Acute aortic dissection (AAD) is characterized by a tear in the aortic intima, with symptoms manifesting within two weeks. Blood flows through this tear into the middle layer of the aorta, forming a true lumen and a false lumen, progressively separating the inner and middle layers of the aorta [1]. Currently, AAD is primarily classified into two types: Stanford type A, which involves the ascending aorta, and Stanford type B, which does not [2]. Clinically, AAD typically presents with acute, severe chest and back pain and is characterized by rapid onset, swift progression, diverse initial symptoms, and high mortality risk [3]. The incidence rate of AAD is about 0.005%, but the mortality rate within 24 hours can reach 33%. Without timely intervention, the mortality rate increases cumulatively by 0.5% each hour, reaching 50% within 48 hours and 74% within one week [4, 5]. The primary treatments for AAD include surgical repair and endovascular treatment, which have been demonstrated to achieve survival rates of up to 90% when timely administered [6]. Acute kidney injury (AKI) is a common complication among AAD patients, occurring either in-hospital or post-surgery, with an incidence rate of 7%–20% [7, 8]. The occurrence of AKI in AAD patients often exacerbates the condition, leading to further complications, prolonged hospital stays, and increased mortality rates [9]. Consequently, it is crucial to establish a robust predictive model for effectively forecasting AKI in hospitalized AAD patients.

Recently, the application of artificial intelligence in the medical field has become increasingly widespread. Machine learning (ML), an important branch of artificial intelligence [10, 11], delves deeper into the intrinsic patterns of data when faced with highly complex, high-dimensional clinical data compared to traditional prediction models. Prediction models developed using machine learning and now widely utilized in clinical predictions exhibit greater stability, higher accuracy, and stronger generalization capabilities [12, 13]. The main ML types include supervised learning, unsupervised learning, and others [14]. The Medical Information Mart for Intensive Care (MIMIC)-IV database is a commonly used large single-center database that contains the clinical data of 382,278 patients at Beth Israel Deaconess Medical Center from 2008 to 2019. The data include demographic characteristics, radiological examination results, laboratory test results, patient vital signs, medication treatment data, various scoring data, in-hospital complications, and clinical outcomes [15, 16]. The MIMIC database is widely used due to its high-quality and comprehensive data records. Previous literature has established a predictive model for in-hospital mortality of AAD patients [17, 18] and reports of predictive models for AKI following acute myocardial infarction [19]; however, there is currently no research on machine learning models related to in-hospital AKI complications in AAD patients. Therefore, this study extracted clinical data of AAD patients from the MIMIC-IV database to establish seven types of machine learning models: support vector machine (SVM), gradient boosting machine (GBM), neural network (NNET), eXtreme gradient boosting (XGBoost), K-nearest neighbors (KNN), light gradient boosting machine (LightGBM), and categorical boosting (CatBoost). Simultaneously, the AAD patient data in the MIMIC-III database were used to externally validate the established optimal model. These machine learning models are employed to screen for risk factors and predict in-hospital AKI complications in AAD patients, aiding clinical decision-making.

This study primarily identified the risk factors associated with AKI complications in hospitalized AAD patients. Through univariate and multivariate logistic regression analyses, clinical features including maximum BUN, BMI, urine output, maximum GLU, minimum BUN, minimum creatinine, maximum creatinine, weight, and APSIII were found to be associated with AKI occurrence during hospitalization in AAD patients. Seven machine learning models were also developed: SVM, GBM, NNET, XGBoost, KNN, LightGBM, and CatBoost. Each model exhibited unique characteristics, while performances varied across different datasets.

In this study, the CatBoost model demonstrated superior performance both in the training set (AUC = 0.876), test set (AUC = 0.723), and validation set (AUC = 0.712) compared to other models. The advantages of the CatBoost model are significant, as it can reduce prediction bias through ordered boosting and unbiased gradient estimation to combat overfitting while using diverse sampling methods to enhance both precision and accuracy, thereby enhancing the model’s generalizability. SHAP visualization analysis was employed to interpret the optimal machine learning model. The incidence of in-hospital AKI in AAD patients was 25.88%, which is consistent with previous studies [7]. Development of in-hospital AKI in AAD patients is associated with worsened outcomes and a poorer prognosis [21, 22, 23]. Therefore, the timely identification of risk factors for in-hospital AKI in AAD patients and the development of effective machine learning models are crucial for identifying high-risk patients and providing timely clinical intervention to prevent further complications.

In 2023, Dai A et al. [24] selected risk factors such as urine output, intraoperative hypotension, and autologous blood transfusion to establish four machine learning models, including XGBoost and SVM, to predict postoperative AKI risk in patients with AAD. However, limitations included the use of fewer models without incorporating the most recent predictive models, as well as missing external validation. In 2022, Luo CC et al. [25] found that variables, such as creatinine levels and extracorporeal circulation time, were closely associated with the occurrence of postoperative AKI in AAD patients; however, the researchers solely developed a nomogram, which showed lower predictive efficacy. In 2023, Zhang C et al. [26] selected risk factors such as hypertension and preoperative renal artery involvement and established a predictive model for in-hospital AKI in postoperative Stanford type A AAD patients, with an AUC of 0.839. However, this study had a small sample size of only 241 cases and utilized a single, simplistic model for prediction without further elaboration. Thus, the reliability of the overall research could not be guaranteed. Previous studies have not established a reliable predictive model for in-hospital AKI in Stanford type A AAD patients, thereby motivating our attempt to develop a more stable model. Utilizing SHAP visualization analysis, we interpreted the optimal CatBoost model, identifying key factors associated with in-hospital AKI, including BUN, BMI, urine output, creatinine, APSIII, etc. In AAD patients, cumulative kidney involvement often leads to renal hypoperfusion, resulting in renal impairment, decreased glomerular filtration rate (GFR), and increased renal reabsorption of water, thereby reducing urine output. Abnormal urine output in AAD patients is indicative of a higher likelihood of developing in-hospital AKI [27, 28]. BUN levels are often increased following the use of nephrotoxic drugs, potentially exacerbating kidney involvement and increasing AKI risk in AAD patients [29, 30]. Creatinine, a metabolic product of phosphocreatine and creatine in muscle tissue, is primarily filtered by the glomeruli into the urine; therefore, elevated creatinine levels often indicate impaired renal function [31, 32]. Obesity is a risk factor for various diseases, such as hypertension and hyperlipidemia [33]. Study has demonstrated that an increase of 5 kg/m² in BMI increases the incidence of AKI by 40% [34], highlighting weight as an important factor in AKI occurrence. The APSIII, a severity-of-disease classification system, is commonly utilized in prognosis studies of respiratory and neurological diseases [35, 36], yet its application in AKI complications still needs to be explored. This study provides insights into the potential clinical utility of the APSIII in predicting in-hospital AKI complications in AAD patients. Based on our findings, it is recommended that clinicians actively prevent in-hospital AKI when there are notable increases in indicators such as BUN, creatinine levels, urine output, GLU, and APSIII in AAD patients or if the patient is obese. This can be performed through medication and symptomatic supportive treatment to manage elevated BUN, creatinine, and GLU levels.

This study has several limitations: (1) the number of patients included, although retrieved under both the ICD-9 and ICD-10 coding systems for AAD diagnosis, was still insufficient, potentially leading to sampling errors and probabilistic biases; (2) our machine learning model relied on a single-center database, the MIMIC, which despite its high quality, may still contain issues such as missing data and errors; (3) the machine learning model focused exclusively on predicting in-hospital AKI in AAD patients, necessitating further research into renal complications post-discharge; (4) the incidence rate of AKI in this study data was only 25.85%, resulting in data imbalance. While oversampling techniques were employed to address these limitations, they may still compromise the effectiveness of the model.

We developed multiple machine learning models using data from the MIMIC-III and MIMIC-IV databases to predict in-hospital AKI in AAD patients. The CatBoost model exhibited superior performance, highlighting its potential clinical implications. This study identified several factors associated with the occurrence of in-hospital AKI in AAD patients, including maximum BUN, BMI, urine output, maximum GLU, minimum BUN, minimum creatinine, maximum creatinine, weight, and APSIII.

The data utilized in this study are accessible via the following online database link: https://physionet.org/content/mimiciv/2.2/, last accessed date on 1 June 2024.

ZW, YC, SL, GL and HL analyzed the data and wrote the manuscript. GL and HL collected the data. YC and YH checked the integrity of the data and processed it. BS and YH designed the experiment and made critical revisions to the manuscript. All authors read and approved the manuscript. All authors have participated sufficiently in the work and agreed to be accountable for all aspects of the work.

Not applicable.

This study was supported by the Natural Science Foundation of Gansu Province (21JR7RA379).

All authors declare that they have no conflicts of interest related to the topics discussed in this manuscript.

Supplementary material associated with this article can be found, in the online version, at https://doi.org/10.31083/RCM25768.