Abstract
Objective
Acute pulmonary embolism (APE) is a major cardiovascular disease, the risk stratification is challenging. This study aims to investigate the feasibility of developing a prediction model for risk stratification of APE patients based on radiomics features of the clots.
Materials and Methods
Computer tomography pulmonary angiography images from 66 APE patients (50% of males, 51.5% of age > 60 years) with different risk levels (33 high-risk and 33 non-high-risk) were analyzed retrospectively. Qanadli and Mastora index was used for evaluating the obstruction degree manually. Radiomics features were extracted from the clots. Independent t-test, least absolute shrinkage selection operator (LASSO) and correlation matrix were used to select the most discriminative features. Support vector machine (SVM), Gaussian Naïve Bayes (GNB), K-Nearest Neighbor (KNN), Decision Tree (DT), Random Forest (RF) and Multi-layer Perceptron-neural net (MLP-NN) were used to build risk stratification models. Mean accuracy and its standard deviation of a 10-fold-cross-validation and the correct rate of these six classifiers were evaluated and compared.
Results
1737 radiomics features were extracted from the segmented clots. 16 features, including 5 shape-based features, 8 texture-based features and 3 histogram-based features, were identified as the most discriminative features after eliminating redundant and irrelevant ones. Mean accuracies and their standard deviations showed that MLP-NN had the best performance (0.9042 ± 0.3029), followed by GNB (0.8625 ± 0.3334), SVM (0.8542 ± 0.4070), RF (0.8542 ± 0.3787), KNN (0.8292 ± 0.3038) and DT (0.7667 ± 0.3122), while the GNB model can predict the highest number of the high risk APE patients. MLP-NN yielded the highest correct prediction rate (86.36%), followed by GNB (84.85%), SVM (84.85%), RF (81.82%) and KNN (81.82%), all improved over the clinical scoring systems of Qanadli and Mastora scoring Indices (72.73% and 77.27%).
Conclusions
The radiomics features combined with MLP-NN can be potentially applied in the clinical risk stratification process to assist the treatment decision for APE patients.