Study design
This retrospective study evaluated clinical, laboratory, and simple CT parameters from the admission of normotensive patients with acute PE. The Bova score and 2019 ESC algorithm were used for risk stratification. The outcomes of interest were defined as the occurrence of adverse outcomes within 30 days after admission into hospital. Eligible patients were randomized 2:1 into training and validation dataset. Training dataset were used to develop and evaluate multivariable logistic regression models for predicting the outcomes of interest. The discriminatory power was evaluated by comparing the nomogram to the established risk stratification systems. The consistency of the nomograms was evaluated using the validation dataset. The investigators independently collected the data regarding the clinical, laboratory, and CT parameters as well as data regarding the risk stratification scores and outcomes of interest. This research was approved by the Institutional review board of Shengjing Hospital of China Medical University (No.2020PS522K) and informed consent was exempted due no intervention to patient’s treatment.
Patient Selection
Normotensive patients with acute PE were evaluated if they were treated at the Shengjing Hospital of China Medical University between January 2011 and May 2020. The diagnosis and management of acute PE was based on the 2019 ESC guidelines [1]. The inclusion criteria were age of ≥ 18 years and diagnosis of PE based on CT pulmonary angiography. The exclusion criteria were pregnancy, reception of reperfusion treatment before admission, and missing data regarding CT parameters, echocardiography, cardiac troponin I (c-Tn I), and N-terminal-pro brain natriuretic peptide (NT-pro BNP).
Clinical Data And Risk Stratification
The patients’ medical records were reviewed to collect their demographic characteristics and baseline data from their admission regarding heart rate, systolic pressure, history of disease, c-Tn I concentration, and NT-pro BNP concentration. RV dysfunction was diagnosed by a transthoracic echocardiography within 24 hours after admission as anyone or more of the following parameters: RV dilation, an increased RV-left ventricle (LV) diameter ratio (> 0.9), hypokinesia of the free RV wall, increased tricuspid regurgitant jet velocity, and/or decreased tricuspid annular plane systolic excursion [2].
The risk stratification was based on the 2019 ESC algorithm [1] and Bova scores [13], with classifications as “low risk,” “intermediate-low risk,” and “intermediate-high risk” (Additional file1:Table S1 and Additional file 2:Table S2). The 2019 ESC algorithm evaluated c-Tn I (cutoff: 0.04 µg/L), NT-pro BNP (cutoff: 600 pg/mL), RV dysfunction, and simple PE severity index [1]. The Bova scores were calculated based on c-Tn I (cutoff: 0.05 µg/L), RV dysfunction, heart rate (cutoff: 110 beats/min), and systolic pressure (cutoff: 90–100 mmHg).
Outcomes Of Interest
The outcomes of interest were defined as the occurrence of adverse outcomes within 30 days after admission. Adverse outcomes were defined as PE-related death, the need for mechanical ventilation, the need for cardiopulmonary resuscitation, and the need for life-saving vasopressor and reperfusion treatment [12, 16].
Measurement Of Ct Parameters
Three simple CT parameters were selected for the analysis. The first factor was thrombus location, which was categorized as within the central pulmonary artery (CPA embolism) [16, 17], spanning both sides of the bifurcation (saddle-CPA embolism) [18], and outside the CPA (non-CPA embolism) (Additional file 3:Figure S1 a, b and c). The second factor was the RV and LV diameters in the short-axis plane, which were measured as the maximal diameter from the cardiac intima to the interventricular septum [19], as well as the relative ratio of the RV/LV short-axis diameters (Additional file 4:Figure S2 a). The third factor was the maximum chamber diameters, which were measured using a 4-chamber view perpendicular to the atrial and interventricular septum (Additional file 4:Figure S2 b) [10, 20], as well as the relative ratios of the RV/LV and right atrium (RA)/left atrium (LA) 4-chamber diameters. All CT parameters were measured using Mimics Medical software (version 19.0, Mimics Medical software, Leuven, Belgium).
Development Of The Models And Risk-scoring Tools
The models were developed based on three steps: (a) identifying relevant prognostic factors, (b) developing and validating the models, and (c) evaluating the models’ discriminatory power relative to the 2019 ESC algorithm and Bova score. To prevent over-fitting and simplification, classification and regression tree (CART) analysis was used to identify relevant prognostic factors. The optimal cutoff points for significant prognostic factors were determined based on the optimal separation from a penalized discriminant analysis [21, 22].
In the second step, eligible patients were randomized 2:1 into training and validation datasets based on the TRIPOD standard [23]. Univariate and multivariate logistic regression analyses were used to investigate independent prognostic factors using the training dataset, and a nomogram was created by converting each regression coefficient from the multivariate logistic regression onto a scale of 0 points (low) to 100 points (high). The total scores for all variables were summed [24], and the different risk groups were separated based on their total nomogram scores via CART analysis.
In the third step, the validation dataset was used to evaluate the models’ consistency relative to the observed outcomes [22]. The models’ abilities to predict adverse outcomes were also compared to the 2019 ESC algorithm [1] and the Bova score [13] based on receiver operating characteristic curve (ROC) and decision curve analysis (DCA). The final risk-scoring tools were published as free web-based calculators.
Statistical analysis
Continuous variables were expressed as mean ± standard deviation and compared using the Student’s t test. Categorical variables were presented as number (%) and compared using the χ2 test. Recursive partitioning analysis and the CART were used to dichotomize each variable while controlling for confounders and to divide the training dataset into different risk group according the total nomogram scores [25]. Univariate and multivariate logistic regression analyses were used to evaluate the different factors, and the results were expressed as odds ratios (ORs) with the corresponding 95% confidence intervals (CIs). The nomograms’ predictive performances were evaluated based on the concordance index (C-index) and calibration with 1,000 bootstrap resampling [24]. The ROC curves were used to evaluate the sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and the area under the ROC curve (AUC). Clinical utility was evaluated based on net benefits from the DCA. DeLong’s test was used to compare the AUC values [26]. Differences were considered significant at p-values of < 0.05 and all analyses were performed using R software (version 4.0.01; R Foundation, https://www.r-project.org).