Prognostic dynamic nomogram integrated with metabolic acidosis for in-hospital mortality and organ malperfusion in acute type B aortic dissection patients undergoing thoracic endovascular aortic repair

In the derivation cohort, complicated ATBAD patients who underwent TEVAR procedures at our hospital from January 2010 to December 2017 were eligible for inclusion in this study. Patients with the following characteristics were excluded: (1) connective tissue disease, including Marfan and Loeys-Dietz syndromes; (2) blunt traumatic thoracic aortic injury; (3) malignant tumor; (4) previous aortic intervention; (5) pre-existing renal or visceral malperfusion; (6) incomplete data due to missing blood gas analyses. To examine the generalizability of the model, an external validation cohort of patients was prospectively collected separately, using the same inclusion and exclusion criteria as the derivation cohort in the same institution, from January 2018 to December 2019. Finally, the derivation and external validation cohort contained 286 and 77 patients, respectively. All patients underwent computed tomography (CT) scans with contrast enhanced, thin-sliced (range 0.75 to 1.25 mm) spiral CT (64-slice multidetector LightSpeed VCT; General Electric Fairfield, CT). Multiplanar reconstruction was performed by Aquarius iNtuition software (Terarecon, San Mateo, CA, USA). This study was approved by the ethics committee of the Guangdong Provincial People’s Hospital (#201807) and the need for informed consent was waived because of the retrospective nature of the analysis.

TEVAR was performed when ATBAD was complicated [1]. The details of the procedure at our hospital had been previously described [12]. Briefly, the procedures were performed with suitable anatomy in a cardiac catheterization room under local anaesthesia. All stent grafts were deployed retrogradely via percutaneous femoral artery access to obliterate the proximal entry tear. The left subclavian artery (LSA) were covered when necessary to obtain 1.5–2 cm proximal landing zone. The choice of reconstruction of the LSA mainly depends on the vertebrobasilar circulation by operators. The diameters of aortic stent grafts were generally oversized by 5% to 10% according to the aortic pathologies. Most patients were treated with a single endograft prosthesis, and additional pieces were placed only when the initial graft did not produce the desired result of coverage of the entry tear and expansion of the true lumen as determined by angiography but the distal end of the stent-graft was still above the diaphragm. No branched stent-graft was used during the study period and balloon angioplasty of the proximal seal zone was avoided if possible to prevent retrograde extension of the dissection into the aortic arch.

Retrospective data on age, sex, medical history, coexisting medical conditions, imaging features, operation parameters, and follow-up records were collected and analyzed. In our center, the blood gas analyses were performed at admission, in the each morning and at the time of disease progression. Each patient had at least one measurement before the procedure. The lowest pH, the lowest bicarbonate concentration, the nadir BE and the highest lactate were employed.

ATBAD was defined as a type B aortic dissection occurring less than 14 days after the onset of symptoms [1]. Complicated type B aortic dissection was described as persistent or recurrent pain, uncontrolled hypertension despite full medication, early aortic expansion, malperfusion, and signs of rupture (haemothorax, increasing periaortic and mediastinal haematoma) [1]. Patients with renal dysfunction were those with an estimated glomerular filtration rate (eGFR) lower than 60 ml/min/1.73m² at admission [13]. The maximum pre-operative outer to outer aortic diameter was measured orthogonal to the vessel centreline. The supply of the abdominal arterises (coeliac artery, superior mesenteric artery, left renal artery and right renal artery) were assessed by multiplanar reconstructed enchanced-CT. The diagnosis of malperfusion in the context of ATBAD was based on the patient’s presenting symptoms in addition to CT confirmation and intra-operative visualization of obstruction to any aortic branch vessels [6, 14].

The endpoint of interest was described as a composite outcome of in-hospital death or new-onset organ malperfusion including visceral malperfusion, renal malperfusion, lower extremity malperfusion and spinal cord malperfusion according to the description of White et al. [14]. A patient having multiple events was considered as having only one event.

Continuous data were presented as mean ± standard deviation or median (quartiles 1 to 3) and were compared using the Student's t-or the Mann–Whitney U tests depending on distribution. The Shapiro–Wilk test was selected for the normality test. Qualitative data are presented as frequencies (percentages) and compared using the Chi-square or Fisher’s exact tests.

To establish the predictive model, the least absolute shrinkage and selection operator regression (LASSO) was used to select variables in the predictive model. By applying multivariate logistic regression, we established a predictive model. An optimal penalization factor was determined using the “pentrace” function in Harrell’s R package “rms” [15] to avoid overfitting. The bootstrapping approach was used for internal model validation as it is considered more efficient than split-data and cross-validation methodologies [15]. Bootstrapping replicates the process of sample generation from an underlying population by drawing samples with replacement from the original data set. The model can be constructed and validated with 100% of the number of subjects using bootstrapping, which make the prediction and internal validity accurate and stable, especially when the sample size is small. Nevertheless, cross-validation or split-data uses part of subjects for model construction and the other for validation, which might result in unstable and biased estimates of performance. The risk predictive model for in-hospital mortality and organ malperfusion in ATBAD patients undergoing TEVAR was presented using a nomogram. In addition, a web-based dynamic prediction tool based on the nomogram has been created to facilitate the calculation and aid in the decision-making process in clinical practice (https://​sycardiovascular​.​shinyapps.​io/​DynNomappBE_​version2/​).

Model performance was assessed under three aspects: (1) The discriminatory capacity was evaluated by the area under the receiver operating characteristics (ROC) curve (AUC), while the bias-corrected AUC was calculated using bootstrapping 1000 times. (2) The calibration ability was evaluated using the following four different methods: the Hosmer–Lemeshow test; calibration plot; the Brier score, as well as the intercept and slope of the calibration. (3) The clinical effectiveness was assessed using the decision curve analysis (DCA). In addition, we also derived an optimal cut-off threshold to determine the positive predictive value (PPV) and the negative predictive value (NPV) to assess clinical usefulness. A sensitivity analysis was also performed in the entire cohort.

A total of 363 complicated ATBAD patients who underwent TEVAR were enrolled in this study, with 286 of them in the derivation cohort and 77 in the external validation cohort. The majority of the participants were male (n = 326, 89.8%) in both cohorts and median age was 52 years (IQR: 45–62) (Table 1). The derivation and validation cohorts showed relatively well balanced features, except for a higher eGFR level and a lower incidence of maximum aortic diameter ≥ 5.5 cm in the derivation cohort (Table 1). The prevalence of in-hospital mortality or organ malperfusion was similar between the derivation and external validation cohorts (n = 41 [14.3%] vs n = 11 [14.3%], P = 0.991, Table 2).

After the initial selection and the elimination of redundant candidates based on LASSO regression analysis, the final selection of predictors was performed by multivariate logistic regression analysis (Additional file 1: Table S1). To enhance the clinical use, the continuous BE value was divided into four groups (≥ 0; − 5 to 0; − 10 to − 5; ≤ − 10) [5] and enrolled into multivariate analysis (Additional file 1: Table S2). As a result of no significant difference between − 5 to 0 group and ≥ 0 group, these two groups were combined to form the low risk group (> − 5). Therefore, BE was split into three groups: low risk group (> − 5); moderate risk group (− 10 to − 5) and high risk group (≤ − 10). Multivariate analysis demonstrated that moderate and high risk BE group, maximum aortic diameter ≥ 55 mm, renal dysfunction and D-dimer ≥ 5.44 μg/mL were independent risk factors for in-hospital mortality and organ malperfusion in complicated ATBAD patients undergoing TEVAR (Table 3). Each predictor received a score based on the regression coefficient derived from the multivariate logistic regression model and summed to the final risk prediction model.

A Brier score of 0.072 and Hosmer–Lemeshow goodness-of-fit tests with 5.59 of chi-square value (P = 0.471) in the derivation cohort suggest a good fitting of the model. To further detect any deviation between observed and predicted events, we internally validated the model by bootstrapping (1000 iterations) the slope and intercept of the calibration plot and the AUC. The original AUC was 0.87, while the bias-corrected estimate was 0.85. The original intercept and slope of the calibration plot were 0 and 1, respectively. The bias-corrected values were -0.13 and 0.89 indicating a mild overfitting. Therefore, it was added a penalty using the “pentrace” function in Harrell’s R package “rms” [15] to improve the model fit and obtain a new calibration plot (Fig. 1). After applying the penalty factor, the original AUC was 0.87 (Fig. 2) and the bias-corrected AUC was 0.85, while the bias-corrected estimates of the intercept and slope were − 0.02 and 0.98 (Fig. 1), respectively. Based on these results, a nomogram was configured (Fig. 3).

The external validation of the predictive model was performed with data prospectively collected separately from January 2018 to December 2019 at the same institute from which the previous data were collected. Of the 77 complicated ATBAD patients who underwent TEVAR, 11 developed in-hospital mortality or organ malperfusion. The validation AUC was 0.86 (Figs. 2), which was consistent with the derivation cohort AUC of 0.87 (P = 0.927). The Brier score was 0.097, indicating good model calibration. It was also performed a sensitivity analysis on patients from both cohorts and the predictive model continued to perform well to predict the incidence of in-hospital mortality and organ malperfusion (AUC, 0.86; 95% confidence interval [CI] 0.80–0.92).

DCA was used to assess the clinical effectiveness of the predictive model. The net benefits of the risk model were obviously more prominent than “treating-all-patients” or “treating-none”, in the derivation (Fig. 4a) and external validation cohorts (Fig. 4b). In addition, we derived the optimal cut-off threshold for the predictive model providing PPV and NPV that might provide clinically useful information. An optimal model threshold score (0.39) gave an NPV of 0.94 and a PPV 0.71 in the derivation cohort. The threshold was carried forward and the NPV remained high (0.93), whereas the PPV decreased to 0.60 in the external validation cohort.