Prediction of risk stratification in acute pulmonary embolism using a combined CTPA radiomics and machine learning model

This study extracted radiomics features of blood clots from computed tomography pulmonary angiography (CTPA) images of acute pulmonary embolism (APE) patients, and constructs a predictive model for APE patient risk stratification by comparing multiple machine learning (ML) algorithms. This retrospective study analyzed patients with APE, documenting their clinical characteristics (clinical and hematological indicators), conventional CTPA imaging parameters. Patients are stratified into low-risk group and high-risk group based on risk stratification. The data were randomly divided into a training cohort and a validation cohort in a 7:3 ratio. We developed 2 distinct models: nomogram model based on clinical characteristics and an image parameter model based on conventional CTPA imaging parameters. Features were screened by the least absolute shrinkage and selection operator (LASSO) method. The radiomics predictive model was constructed using 7 ML algorithms and selected the one with the best performance.The combined model was constructed by integrating radiomics features with clinical characteristics and image parameter using an optimal algorithm. Model performance was evaluated using receiver operating characteristic curves and decision curves analysis, with the area under the curve (AUC) compared via the DeLong test. A total of 202 patients were included. The training cohort contained 141 cases, while the validation cohort included 61 cases. Multivariate logistic analysis revealed that the pulmonary embolism severity index, dyspnea, troponin, and right ventricle/ left ventricle short-axis maximum diameter ratio were independent clinical predictors (P < .05). A total of 12 radiomics features were selected through LASSO screening. Logistic regression demonstrated the best performance among the 7 ML algorithms. In the validation cohort, the combined model demonstrated significantly superior predictive performance (AUC = 0.935) compared to the nomogram model (AUC = 0.805), the image parameter model (AUC = 0.788), and the radiomics model (AUC = 0.860) (P < .05). Decision curve analysis indicated that the combined model demonstrated higher clinical net benefit when the risk threshold exceeded 0.04. Calibration curves and Hosmer-Lemeshow test further confirmed the combined model's goodness of fit in both the training and validation cohorts. The combined model combining clinical characteristics, conventional CTPA imaging parameters, and radiomics features demonstrated outstanding predictive performance and clinical applicability in risk stratification for APE patients.