Machine learning is broadly defined as a system’s ability to autonomously acquire knowledge by identifying patterns within large data sets [17]. Machine learning-based models perform continuous basic operations, and the complex and unpredictable nature of human physiology has been shown to be better captured by machine learning algorithms [18, 19]. This capability allows for the exploration of intricate correlation features, contributing to the increasing adoption of machine learning in the field of medicine [20, 21]. Significant advancements in machine learning have demonstrated substantial potential in disease diagnosis, complication monitoring, and prognosis prediction [22]. In this study, AD patients were divided into either the normal or prolonged LOS group according to binary classification (a threshold value of 30 days of LOS). The variables associated with prolonged LOS were screened and selected by the Boruta algorithm. Subsequently, these variables were used to develop machine learning-based models, including XGBoost, AdaBoost, KNN, logistic regression, LightGBM, GNB, MLP, CNB, and SVM. Our findings revealed that the XGBoost model had outstanding predictive performance in prolonged LOS risk in AD patients, evidenced by its highest AUC values in both the training and validation subsets. Additionally, other parameters of the model further supported this finding. Therefore, the XGBoost model was considered the final prediction model for prolonged LOS in AD patients.

The SHAP analysis can offer a novel approach to interpreting machine learning models, proving effective for both local and global interpretability. Furthermore, the SHAP provides a more pragmatic elucidation of the effectiveness of machine learning algorithms in delivering precise predictions for specific patient cohorts. In our study, machine learning algorithms-based models and the SHAP method was combined to improve the accuracy in predicting risk factors for prolonged LOS in AD patients. Moreover, the SHAP analysis showed intuitive explanations to assist clinicians in comprehensive understanding of the decision-making process for assessing disease severity and optimizing opportunities for early intervention [23]. In this study, the SHAP analysis for the XGBoost identified HDL-C, SBP, LymP, ALT, and OT as the top five variables affecting prolonged LOS in AD patients.

HDL-C can efficiently transport and remove excess cholesterol from peripheral tissues, thereby facilitating its return to the liver [24]. The study has reported that low levels of HDL-C increase the susceptibility to cardiovascular and cerebrovascular events in patients following cardiac intervention therapy [25]. A prospective cohort study involving 3 million participants revealed a ‘J-shaped’ association between HDL-C levels and cardiovascular disease mortality, where both high and low HDL-C levels were related with increased risk [26]. In this study, HDL-C levels in the prolonged LOS group were significantly lower than those in the normal LOS group. Although we did not observe a ‘J-shaped’ relationship between HDL-C levels and LOS in AD patients, our study identified that reduced HDL-C levels might contribute to prolonged LOS in these patients. Additionally, elevated HDL-C levels have been shown to reduce the risk of delirium and shorter LOS in patients following cardiac surgery [27]. Our findings of the SHAP analysis revealed that HDL-C was identified as the most crucial risk variable for prolonged LOS. These findings implied that HDL-C of patients should to be noticed by clinicians in clinical practice.

SBP is the highest blood pressure exerted on the active artery when the left ventricle contracts. A 2023 Mendelian randomization analysis demonstrated that elevated SBP was significantly associated with an increased risk of both aortic aneurysm and AD in patients [28]. Indeed, SBP can serve as a robust prognostic indicator for mortality in various cardiovascular conditions, including acute coronary syndrome, cardiogenic shock, and acute heart failure. Likewise, a study conducted by Bossone, E et al. revealed a logarithmic correlation between SBP and cardiovascular risk [29]. The inadequate control of preoperative SBP may elevate the susceptibility to perioperative nervous system complications [30]. Additionally, maintaining appropriate preoperative SBP is crucial for ensuring intraoperative hemodynamic stability, shortening OT, reducing the risk of stroke and postoperative delirium, as well as other neurologic complications and mortality [31]. Simultaneously, elevated SBP levels heighten susceptibility to arterial sclerosis, aneurysm formation, and acute aortic syndrome [32, 33]. Although there are limited studies that have included SBP as an independent indicator, the results of the SHAP analysis showed that elevated SBP levels were important predictors of prolonged LOS in patients with AD. These results might offer valuable insights for future blood pressure control programs.

LymP serves as a crucial indicator for monitoring immune function and reflects the level of inflammation in the body [34, 35]. Current studies in cardiovascular disease primarily focus on T lymphocytes, B lymphocytes, and NK cells. The key role of LymP in cardiovascular disease is undisputed, as it accelerates the pathological process through the secretion of proinflammatory cytokines. Equally well-established is the fact that autoimmune diseases, mediated by autoreactive T cells, significantly increase the risk of developing cardiovascular disease [36]. NLR has been utilized as a prognostic indicator for cardiovascular surgery, with numerous studies highlighting its correlation with the duration of postoperative hospitalization [37]. In this study, NLR and LymP exhibited statistical significance between the prolonged and normal LOS groups in AD patients. However, the SHAP analysis suggested that LymP might have a more prominent role than NLR. Previous studies found that lymphopenia may lead to elevated creatinine levels, prolonged mechanical ventilation duration, increased risk of arrhythmia following cardiac surgery, and extended LOS [38,39,40]. Consistently, our study identified LymP as a significant variable for predicting prolonged LOS in patients with AD. With the emergence of novel indicators, greater attention should be devoted to the role of LymP in clinical practice in the future.

ALT, an enzyme primarily synthesized by the liver, serves as a clinical indicator for hepatic injury and an independent determinant of cardiovascular disease [41, 42]. A recent study involving 2,565 patients found that elevated levels of ALT were significantly associated with prolonged LOS and increased mortality in individuals undergoing cardiac surgery [43]. Furthermore, higher preoperative ALT levels are correlated with increased drainage volume following aortic arch surgery, highlighting the significant prognostic relevance of ALT levels for patient outcomes [44]. As ALT levels increased in our study, AD patients were more likely to experience a prolonged LOS, suggesting that monitoring this index could provide valuable guidance for future clinical treatment strategies. Notably, the SHAP analysis identified ALT as an important risk factor, following LymP.

OT in patients undergoing cardiac surgery is determined by multiple factors, primarily involving disease severity, extracorporeal circulation time, and surgical technique. Numerous studies have demonstrated that OT closely correlates with the occurrence of postoperative complications, and increased OT may raise the risk of postoperative infection, mechanical ventilation, delirium, and prolonged LICU [45,46,47,48]. Indeed, the specific timing for defining aortic dissection surgery lacks relevant research. Our study employed a machine learning prediction model to effectively examine risk factors associated with prolonged LOS, identifying OT as a key contributing factor. Prolonged OT was detrimental to the recovery of patients with AD, indicating the necessity for medical practitioners to be mindful of this aspect in clinical practice. The prolongation of OT is associated with an elevated postoperative risk in open surgery, which hinders patients' recovery and leads to prolonged LOS and increased treatment expenses. Future treatment methods should prioritize tailoring the specific timing of operations based on individual conditions, thereby reducing OT, facilitating patient recovery, and minimizing the risk of postoperative complications.

AD patients undergoing surgery do not imply the completion of treatment program; instead, it means that they may be entering a period characterized by high risk and multifaceted challenges. These situations require us to be better prepared before surgery, and the lack of preoperative assessment of prolonged LOS may be a prominent issue. The application of AI-based machine learning in the medical field is important and necessary. It can not only help doctors improve their work efficiency but also provide more accurate diagnosis and treatment plans, thereby improving the quality of patients’ life. Our study aimed to identify preoperative risk variables associated with prolonged LOS in AD patients using machine learning-based models. With the help of machine learning AI tools, the XGBoost model was successfully simplified into easy-to-understand graphs, which enhanced its practicality for clinicians.

However, although the machine learning models exhibited favorable predictive performance in this study, it is imperative to acknowledge certain limitations. Firstly, the data of patients in this study were obtained solely from a single center, leading to some biases in the results. This might affect the development of the models and risk identification to some extent. Secondly, surgical manners and therapeutic strategies adopted by the doctors might affect the final prognosis. Although all the subjects in this study were diagnosed with AD, different surgical manners and therapeutic strategies might lead to different LOS. Finally, there was a lack of utilization of the external data for model validation, potentially impacting the generalizability of the models. The above limitations might restrict the applicability and generalizability of machine learning models in individuals diagnosed with AD. Factually, machine learning-based AI tools primarily rely on computer systems, thereby exhibiting characteristics of simplicity and operability. This feature facilitates convenience for scientific research or clinical work. The future plans for this study involve expanding the sample size and establishing multi-center cooperation, we hope that external data can be obtained and used for the external validation. This will enhance the reliability of the relevant conclusions. Simultaneously, machine learning will be employed to further investigate clinical practice, including complementary imaging data and clinical intervention, with the aim of providing clinicians with a more robust decision-making framework and improving medical care quality.