count, ×109/L	0.54 (0.38, 0.75)	0.555 (0.43, 0.85)	0.54 (0.38, 0.74)	0.61
Neutrophil count, ×109/L	7.985 (4.81, 11.61)	9.905 (5.33, 13.28)	7.9 (4.76, 11.47)	0.18
C-reactive protein, mg/L	25.67 (4.28, 85.30)	55.37 (23.62, 109.10)	24.805 (4.13, 84.47)	0.048
ALT, U/L	21 (14, 35)	20.5 (20.5, 49)	21 (14, 35)	0.99
AST, U/L	22 (17, 33.75)	21.5 (16, 39.25)	22 (17, 33.25)	0.79
Urea, mmol/L	6.36 (5.2, 8.4)	8.51 (5.9525, 10.5475)	6.31 (5.14, 8.3)	0.02
Creatinine, μmol/L	80.59 (65.53, 102.55)	87.59 (63.77, 117.53)	80.5 (65.57, 102.45)	0.54
Uric acid, μmol/L	365.4 (290.55, 459.3)	315.25 (174.6, 494.68)	365.85 (291.95, 458.9)	0.24
Glucose, mmol/L	6.68 (5.52, 8.10)	7.7 (5.83, 9.83)	6.63 (5.49, 8.06)	0.10
CK, U/L	95 (60, 191)	118 (75.25, 257)	95 (60, 190.25)	0.33
CK-MB, U/L	1.7 (0.9, 4.175)	5.25 (1.925, 10.9725)	1.6 (0.9, 3.8)	<0.001
LDH, U/L	227 (185, 312.5)	272.5 (205.75, 439.5)	227 (184.75, 309.5)	0.07
MB, ng/mL	42.2 (20.73, 129.48)	110.5 (56.05, 406.4)	40.245 (20.58, 124.45)	0.004

Data are presented as n (%) or median (interquartile range). ALT, alanine aminotransferase; AST, aspartate aminotransferase; ATAAD, acute type A aortic dissection; BMI, body mass index; CABG, coronary artery bypass grafting; CK, creatine kinase; CK-MB, creatine kinase-MB isoenzyme; CPB, cardiopulmonary bypass; Hb, hemoglobin; LDH, lactate dehydrogenase; MB, myoglobin; MVR, mitral valve replacement; PLT, platelet; WBC, white blood cell.

Feature variable selection

In this study, potentially relevant variables were included in the Lasso regression. The Lambda value corresponding to the minimum deviation point was selected as 0.011 (Figures 2,3). At this point, the model error was minimized and the predictive performance was optimal. Seven risk factors associated with the primary outcome were identified using the Lambda-min value: pancreatic amylase, pancreatic lipase, secondary surgery, left subclavian artery involvement, Sun’s procedure, Hb, and age. These predictive features were incorporated into the risk prediction model as key variables.

Figure 2.

Lasso regression coefficient distribution plot. Lasso, least absolute shrinkage and selection operator.

Figure 3.

Cross-validation deviation curve of Lasso regression. Lasso, least absolute shrinkage and selection operator.

Comparative analysis of multiple models

To distinguish the performance of various ML models, we generated receiver operating characteristic (ROC) curves, calibration curves, and decision curve analysis. Meanwhile, we calculated the confusion matrix of the validation set (Table 2). In the training set (Figure 4), the Random Forest model achieved the highest AUC of 1, with a 95% confidence interval (CI) of (1–1). In the validation set (Figure 5), the Neural Networks model performed the best, with an AUC value of 0.829 (95% CI: 0.684–0.974). After comprehensively evaluating indicators such as AUC, Kappa value, F1 score (Figure 6) and Brier score (Figure 7) in the confusion matrix, it was considered that the Random Forest model might have an overfitting phenomenon, while the Neural Networks model showed relatively better stability. On the calibration curve (Figure 8), the Neural Networks model exhibited higher calibration accuracy; the clinical decision curve (Figure 9) also indicated that the Neural Networks model had good clinical applicability. Its AUC (95% CI), accuracy, sensitivity, specificity, Brier score, F1 score, and Kappa value were 0.829 (0.684–0.974), 0.924, 0.333, 0.945, 0.059, 0.235, and 0.199, respectively.

Table 2. Confusion matrices of machine learning models.

Models	AUC (95% CI)	Accuracy	Sensitivity	Specificity	PPV	NPV	Brier	F1	Kappa
Logistic regression	0.816 (0.719–0.913)	0.930	0.000	0.964	0.000	0.964	0.060	0.000	−0.036
Random Forest	0.767 (0.581–0.952)	0.959	0.167	0.988	0.333	0.970	0.043	0.222	0.204
Neural Network	0.829 (0.684–0.974)	0.924	0.333	0.945	0.182	0.975	0.059	0.235	0.199
Decision Tree	0.460 (0.123–0.800)	0.924	0.167	0.952	0.111	0.969	0.071	0.133	0.095
XGboost	0.716 (0.530–0.903)	0.947	0.167	0.976	0.200	0.970	0.047	0.182	0.155
KNN	0.731 (0.513–0.949)	0.936	0.167	0.964	0.143	0.970	0.052	0.154	0.121
SVM	0.753 (0.443–1.000)	0.953	0.167	0.982	0.250	0.970	0.037	0.200	0.177

AUC, area under the curve; CI, confidence interval; KNN, K-Nearest Neighbor; NPV, negative predictive value; PPV, positive predictive value; SVM, Support Vector Machine.

Figure 4.

ROC curve of the training set. AUC, area under the curve; KNN, K-Nearest Neighbor; ROC, receiver operating characteristic; SVM, Support Vector Machine.

Figure 5.

ROC curve of the test set. AUC, area under the curve; KNN, K-Nearest Neighbor; ROC, receiver operating characteristic; SVM, Support Vector Machine.

Figure 6.

Bar chart of F1 score comparison on the test set. KNN, K-Nearest Neighbor; SVM, Support Vector Machine.

Figure 7.

Bar chart of Brier score comparison. KNN, K-Nearest Neighbor; SVM, Support Vector Machine.

Figure 8.

Calibration curve plot of all models. KNN, K-Nearest Neighbor; SVM, Support Vector Machine.

Figure 9.

Decision curve analysis plot of all models. KNN, K-Nearest Neighbor; SVM, Support Vector Machine.

Interpretability analysis

To visually illustrate the predictions of the Neural Networks model, we generated SHAP feature density scatter plots and feature importance plots to interpret the model (Figures 10,11).

Figure 10.

Bar chart of mean absolute SHAP values. Hb, hemoglobin; SHAP, SHapley Additive exPlanations.

Figure 11.

SHAP value swarm plot. Hb, hemoglobin; SHAP, SHapley Additive exPlanations.

The results in Figure 10 show that the variables influencing the model, ranked by importance from highest to lowest, are: pancreatic lipase, left subclavian artery involvement, Sun’s procedure, age, pancreatic amylase, Hb, and secondary surgery.

Discussion

ATAAD is an extremely life-threatening cardiovascular disease in cardiovascular surgery, characterized by abrupt onset and high mortality. At least 50% of ATAAD patients who do not undergo surgical treatment die from aortic rupture (10). Approximately 2% to 4% of ATAAD patients develop paraplegia after surgery, which manifests as impairment or loss of motor and sensory function at the corresponding spinal segments. Once paraplegia occurs, patients experience prolonged hospital stay and intensive care unit (ICU) length of stay (11), increased hospitalization costs, and significantly poor prognosis.

In this study, a Neural Networks ML risk prediction model was established for postoperative paraplegia in patients undergoing surgery for ATAAD, and SHAP analysis was applied to enhance model interpretability. Neural Networks are ML models designed to mimic biological nervous systems (e.g., the connection patterns of neurons in the human brain). Their core principle involves collaborative computation across multiple layers of “artificial neurons” to learn linear and nonlinear relationships in data, enabling tasks such as classification, regression, and pattern recognition (12). Compared with traditional logistic regression, Neural Networks offer advantages including robust nonlinear modeling capabilities and automatic feature extraction. In a study evaluating the performance of Neural Networks versus traditional logistic regression in predicting 30-day postoperative mortality in ATAAD patients, the Neural Networks model achieved higher overall predictive accuracy (13).

Age is an unavoidable factor when predicting surgical risk in ATAAD patients. Multiple studies have identified age as a key risk factor for poor prognosis in ATAAD patients (14-16). Compared with young patients, elderly patients are more likely to have underlying diseases such as cerebrovascular disease and coronary atherosclerotic heart disease, with poorer baseline health status and lower tolerance of the spinal cord to ischemia-hypoxia injury (17). Additionally, concomitant surgeries such as coronary artery bypass grafting (CABG) may be required during the operation, further prolonging the surgical duration (18). Additionally, elderly patients have longer postoperative awakening times (19), making early identification of paraplegic symptoms more difficult.

Hb serves as the oxygen carrier in human blood circulation. Low Hb levels reduce the oxygen-carrying capacity of red blood cells, leading to further spinal cord ischemia-hypoxia (20). According to relevant studies, the proportion of patients with Hb <120 g/L is significantly higher in the spinal cord ischemia group than in the non-ischemia group (21). Given the spinal cord’s poor tolerance to ischemia-hypoxia, it is recommended to maintain Hb levels above 10 g/dL during the perioperative period. This, combined with cerebrospinal fluid (CSF) drainage and a mean arterial pressure (MAP) ≥90 mmHg, is regarded as one of the three core oxygen supply indicators for spinal cord protection (22).

Sun’s procedure is currently the most commonly used surgical approach for managing aortic arch lesions. Its principle involves aortic arch replacement with distal frozen elephant trunk stent implantation under DHCA for cerebral protection (23). During surgery, ATAAD patients undergo a period of circulatory arrest, which may result in prolonged spinal cord hypoperfusion. Furthermore, after Sun’s procedure, deployment of the descending aortic stent rapidly restores blood flow in the true lumen of the distal aorta and accelerates thrombosis in the false lumen. Collateral blood supply to the spinal arteries in the affected segments cannot compensate promptly, leading to acute spinal cord ischemia and subsequent paraplegia (24). Relevant studies have shown that when the frozen elephant trunk stent covers ≥8 intercostal arteries, perfusion of the Adamkiewicz artery is reduced, increasing the risk of paraplegia (25).

Secondary surgery is a crucial rescue measure for patients with ATAAD who develop persistent active bleeding, cardiac tamponade, or refractory circulatory instability after initial surgery (26). Active bleeding can cause circulatory fluctuations in patients, leading to manifestations such as hypotension and anemia. This further reduces spinal cord perfusion pressure and exacerbates spinal cord ischemia-hypoxia. Performing secondary surgery within the “golden 6-hour window” after identifying these issues may significantly improve patient prognosis (27).

Postoperatively, close monitoring of patients’ neurological responses—including consciousness, limb movement, and pathological reflexes—is essential. Upon detecting paraplegic symptoms, urgent aortic CTA and spinal magnetic resonance imaging (MRI) should be performed to assess aortic blood flow and spinal cord ischemia, which helps determine patient prognosis. Acute management strategies should focus on maximizing spinal cord perfusion pressure: administering intravenous vasopressors to increase MAP and ensuring adequate circulating blood volume; performing CSF drainage to reduce CSF pressure; and providing symptomatic support such as glucocorticoids, mannitol, and neurotrophic agents. These measures help alleviate paraplegic symptoms and improve patient prognosis.

Currently, research focusing on the role of pancreatic enzymes in SCI remains limited. According to the few existing studies, the incidence of elevated pancreatic enzymes in patients with acute spinal cord injury (aSCI) is approximately 50% (28). Elevated pancreatic enzymes can reflect, to a certain extent, the ischemia of branch vessels of the descending aorta, thereby indicating the involvement of other branch vessels.

Furthermore, basic research has demonstrated that the elevation of pancreatic enzymes may trigger a systemic inflammatory cascade reaction (29). Aortic dissection patients who have already developed spinal cord ischemia are less tolerant of additional ischemic insults. Collectively, the elevation of pancreatic enzymes can serve as an indicator of multiple organ malperfusion.

Paraplegia following surgical intervention for ATAAD is not attributed to a single factor, but rather the synergistic effect of multiple factors including surgical approach, underlying diseases, and laboratory indicators. Due to the “independent variable assumption”, logistic regression fails to quantify the interaction effects between factors. In contrast, ML models can identify which factor combinations significantly amplify the risk through “feature interaction detection”.

The mechanism underlying this synergistic effect lies in the following aspects: Sun’s procedure requires prolonged circulatory arrest; elderly patients exhibit poor elasticity of spinal blood vessels; and hypertension further exacerbates spinal microvascular spasm—these three factors together induce a “vicious cycle” of spinal cord ischemia and hypoxia. Based on this perspective, clinical surgical strategies can be adjusted accordingly: for patients with “advanced age plus hypertension”, surgical approaches with shorter circulatory arrest time (e.g., Bentall procedure) should be prioritized instead of the routinely used Sun’s procedure.

The incidence of paraplegia after surgical treatment of ATAAD remains relatively low, and few studies have been conducted to analyze this specific issue. Compared with existing paraplegia prediction studies, this study adopts emerging research methodologies such as Lasso regression, cross-validation with grid search, and ML algorithms. These approaches effectively reduce the extreme conclusions that may be caused by sampling bias in previous studies, thereby providing better guidance for clinical practice.

Patients with ATAAD present with critical conditions, often requiring rapid assessment of surgical feasibility within a short timeframe. Consequently, variables influencing clinical outcomes must be accessible promptly before surgery. Basic patient information (such as age and history of hypertension) can be directly extracted through preoperative interviews and electronic medical record systems; preoperative emergency routine blood tests (with a turnaround time of within 30 minutes) can cover relevant laboratory indicators, including Hb, PLT, and LYM; and the treatment strategy is determined by the surgical team based on preoperative imaging examinations (aortic CTA).

By analyzing these variables and establishing a predictive model, this study can provide guiding recommendations for surgeons planning surgical intervention. Specifically, it assists in comprehensively determining whether surgery is indicated, and if so, whether a more aggressive or conservative surgical strategy should be adopted, as well as predicting the probability of postoperative paraplegia. These objectives represent the core intended outcomes of this study.

There are certain limitations in this study. First, the patient data collected in this study were all from a single center, which may lead to selection bias. In the future, multi-center data will be needed for external validation of the model to enhance its predictive performance. Additionally, the clinical data in this study were collected retrospectively, and the conclusions drawn therefore require further verification through prospective studies.

On the whole, this study developed a risk prediction model based on Neural Networks to assess the probability of postoperative paraplegia in patients with ATAAD. The model exhibits good predictive performance, enabling effective individualized evaluation of patients’ paraplegia risk and providing valuable guidance for clinicians to make clinical decisions that are most beneficial to patients’ prognosis. However, due to issues such as the low incidence of this disease, the model still requires further validation.

Conclusions

This study successfully established and validated seven ML-based risk prediction models for postoperative paraplegia following surgical treatment of ATAAD. Based on confusion matrix metrics including the AUC, calibration curves, decision curve analysis, F1 score, Brier score, and Kappa value, the Neural Networks model was selected as the optimal model.

Supplementary

The article’s supplementary files as

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. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Ethics Committee of Beijing Anzhen Hospital Affiliated to Capital Medical University (No. 2026013x). In view of the retrospective case-control design of this study, the requirement for individual informed consent from patients was waived.

Data Sharing Statement

Available at https://jtd.amegroups.com/article/view/10.21037/jtd-2025-2000/dss

jtd-18-02-102-dss.pdf ^{(68.3KB, pdf)}

DOI: 10.21037/jtd-2025-2000