Development and validation of a predictive model for postoperative acute respiratory distress syndrome in patients with type A aortic dissection based on the 2023 updated definition Authors: ChengBin Tang, Tianwei Wang, Haiqing Diao, Lulu Zhou, Haoran Wang, Tingting Yu, Jichao Zhai, Aipeng Hu, Jing Yuan, Jing Hang, Hailong Yu, Yuping Li, Ruiqiang Zheng, Jun Shao Published: 2025 ABSTRACT BackgroundAcute respiratory distress syndrome (ARDS) is a common complication after type A aortic dissection surgery and often leads to worsened clinical outcomes for patients. The early prediction of postoperative ARDS is a crucial challenge in clinical practice; however, there have been few reports on related studies based on the 2023 global new definition.MethodsA retrospective analysis was conducted on the clinical data of 423 patients who were diagnosed with type A aortic dissection and who underwent surgery at Northern Jiangsu People’s Hospital in Jiangsu Province from November 2019 to April 2025. A 7:3 random division was applied to the patients, resulting in a training set n = 296 and a validation set n = 127. Risk factors were identified via LASSO analysis, and a comprehensive risk prediction model was subsequently constructed by integrating five machine learning algorithms. The receiver operating characteristic (ROC) curve was utilised, and the model with the best predictive performance was selected based on the area under the curve (AUC).ResultsAmong the 423 included patients, 192 developed ARDS, with an incidence rate of 45.39%. LASSO analysis revealed 13 risk factors. Among the five machine learning models constructed based on these factors, the random forest model demonstrated the highest prediction efficiency for ARDS (AUC = 0.978), followed by the logistic regression (AUC = 0.965), decision tree (AUC = 0.881), support vector machine (AUC = 0.835), and K-nearest neighbour (AUC = 0.807) models.ConclusionThe development of a nomogram model using machine learning algorithms for predicting ARDS risk in patients with type A aortic dissection after surgery could identifying high-risk patients at an early stage and enable timely implementations of preventive strategies.Trial registrationThe medical research ethics committee of the Northern Jiangsu People’s Hospital provided approval for this study (ethics number: 2024ky314). This study is registered in the Chinese Clinical Trial Registry under registration number ChiCTR2500099730.The registration date was March 27,2025.Supplementary InformationThe online version contains supplementary material available at 10.1186/s12931-025-03415-2. Introduction Stanford type A aortic dissection (AAD) is a serious cardiovascular surgical disease in which a vicious cycle is elicited due to the interaction between structural abnormalities of the aortic media itself and haemodynamic abnormalities [1]. Acute respiratory distress syndrome (ARDS) is a pulmonary disease characterised by intractable hypoxemia that seriously threatens the lives of patients [2, 3]; moreover, it is a common complication occurring after aortic dissection surgery [4, 5]. ARDS is associated with a significant increase in postoperative mortality. Moreover, it strongly affects prognosis, prolongs the durations of ventilator use and postoperative intensive care unit (ICU) stay, and may induce multiple organ failure, thereby potentially leading to mortality [6–8]. Therefore, an effective method for predicting the occurrence of postoperative ARDS in patients with AAD at an early stage is urgently needed to guide clinicians in early diagnosis and treatment, thereby improving patient prognosis. Earlier research has confirmed that a (BMI) ≥ 25 kg/m², a time interval from symptom onset to surgery ≥ 24 h, a preoperative white blood cell count ≥ 10 × 10⁹/L, and a preoperative ratio of partial pressure of arterial oxygen to fraction of inspired oxygen (PaO₂/FiO₂) < 300 mmHg act as independent risk factors for ARDS [9, 10]. However, a large amount of complex and multidimensional clinical data can be obtained. Moreover, the sole identification of independent risk factors provides limited guidance for clinicians. Machine learning is a new type of artificial intelligence; specifically, this technology demonstrates excellent automated learning and predictive capabilities and can handle large clinical datasets in a more effective manner [11]. Therefore, various machine learning algorithms have been applied to the development of ARDS prediction models, such as XGBoost, random forest, and support vector machine models. These algorithms can process substantial clinical and imaging data and mine the potential patterns within these data, thereby enabling the early prediction of ARDS [12]. A recent study revealed that a nomogram model constructed on the basis of machine learning could effectively predict the occurrence of ARDS after AAD, with an ROC value of 0.809 [13]; however, the small sample size of this study may have resulted in considerable bias. Previous studies have examined the risk factors for ARDS after AAD surgery; however, all of these studies have relied solely on the “Berlin Definition” published in 2012. Advancements in clinical practice and scientific research have increasingly highlighted the deficiencies of the “Berlin Standard”. In 2023, a new global definition of ARDS was released, thereby overcoming the limitations of the “Berlin Standard”. However, no studies have focused on a predictive model for postoperative ARDS in patients with AAD based on the newly established definition. Thus, the aim of this study was to establish and validate a predictive model for the occurrence of ARDS following AAD surgery through analyses of perioperative clinical indicators and the use of five machine learning techniques. Methods Participants This was a single-centre retrospective study. All patients diagnosed with AAD who underwent surgery between November 2019 and April 2025 were included in the study. All of the patients were treated at the Cardiac and Cardiovascular Disease Center of Northern Jiangsu People’s Hospital in Jiangsu Province. According to whether the patients met the diagnostic criteria of the 2023 global new definition of ARDS [14] after the operation, the patients were divided into the ARDS group or the non-ARDS group. The inclusion criteria were as follows: (1) aged ≥ 18 years and (2) diagnosed with AAD and undergoing surgical treatment. The exclusion criteria were as follows: (1) hospitalised or survived in the ICU for no more than 24 h; (2) severe heart failure (NYHA grade ≥ III); (3) cardiogenic pulmonary oedema; (4) extensive atelectasis; (5) pulmonary arterial hypertension; and (6) >30% missing clinical data. Diagnostic criteria of ARDS Patients meeting the following four criteria was diagnosed with ARDS: (1) acute or worsening respiratory symptoms occurring within one week of a known clinical insult; (2) bilateral opacities on chest X-ray or CT not fully explained by effusion, lobar/lung collapse or nodules; (3)respiratory failure not fully explained by cardiac failure or fluid overload; (4)hypoxemia: PaO₂/FiO₂ ≤ 300 mmHg with PEEP ≥ 5 cmH₂O (or CPAP); or, if on high-flow nasal oxygen (HFNO) > 30 L/min or NIV/CPAP with at least 5 cmH₂O end-expiratory pressure, PaO₂/FiO₂ < 300 mmHg or SpO₂/FiO₂ < 315 (when SpO₂ < 97%). Data collection Patient data, including age, sex, height, weight, BMI, smoking history, hypertension history, diabetes history, and admission vital signs, were subjected to statistical analysis. After admission, the patients’ imaging examination results were recorded, the presence of pleural and pericardial effusion in the images was noted, and pulmonary exudation was quantitatively evaluated by using CT scans. Statistical analysis of various laboratory indicators of patients obtained during the perioperative period, such as blood cell (white blood cell count, haemoglobin, levels, etc.), liver function (alanine aminotransferase levels, aspartate aminotransferase levels, etc.), renal function (uric acid levels, creatinine levels, etc.), coagulation function (prothrombin time, activated partial thromboplastin time, etc.), blood gas (PaO2, FiO2, etc.), and surgery-related data (CPB time, intraoperative blood loss, intraoperative blood transfusion volume, etc.), was performed. Data processing If the proportion of missing data is less than 10% of the total number of values, the mean imputation method is used. If the proportion of missing data is greater than 10% but less than 20%, multiple imputation is applied. If the proportion of missing data exceeds 20%, the variable is deleted. Statistical analysis In this study, Stata 17.0 and R 4.3.2 software were used for data collation and statistical analysis. The Kolmogorov-Smirnov test was used to assess the goodness-of-fit of the continuous variables to a normal distribution. Normally distributed data are presented as the mean ± standard deviation (x ± s), and an independent sample t test was used to compare the mean values between the two groups. Those data with a skewed distribution were described as [M (P25, P75)], and comparisons between the groups were performed via the Mann-Whitney U test. Categorical variables are described as [n (%)], and the Fisher’s exact test or chi-square test was used for comparisons between the groups. The training set and validation set are randomly divided in a 7:3 ratio using the random seed method, with the seed set to 42 to ensure reproducibility. The training set was used to construct the model, whereas the validation set was used to evaluate the model’s performance. The least absolute shrinkage and selection operator (LASSO) regression method was adopted to screen meaningful predictors. Prior to LASSO regression, all continuous variables were standardized using Z-score normalization. Multicollinearity was assessed through Variance Inflation Factor (VIF) evaluation, with all variables included in the LASSO model showing VIF values < 5, indicating minimal multicollinearity. The model selection process employed 10-fold cross-validation to identify optimal λ values, ultimately selecting 13 non-zero coefficient variables for modeling.Logistic regression analysis and algorithms such as the decision tree (DT), K-nearest neighbour (KNN), random forest (RF), and support vector machine (SVM) were utilised to develop ARDS risk prediction models on the basis of the training dataset. A receiver operating characteristic (ROC) curve was constructed, and the predictive efficacies of the risk prediction models for ARDS were evaluated through the area under the curve (AUC), as demonstrated in a previous study in which the model performance was assessed by using ROC analysis and the AUC. The model with the highest predictive efficiency was selected as the optimal model for this study. Calibration curves were used to determine the calibration capability of the prediction model.The clinical practicability and net profit of the predictive model were evaluated by Decision Curve Analysis (DCA). Results Demographic and clinical parameters for the training and validation sets As illustrated in Fig. 1 462 patients diagnosed with AAD, a condition characterised by clinical symptoms such as abnormal body temperature, increased heart rate, rapid breathing, and abnormal white blood cell counts, as well as laboratory indicators such as positive blood cultures, were identified. Among the identified patients, 11 patients were excluded because their ICU stay or survival time did not exceed 24 h; 12 patients were excluded because of severe heart failure (NYHA classification ≥ III grade); and 16 patients were excluded due to > 30% missing clinical data. A total of 423 patients were ultimately included for analysis, with the data of 296 and 127 patients included in the training and validation sets, respectively.Fig. 1Flowchart of patient inclusion and exclusion. AAD, A aortic dissection Flowchart of patient inclusion and exclusion. AAD, A aortic dissection As shown in Tables 1 and 2, and Table 3, statistical analysis revealed significant differences in BMI, blood transfusion volume, and ARDS incidence between the two groups (P < 0.05), thereby indicating that these variables may be important factors affecting the observed outcomes. Table 1Comparison of clinical data between the two groups of patientsVariablesTotal (n = 423)TrainingSet (n = 296)ValidationSet (n = 127)Statistic P Age, years58.00 (48.00, 69.00)58.00 (47.75, 69.00)59.00 (48.00, 68.50)Z=−0.300.762Sex, n (%)χ²=1.310.253 Female130 (30.73)86 (29.05)44 (34.65) Male293 (69.27)210 (70.95)83 (65.35)BMI, kg/m224.80 (22.82, 27.68)25.11 (22.98, 27.68)24.16 (21.34, 27.61)Z=−2.070.038 *Sbp, mmHg133.00 (114.00, 152.00)131.50 (112.00, 152.00)135.00 (120.00, 152.50)Z=−1.480.138Dbp, mmHg72.00 (60.00, 86.00)71.00 (59.00, 86.00)74.00 (64.00, 86.00)Z=−0.660.509Map, mmHg92.00 (78.83, 106.83)92.00 (77.50, 106.67)92.67 (82.00, 108.00)Z=−1.150.252Shock index0.57 (0.47, 0.70)0.58 (0.49, 0.71)0.56 (0.46, 0.69)Z=−1.410.158Apache II score,8.00 (6.00, 11.00)8.00 (6.00, 11.00)8.00 (6.00, 11.00)Z=−0.510.612Hospital stay, days20.00 (16.00, 25.00)20.00 (16.00, 26.00)19.00 (17.00, 24.00)Z=−0.160.872ICU stay, days5.91 (4.47, 9.22)5.90 (4.37, 9.10)5.93 (4.53, 9.33)Z=−0.110.915Mechanical ventilation time, days3.46 (1.38, 5.67)3.66 (1.29, 5.64)3.10 (1.42, 5.74)Z=−0.040.969Secondary intubation, n (%)χ²=0.000.997 NO383 (90.54)268 (90.54)115 (90.55) YES40 (9.46)28 (9.46)12 (9.45)ARDS Ao, n (%)χ²=4.220.040 * NO231 (54.61)152 (51.35)79 (62.20) YES192 (45.39)144 (48.65)48 (37.80)Shock, n (%)χ²=1.360.243 NO360 (85.11)248 (83.78)112 (88.19) YES63 (14.89)48 (16.22)15 (11.81)Hypertension, n (%)χ²=0.050.819 NO120 (28.37)83 (28.04)37 (29.13) YES303 (71.63)213 (71.96)90 (70.87)Diabetes, n (%)χ²=0.030.866 NO407 (96.22)284 (95.95)123 (96.85) YES16 (3.78)12 (4.05)4 (3.15)Coronary heart disease, n (%)χ²=0.001.000 NO414 (97.87)290 (97.97)124 (97.64) YES9 (2.13)6 (2.03)3 (2.36)Stroke, n (%)χ²=0.000.998 NO403 (95.27)282 (95.27)121 (95.28) YES20 (4.73)14 (4.73)6 (4.72)Smoking, n (%)χ²=0.000.991 NO230 (54.37)161 (54.39)69 (54.33) YES193 (45.63)135 (45.61)58 (45.67)Drinking, n (%)χ²=0.310.579 NO255 (60.28)181 (61.15)74 (58.27) YES168 (39.72)115 (38.85)53 (41.73)BMI body mass index, Sbp systolic blood pressure, Dbp diastolic blood pressure, MAP mean arterial pressure, Apache II acute physiology and chronic health evaluation, Ao after operation Table 1Comparison of clinical data between the two groups of patientsVariablesTotal (n = 423)TrainingSet (n = 296)ValidationSet (n = 127)Statistic P Age, years58.00 (48.00, 69.00)58.00 (47.75, 69.00)59.00 (48.00, 68.50)Z=−0.300.762Sex, n (%)χ²=1.310.253 Female130 (30.73)86 (29.05)44 (34.65) Male293 (69.27)210 (70.95)83 (65.35)BMI, kg/m224.80 (22.82, 27.68)25.11 (22.98, 27.68)24.16 (21.34, 27.61)Z=−2.070.038 *Sbp, mmHg133.00 (114.00, 152.00)131.50 (112.00, 152.00)135.00 (120.00, 152.50)Z=−1.480.138Dbp, mmHg72.00 (60.00, 86.00)71.00 (59.00, 86.00)74.00 (64.00, 86.00)Z=−0.660.509Map, mmHg92.00 (78.83, 106.83)92.00 (77.50, 106.67)92.67 (82.00, 108.00)Z=−1.150.252Shock index0.57 (0.47, 0.70)0.58 (0.49, 0.71)0.56 (0.46, 0.69)Z=−1.410.158Apache II score,8.00 (6.00, 11.00)8.00 (6.00, 11.00)8.00 (6.00, 11.00)Z=−0.510.612Hospital stay, days20.00 (16.00, 25.00)20.00 (16.00, 26.00)19.00 (17.00, 24.00)Z=−0.160.872ICU stay, days5.91 (4.47, 9.22)5.90 (4.37, 9.10)5.93 (4.53, 9.33)Z=−0.110.915Mechanical ventilation time, days3.46 (1.38, 5.67)3.66 (1.29, 5.64)3.10 (1.42, 5.74)Z=−0.040.969Secondary intubation, n (%)χ²=0.000.997 NO383 (90.54)268 (90.54)115 (90.55) YES40 (9.46)28 (9.46)12 (9.45)ARDS Ao, n (%)χ²=4.220.040 * NO231 (54.61)152 (51.35)79 (62.20) YES192 (45.39)144 (48.65)48 (37.80)Shock, n (%)χ²=1.360.243 NO360 (85.11)248 (83.78)112 (88.19) YES63 (14.89)48 (16.22)15 (11.81)Hypertension, n (%)χ²=0.050.819 NO120 (28.37)83 (28.04)37 (29.13) YES303 (71.63)213 (71.96)90 (70.87)Diabetes, n (%)χ²=0.030.866 NO407 (96.22)284 (95.95)123 (96.85) YES16 (3.78)12 (4.05)4 (3.15)Coronary heart disease, n (%)χ²=0.001.000 NO414 (97.87)290 (97.97)124 (97.64) YES9 (2.13)6 (2.03)3 (2.36)Stroke, n (%)χ²=0.000.998 NO403 (95.27)282 (95.27)121 (95.28) YES20 (4.73)14 (4.73)6 (4.72)Smoking, n (%)χ²=0.000.991 NO230 (54.37)161 (54.39)69 (54.33) YES193 (45.63)135 (45.61)58 (45.67)Drinking, n (%)χ²=0.310.579 NO255 (60.28)181 (61.15)74 (58.27) YES168 (39.72)115 (38.85)53 (41.73)BMI body mass index, Sbp systolic blood pressure, Dbp diastolic blood pressure, MAP mean arterial pressure, Apache II acute physiology and chronic health evaluation, Ao after operation Comparison of clinical data between the two groups of patients BMI body mass index, Sbp systolic blood pressure, Dbp diastolic blood pressure, MAP mean arterial pressure, Apache II acute physiology and chronic health evaluation, Ao after operation Table 2Comparison of laboratory tests and imaging data between the two groups of patientsVariablesTotal (n = 423)TrainingSet (n = 296)ValidationSet (n = 127)Statistic P PCO2 Bo, mmHg41.78 (37.25, 45.00)41.78 (37.72, 45.05)41.78 (36.95, 45.00)Z=−0.100.923PO2 Bo, mmHg172.70 (103.00, 221.00)172.85 (101.35, 220.05)171.00 (114.60, 225.35)Z=−0.360.721FiO2 Bo0.60 (0.41, 0.61)0.60 (0.41, 0.61)0.60 (0.41, 0.80)Z=−0.090.930PH Bo7.36 (7.33, 7.40)7.35 (7.32, 7.40)7.36 (7.33, 7.40)Z=−0.270.783WBC Bo, 109/L12.21 (9.07, 14.20)12.21 (9.06, 13.63)12.21 (9.07, 14.75)Z=−0.880.380Ne Bo, 109/L10.55 (7.36, 12.39)10.55 (7.37, 12.02)10.55 (6.86, 13.04)Z=−1.210.225Ne percent Bo, %0.87 (0.81, 0.89)0.87 (0.81, 0.89)0.87 (0.82, 0.89)Z=−0.640.523Hb Bo, g/L131.00 (110.50, 151.50)132.50 (110.00, 153.00)129.00 (111.50, 149.50)Z=−1.070.284PLT Bo, 109/L152.66 (114.00, 177.00)152.66 (115.00, 178.50)152.66 (108.50, 171.50)Z=−0.350.726Pt Bo, s13.20 (12.10, 13.70)13.10 (12.10, 13.70)13.20 (12.10, 13.70)Z=−0.510.610Aptt Bo, s30.00 (27.10, 32.92)29.95 (27.08, 32.92)30.20 (27.10, 34.20)Z=−0.750.451INR Bo1.14 (1.05, 1.20)1.14 (1.05, 1.20)1.15 (1.05, 1.21)Z=−0.850.396Alb Bo, g/L37.03 (33.50, 40.80)37.03 (33.70, 40.82)37.03 (33.15, 40.30)Z=−0.490.625Alo Bo, g/L25.86 (23.30, 28.15)25.86 (23.30, 28.13)25.86 (23.25, 28.15)Z=−0.090.931Tbil Bo, µmol/L18.30 (11.90, 22.20)18.10 (12.00, 21.87)18.86 (11.85, 22.45)Z=−0.080.939Idil Bo, µmol/L16.80 (10.20, 20.00)16.85 (10.35, 20.10)16.70 (9.80, 19.75)Z=−0.380.702Dbil Bo, µmol/L0.00 (0.00, 1.01)0.00 (0.00, 1.01)0.00 (0.00, 1.01)Z=−0.220.825ALT Bo, U/L27.00 (19.00, 48.65)27.00 (19.00, 48.65)26.00 (18.50, 48.65)Z=−0.460.644AST Bo, U/L34.00 (25.00, 73.65)35.50 (25.88, 71.66)33.00 (25.00, 73.65)Z=−0.130.894UA Bo, µmol/L363.47 (298.15, 418.50)363.47 (293.85, 420.55)363.47 (303.10, 411.50)Z=−0.390.696Cr Bo, µmol/L88.70 (67.50, 105.20)88.65 (67.50, 105.00)89.90 (67.45, 106.20)Z = 0.001.000PCO2 Ao, mmHg42.40 (37.10, 47.15)42.55 (37.38, 47.40)41.90 (36.85, 46.35)Z=−1.520.127PO2 Ao, mmHg152.30 (108.15, 195.30)151.15 (106.47, 191.13)154.80 (123.00, 204.10)Z=−0.990.324FiO2 Ao, mmHg0.60 (0.50, 0.97)0.60 (0.50, 0.91)0.60 (0.50, 1.00)Z=−0.650.514PH Ao, mmHg7.38 (7.34, 7.43)7.38 (7.33, 7.42)7.39 (7.35, 7.43)Z=−1.390.164Lac Ao, mmol/L2.30 (1.50, 3.70)2.20 (1.50, 3.50)2.50 (1.50, 4.05)Z=−1.580.113WBC Ao, 109/L9.92 (7.61, 12.54)9.92 (7.64, 12.16)9.92 (7.55, 13.09)Z=−0.310.756Ne Ao, 109/L8.77 (6.72, 11.41)8.71 (6.75, 11.09)8.88 (6.65, 11.97)Z=−0.590.558Ne percent Ao0.89 (0.86, 0.92)0.89 (0.86, 0.92)0.89 (0.85, 0.92)Z=−0.600.550Hb Ao, g/L96.00 (86.00, 106.00)96.07 (87.00, 106.00)94.00 (84.00, 104.50)Z=−1.000.318PLT Ao, 109/L95.00 (66.50, 122.00)94.50 (69.00, 122.00)96.00 (62.00, 122.50)Z=−0.170.867Pt Ao, s14.10 (13.20, 15.20)14.00 (13.20, 15.12)14.30 (13.25, 15.35)Z=−1.240.216Aptt Ao, s33.00 (29.30, 38.75)32.50 (28.80, 38.33)34.50 (30.00, 39.10)Z=−1.410.158INRAo1.22 (1.15, 1.32)1.22 (1.15, 1.31)1.24 (1.15, 1.33)Z=−0.770.440Alb Ao, g/L33.10 (29.95, 36.70)33.00 (30.10, 36.70)33.90 (29.70, 36.55)Z=−0.380.707Glo Ao, g/L21.60 (19.40, 23.80)21.60 (19.40, 23.73)21.70 (19.35, 23.80)Z=−0.040.972Tbil Ao, µmol/L23.30 (17.40, 31.40)23.85 (18.05, 31.13)22.40 (16.50, 33.35)Z=−0.410.682Idil Ao, µmol/L19.00 (12.95, 28.00)19.05 (14.07, 27.85)18.10 (11.60, 28.35)Z=−0.960.336Dbil Ao, µmol/L0.00 (0.00, 5.90)0.00 (0.00, 5.43)0.00 (0.00, 7.50)Z=−0.040.966ALT Ao, U/L23.00 (17.00, 39.00)23.00 (17.00, 39.00)22.00 (16.00, 37.50)Z=−0.870.383AST Ao, U/L49.00 (37.00, 79.50)49.00 (37.72, 73.25)48.00 (35.45, 82.50)Z=−0.480.635Ua Ao, µmol/L315.40 (246.30, 379.25)315.70 (248.25, 379.95)315.10 (246.30, 377.35)Z=−0.190.848Cr Ao, µmol/L90.00 (67.95, 124.15)89.00 (67.98, 119.85)90.90 (68.00, 130.30)Z=−0.630.531Pulmonary exudation, n (%)χ²=0.170.677 NO101 (23.88)69 (23.31)32 (25.20) YES322 (76.12)227 (76.69)95 (74.80)Hydrothorax, n (%)χ²=0.000.958 NO184 (43.50)129 (43.58)55 (43.31) YES239 (56.50)167 (56.42)72 (56.69)Hydropericardium, n (%)χ²=0.220.637 NO389 (91.96)271 (91.55)118 (92.91) YES34 (8.04)25 (8.45)9 (7.09)PCO2 Partial Pressure of Carbon Dioxide, PO2 Partial Pressure of Oxygen, FiO2 Fraction of Inspired Oxygen, WBC White Blood Cell Count, Ne Neutrophils, Hb Hemoglobin, PLT Platelet Count, PT Prothrombin Time, APTT Activated Partial Thromboplastin Time, INR International Normalized Ratio, Alb Albumin, Tbil Total Bilirubin, Idil Indirect Bilirubin, Dbil Direct Bilirubin, ALT Alanine Aminotransferase, AST Aspartate Aminotransferase, UA Uric Acid, Cr Serum Creatinine, Ao after operation, Bo before operation Table 2Comparison of laboratory tests and imaging data between the two groups of patientsVariablesTotal (n = 423)TrainingSet (n = 296)ValidationSet (n = 127)Statistic P PCO2 Bo, mmHg41.78 (37.25, 45.00)41.78 (37.72, 45.05)41.78 (36.95, 45.00)Z=−0.100.923PO2 Bo, mmHg172.70 (103.00, 221.00)172.85 (101.35, 220.05)171.00 (114.60, 225.35)Z=−0.360.721FiO2 Bo0.60 (0.41, 0.61)0.60 (0.41, 0.61)0.60 (0.41, 0.80)Z=−0.090.930PH Bo7.36 (7.33, 7.40)7.35 (7.32, 7.40)7.36 (7.33, 7.40)Z=−0.270.783WBC Bo, 109/L12.21 (9.07, 14.20)12.21 (9.06, 13.63)12.21 (9.07, 14.75)Z=−0.880.380Ne Bo, 109/L10.55 (7.36, 12.39)10.55 (7.37, 12.02)10.55 (6.86, 13.04)Z=−1.210.225Ne percent Bo, %0.87 (0.81, 0.89)0.87 (0.81, 0.89)0.87 (0.82, 0.89)Z=−0.640.523Hb Bo, g/L131.00 (110.50, 151.50)132.50 (110.00, 153.00)129.00 (111.50, 149.50)Z=−1.070.284PLT Bo, 109/L152.66 (114.00, 177.00)152.66 (115.00, 178.50)152.66 (108.50, 171.50)Z=−0.350.726Pt Bo, s13.20 (12.10, 13.70)13.10 (12.10, 13.70)13.20 (12.10, 13.70)Z=−0.510.610Aptt Bo, s30.00 (27.10, 32.92)29.95 (27.08, 32.92)30.20 (27.10, 34.20)Z=−0.750.451INR Bo1.14 (1.05, 1.20)1.14 (1.05, 1.20)1.15 (1.05, 1.21)Z=−0.850.396Alb Bo, g/L37.03 (33.50, 40.80)37.03 (33.70, 40.82)37.03 (33.15, 40.30)Z=−0.490.625Alo Bo, g/L25.86 (23.30, 28.15)25.86 (23.30, 28.13)25.86 (23.25, 28.15)Z=−0.090.931Tbil Bo, µmol/L18.30 (11.90, 22.20)18.10 (12.00, 21.87)18.86 (11.85, 22.45)Z=−0.080.939Idil Bo, µmol/L16.80 (10.20, 20.00)16.85 (10.35, 20.10)16.70 (9.80, 19.75)Z=−0.380.702Dbil Bo, µmol/L0.00 (0.00, 1.01)0.00 (0.00, 1.01)0.00 (0.00, 1.01)Z=−0.220.825ALT Bo, U/L27.00 (19.00, 48.65)27.00 (19.00, 48.65)26.00 (18.50, 48.65)Z=−0.460.644AST Bo, U/L34.00 (25.00, 73.65)35.50 (25.88, 71.66)33.00 (25.00, 73.65)Z=−0.130.894UA Bo, µmol/L363.47 (298.15, 418.50)363.47 (293.85, 420.55)363.47 (303.10, 411.50)Z=−0.390.696Cr Bo, µmol/L88.70 (67.50, 105.20)88.65 (67.50, 105.00)89.90 (67.45, 106.20)Z = 0.001.000PCO2 Ao, mmHg42.40 (37.10, 47.15)42.55 (37.38, 47.40)41.90 (36.85, 46.35)Z=−1.520.127PO2 Ao, mmHg152.30 (108.15, 195.30)151.15 (106.47, 191.13)154.80 (123.00, 204.10)Z=−0.990.324FiO2 Ao, mmHg0.60 (0.50, 0.97)0.60 (0.50, 0.91)0.60 (0.50, 1.00)Z=−0.650.514PH Ao, mmHg7.38 (7.34, 7.43)7.38 (7.33, 7.42)7.39 (7.35, 7.43)Z=−1.390.164Lac Ao, mmol/L2.30 (1.50, 3.70)2.20 (1.50, 3.50)2.50 (1.50, 4.05)Z=−1.580.113WBC Ao, 109/L9.92 (7.61, 12.54)9.92 (7.64, 12.16)9.92 (7.55, 13.09)Z=−0.310.756Ne Ao, 109/L8.77 (6.72, 11.41)8.71 (6.75, 11.09)8.88 (6.65, 11.97)Z=−0.590.558Ne percent Ao0.89 (0.86, 0.92)0.89 (0.86, 0.92)0.89 (0.85, 0.92)Z=−0.600.550Hb Ao, g/L96.00 (86.00, 106.00)96.07 (87.00, 106.00)94.00 (84.00, 104.50)Z=−1.000.318PLT Ao, 109/L95.00 (66.50, 122.00)94.50 (69.00, 122.00)96.00 (62.00, 122.50)Z=−0.170.867Pt Ao, s14.10 (13.20, 15.20)14.00 (13.20, 15.12)14.30 (13.25, 15.35)Z=−1.240.216Aptt Ao, s33.00 (29.30, 38.75)32.50 (28.80, 38.33)34.50 (30.00, 39.10)Z=−1.410.158INRAo1.22 (1.15, 1.32)1.22 (1.15, 1.31)1.24 (1.15, 1.33)Z=−0.770.440Alb Ao, g/L33.10 (29.95, 36.70)33.00 (30.10, 36.70)33.90 (29.70, 36.55)Z=−0.380.707Glo Ao, g/L21.60 (19.40, 23.80)21.60 (19.40, 23.73)21.70 (19.35, 23.80)Z=−0.040.972Tbil Ao, µmol/L23.30 (17.40, 31.40)23.85 (18.05, 31.13)22.40 (16.50, 33.35)Z=−0.410.682Idil Ao, µmol/L19.00 (12.95, 28.00)19.05 (14.07, 27.85)18.10 (11.60, 28.35)Z=−0.960.336Dbil Ao, µmol/L0.00 (0.00, 5.90)0.00 (0.00, 5.43)0.00 (0.00, 7.50)Z=−0.040.966ALT Ao, U/L23.00 (17.00, 39.00)23.00 (17.00, 39.00)22.00 (16.00, 37.50)Z=−0.870.383AST Ao, U/L49.00 (37.00, 79.50)49.00 (37.72, 73.25)48.00 (35.45, 82.50)Z=−0.480.635Ua Ao, µmol/L315.40 (246.30, 379.25)315.70 (248.25, 379.95)315.10 (246.30, 377.35)Z=−0.190.848Cr Ao, µmol/L90.00 (67.95, 124.15)89.00 (67.98, 119.85)90.90 (68.00, 130.30)Z=−0.630.531Pulmonary exudation, n (%)χ²=0.170.677 NO101 (23.88)69 (23.31)32 (25.20) YES322 (76.12)227 (76.69)95 (74.80)Hydrothorax, n (%)χ²=0.000.958 NO184 (43.50)129 (43.58)55 (43.31) YES239 (56.50)167 (56.42)72 (56.69)Hydropericardium, n (%)χ²=0.220.637 NO389 (91.96)271 (91.55)118 (92.91) YES34 (8.04)25 (8.45)9 (7.09)PCO2 Partial Pressure of Carbon Dioxide, PO2 Partial Pressure of Oxygen, FiO2 Fraction of Inspired Oxygen, WBC White Blood Cell Count, Ne Neutrophils, Hb Hemoglobin, PLT Platelet Count, PT Prothrombin Time, APTT Activated Partial Thromboplastin Time, INR International Normalized Ratio, Alb Albumin, Tbil Total Bilirubin, Idil Indirect Bilirubin, Dbil Direct Bilirubin, ALT Alanine Aminotransferase, AST Aspartate Aminotransferase, UA Uric Acid, Cr Serum Creatinine, Ao after operation, Bo before operation Comparison of laboratory tests and imaging data between the two groups of patients PCO2 Partial Pressure of Carbon Dioxide, PO2 Partial Pressure of Oxygen, FiO2 Fraction of Inspired Oxygen, WBC White Blood Cell Count, Ne Neutrophils, Hb Hemoglobin, PLT Platelet Count, PT Prothrombin Time, APTT Activated Partial Thromboplastin Time, INR International Normalized Ratio, Alb Albumin, Tbil Total Bilirubin, Idil Indirect Bilirubin, Dbil Direct Bilirubin, ALT Alanine Aminotransferase, AST Aspartate Aminotransferase, UA Uric Acid, Cr Serum Creatinine, Ao after operation, Bo before operation Table 3Comparison of surgical data between the two groups of patientsVariablesTotal (n = 423)0 (n = 296)1 (n = 127)Statistic P Time from onset to operation, h0.10 (0.05, 0.34)0.10 (0.05, 0.34)0.10 (0.05, 0.32)Z=−0.050.960Operation time, min255.00 (210.00, 315.00)250.00 (210.00, 310.00)260.00 (212.00, 332.50)Z=−0.540.587CPB time, min110.00 (85.00, 153.00)110.50 (85.00, 153.50)107.00 (85.50, 149.50)Z=−0.060.952Amount of bleeding, ml500.00 (300.00, 500.00)500.00 (300.00, 500.00)500.00 (300.00, 500.00)Z=−0.670.503Blood transfusion volume, ml500.00 (400.00, 750.00)500.00 (400.00, 800.00)400.00 (300.00, 600.00)Z=−2.37 0.018 * Drainage of 24 h Ao, ml432.00 (265.00, 605.00)448.50 (275.75, 620.00)400.00 (255.00, 575.00)Z=−1.030.303Emergency operation, n (%)χ²=0.180.669 NO51 (12.06)37 (12.50)14 (11.02) YES372 (87.94)259 (87.50)113 (88.98)Secondary thoracotomy, n (%)χ²=0.180.672 NO400 (94.56)279 (94.26)121 (95.28) YES23 (5.44)17 (5.74)6 (4.72)Ao after operation, CPB cardiopulmonary bypass Table 3Comparison of surgical data between the two groups of patientsVariablesTotal (n = 423)0 (n = 296)1 (n = 127)Statistic P Time from onset to operation, h0.10 (0.05, 0.34)0.10 (0.05, 0.34)0.10 (0.05, 0.32)Z=−0.050.960Operation time, min255.00 (210.00, 315.00)250.00 (210.00, 310.00)260.00 (212.00, 332.50)Z=−0.540.587CPB time, min110.00 (85.00, 153.00)110.50 (85.00, 153.50)107.00 (85.50, 149.50)Z=−0.060.952Amount of bleeding, ml500.00 (300.00, 500.00)500.00 (300.00, 500.00)500.00 (300.00, 500.00)Z=−0.670.503Blood transfusion volume, ml500.00 (400.00, 750.00)500.00 (400.00, 800.00)400.00 (300.00, 600.00)Z=−2.37 0.018 * Drainage of 24 h Ao, ml432.00 (265.00, 605.00)448.50 (275.75, 620.00)400.00 (255.00, 575.00)Z=−1.030.303Emergency operation, n (%)χ²=0.180.669 NO51 (12.06)37 (12.50)14 (11.02) YES372 (87.94)259 (87.50)113 (88.98)Secondary thoracotomy, n (%)χ²=0.180.672 NO400 (94.56)279 (94.26)121 (95.28) YES23 (5.44)17 (5.74)6 (4.72)Ao after operation, CPB cardiopulmonary bypass Comparison of surgical data between the two groups of patients Ao after operation, CPB cardiopulmonary bypass Screening of predictors The predictors were included in the LASSO regression for variable screening, and 13 predictors were identified, including BMI, alcohol consumption, PO2-bo, FiO2-bo, PLT-bo, Alb-bo, PO2-ao, FiO2-ao, Ne-ao, APTT-ao, pulmonary exudation, CPB time and blood transfusion volume (Fig. 2).Fig. 2LASSO regression for feature selection. A Trajectories of change in the coefficients for each variable in the LASSO regression. B Cross-validation plot of LASSO regression. LASSO, least absolute shrinkage and selection operator LASSO regression for feature selection. A Trajectories of change in the coefficients for each variable in the LASSO regression. B Cross-validation plot of LASSO regression. LASSO, least absolute shrinkage and selection operator Comparison of the predictive models As shown in Fig. 3a, in the decision tree model constructed based on the training set, PO2-ao and Fio2-ao were the earliest variables used for node partitioning. The ranking of the importance of the characteristics in the decision tree model is illustrated in Fig. 3b, and PO2-ao and FiO2-ao were determined to be the two most important characteristics. The ranking of the importance of the variables in the random forest model is shown in Fig. 4. Thirteen variables, such as PO2-ao, Fio2-ao, PO2-bo, and pulmonary exudation, were included. Fig. 3Decision tree model. PO2, Partial Pressure of Oxygen; FiO2, Fraction of Inspired Oxygen; BMI, body mass index; ALB, Albumin; Ne, Neutrophils; APTT, Activated Partial Thromboplastin Time; CPB, cardiopulmonary bypass; Ao, after operation; Bo, before operation Decision tree model. PO2, Partial Pressure of Oxygen; FiO2, Fraction of Inspired Oxygen; BMI, body mass index; ALB, Albumin; Ne, Neutrophils; APTT, Activated Partial Thromboplastin Time; CPB, cardiopulmonary bypass; Ao, after operation; Bo, before operation Fig. 4Random forest feature ranking. PO2, Partial Pressure of Oxygen; FiO2, Fraction of Inspired Oxygen; BMI, body mass index; ALB, Albumin; Ne, Neutrophils; APTT, Activated Partial Thromboplastin Time; CPB, cardiopulmonary bypass; PLT, Platelet Count; Ao, after operation; Bo, before operation Random forest feature ranking. PO2, Partial Pressure of Oxygen; FiO2, Fraction of Inspired Oxygen; BMI, body mass index; ALB, Albumin; Ne, Neutrophils; APTT, Activated Partial Thromboplastin Time; CPB, cardiopulmonary bypass; PLT, Platelet Count; Ao, after operation; Bo, before operation The model construction results demonstrated that the random forest model outperformed the other machine learning algorithms in predicting ARDS following AAD surgery, with an impressive AUC value of 0.978. The logistic regression model demonstrated an AUC value of 0.965, followed by the decision tree model (with an AUC value of 0.881), the support vector machine model (with an AUC value of 0.835), and the K-nearest neighbour model (with an AUC value of 0.807), as shown in Fig. 5; Table 4.Fig. 5ROC curves of the five machine learning algorithm models for predicting the occurrence of ARDS after AAD surgery ROC curves of the five machine learning algorithm models for predicting the occurrence of ARDS after AAD surgery Table 4ROC curves of the five machine learning algorithms for predicting ARDS after AAD surgeryAlgorithmsAUCSensitivitySpecificityAccuracyPPVNPV95% CIRandom Forest0.9780.9620.9580.9610.9740.9390.924–0.991Logistic Regression0.9650.9620.9580.9600.9740.9390.930–0.999Decision Tree0.8810.8980.8540.8820.9100.8370.806–0.956Support Vector Machine0.8350.8990.7710.8500.8660.8220.805–0.931K-Nearest Neighbour0.8070.8230.7920.8110.8670.7310.735–0.879 Table 4ROC curves of the five machine learning algorithms for predicting ARDS after AAD surgeryAlgorithmsAUCSensitivitySpecificityAccuracyPPVNPV95% CIRandom Forest0.9780.9620.9580.9610.9740.9390.924–0.991Logistic Regression0.9650.9620.9580.9600.9740.9390.930–0.999Decision Tree0.8810.8980.8540.8820.9100.8370.806–0.956Support Vector Machine0.8350.8990.7710.8500.8660.8220.805–0.931K-Nearest Neighbour0.8070.8230.7920.8110.8670.7310.735–0.879 ROC curves of the five machine learning algorithms for predicting ARDS after AAD surgery Calibration and clinical utility The calibration curve closely followed the ideal line, with Brier scores of 0.472, indicating that the predicted probabilities of the random-forest model differ minimally from the observed event probabilities and thus demonstrating high calibration (Fig. 6). Decision-curve analysis showed that the random-forest model yields a positive net clinical benefit across a wide range of threshold probabilities, affirming its substantial clinical utility (Fig. 7). Fig. 6Calibration of the predictive model Calibration of the predictive model Fig. 7DCA curves of the predictive model DCA curves of the predictive model Discussion ARDS is an acute respiratory failure syndrome with a complex pathogenesis. The early identification of and intervention to address this condition are crucial for improving the prognosis of patients [9, 15]. The new definition of ARDS in 2023 revised the “Berlin Standard” in terms of both the conceptual model and diagnostic criteria [14]. Based on the new global definition of ARDS established in 2023, in this study, a predictive model for the occurrence of ARDS in patients with AAD after surgery was constructed by using five machine learning methods. Based on recent research and clinical guidelines, clinicians can use advanced methods to identify high-risk patients in the early preoperative and postoperative stages. These methods are designed to implement targeted preventive and therapeutic measures, with the aim of significantly reducing the incidence and mortality rates of ARDS, which is a severe pulmonary condition characterised by acute respiratory failure. Thoracotomy for AAD is characterised by significant surgical trauma and the need for cardiopulmonary bypass. Owing to the prolonged exposure of the thoracic cavity, the use of anaesthesia, and hypothermia during surgery, patients may experience insufficient pulmonary ventilation, weakened diaphragm activity, and an impaired or inhibited cough reflex after surgery [16]. Ischaemia-reperfusion injury and the release of substantial inflammatory mediators during CPB are among the causes of ARDS after cardiac surgery [17]. Thoracotomy with cardiopulmonary bypass can lead to a systemic inflammatory response, which activates the complement system, neutrophils, macrophages and various inflammatory mediators while weakening the immune response capacity, thereby causing pulmonary dysfunction [10]. A previous study focusing on the risk factors for postoperative ARDS in patients with Stanford type A acute aortic dissection revealed that BMI, time from onset to surgery, preoperative white blood cell count, and preoperative oxygenation index are independent risk factors for postoperative ARDS [2]. This study revealed that NE-ao, PO2-bo, FiO2-bo, PO2-ao and FiO2-ao are associated with an increased risk for the occurrence of ARDS, which aligns with previous findings that highlight the significance of oxygenation indices such as PaO2/FiO2 in the diagnosis and prognosis of ARDS. Obesity itself is a chronic inflammatory disease. Obese patients demonstrate continuous inflammatory responses, and the levels of neutrophils and inflammatory cytokines in their blood are relatively high [18]. This chronic inflammatory state can impair the homeostasis of pulmonary blood vessels, lead to the dysfunction of pulmonary endothelial cells, promote the occurrence of ARDS [19]. A high BMI may also exacerbate lung injury by increasing the recruitment of neutrophils to the lungs and promoting the activation of macrophages [20, 21]. Serum albumin levels are closely related to the occurrence of ARDS. Research has indicated that low serum albumin levels represent an independent risk factor for ARDS, and a linear dose-response relationship exists, wherein decreased serum albumin levels are associated with an increased risk of ARDS [22]. Specifically, for every 1 g/L decrease in the serum albumin level, the risk of ARDS increases by 17% (OR = 0.83; 95% CI: 0.75–0.91; p < 0.001) [22]. In addition, low serum albumin levels can lead to an increase in interstitial water content in the lungs and a decrease in lung compliance, resulting in pulmonary oedema. These low levels may also exacerbate lung oxidative damage by reducing antioxidant capacity (such as by reducing free radical scavenging), thereby promoting the occurrence of ARDS [23]. When the serum albumin level was lower than 35 g/L, the risk of ARDS significantly increased (OR = 2.16; 95% CI: 1.47–4.06; p = 0.017). Excessive blood transfusions performed during aortic dissection surgery not only endanger patient health but also significantly increase the incidence of postoperative ARDS. Due to the large amounts of cell debris and exogenous proteins found in the blood that is being transfused into the body, the patient’s lung function may be affected, and ARDS could be elicited, which is consistent with previous research results [9]. Pulmonary exudation is among the core pathophysiological features of ARDS and is closely related to the occurrence, development and prognosis of ARDS [24]. One of the main pathophysiological mechanisms of ARDS involves the impaired integrity of the alveolar capillary barrier, which causes fluid, protein, and inflammatory cells to seep from the blood vessels into pulmonary interstitial and alveolar spaces [25]. This exudation process is caused by damage to alveolar capillary endothelial cells and alveolar epithelial cells [26]. In patients with ARDS, the alveolar cavities are filled with protein-rich fluids, inflammatory cells (such as neutrophils and macrophages), and the cytokines and enzymes that are secreted by these cells. These exudates can lead to alveolar collapse, reduced lung compliance and gas exchange disorders, thereby causing severe hypoxemia [26, 27]. Clinicians should focus on the pulmonary exudation of patients during the perioperative period and adjust the treatment strategy in a timely manner according to the actual clinical situation. Inflammation serves as a pivotal mechanism in ARDS and triggers the release of inflammatory factors that activate the coagulation system, which results in abnormal changes in indicators such as APTT. This interplay between inflammation and coagulation exacerbates lung damage, thereby facilitating the onset of ARDS. Inflammatory response-induced vascular endothelial cell injury constitutes a crucial pathophysiological process in ARDS development. After endothelial cells are injured, they release tissue factors, which activate the coagulation system and lead to abnormal coagulation changes in indicators such as APTT. Patients with ARDS often demonstrate coagulation dysfunction, which manifests in abnormal changes in coagulation indicators such as PT, APTT and PLT. Coagulation dysfunction can lead to the formation of microthrombi, which can block pulmonary capillaries, thereby causing local ischaemia and hypoxia, which further aggravates lung damage and induces ARDS. With the advancement of artificial intelligence technology, deep learning algorithms have been applied to CT image analysis to automatically identify lung lesions and predict ARDS. For instance, a team led by Gao Yuan from Renji Hospital, School of Medicine, Shanghai Jiao Tong University, developed an AI-based UNETR model. This model was trained with over 276,623 CT slices and achieved early prediction of ARDS with an AUC value of 0.916 [28]. In addition to CT images, the study also attempted to combine various imaging data (such as ultrasound, magnetic resonance imaging, etc.) with clinical data to construct a multimodal imaging fusion prediction model to improve the accuracy and reliability of ARDS prediction [28]. Significant progress has also been made in predicting ARDS using deep learning models. In addition to the aforementioned UNETR model, there are models based on Convolutional Neural Networks (CNN) that are used to automatically extract imaging features and predict the risk of ARDS occurrence. This model also demonstrated good predictive performance, with a sensitivity of 96.3% and a specificity of 96.6% [29]. Despite the progress made in ARDS prediction models, there are still some challenges. First, the development of models requires a large amount of high-quality data, and there are difficulties in obtaining and annotating data. Second, differences between patient populations and clinical environments may lead to inconsistencies in model performance. Additionally, the interpretability and clinical operability of the models are issues that need to be further addressed. Conclusion In summary, prediction models for ARDS in patients with AAD after surgery were developed by using five machine learning algorithms. Among the five machine learning algorithm models that were adopted in this study, the random forest model demonstrated the highest prediction efficiency for ARDS (AUC = 0.978). This study enables medical staff to promptly identify high-risk patients, thus facilitating the targeted prevention of ARDS onset and progression. Although the model performed well in this study, we acknowledge the potential risk of overfitting, especially This study was a single-centre retrospective study, and no external validation was conducted to further verify the predictive efficacy of this model. Future multicentre, large-sample prospective studies are needed for further verification. Supplementary Information Supplementary Material 1. Supplementary Material 1.