Improvement of artificial intelligence-based computed tomography pulmonary angiography in identifying acute pulmonary embolism.

Background: Acute pulmonary embolism (APE) is a potentially fatal condition. Although artificial intelligence (AI) algorithms show promise in diagnosis, overreliance on them may lead to critical errors. This study aimed to improve diagnostic accuracy by integrating AI algorithms with clinical indicators. Methods: A retrospective analysis was conducted on 329 patients with suspected APE who underwent computed tomography pulmonary angiography (CTPA), lower-extremity venous ultrasound, and laboratory tests. Independent risk factors were identified through logistic regression. The diagnostic performance of individual and combined predictors was assessed via receiver operating characteristic (ROC) curves. Interrater agreement between the AI algorithms and radiologists was evaluated via Kappa analysis. Additionally, subgroup analysis was performed on patients with discordant AI predictions and clinical indicators. Differences in confirmed APE rates among these subgroups were calculated to assess the diagnostic value of AI algorithms in this patient population. Results: The AI algorithms demonstrated excellent standalone diagnostic performance, with an area under the curve (AUC) of 0.933 [95% confidence intervals (CI): 0.894-0.973; P<0.001] and an odds ratio of 803.28 (95% CI: 163.05-3,957.36; P<0.001). They showed strong agreement with radiologists (κ=0.87; P<0.001). Subgroup analysis revealed that among patients positive for deep vein thrombosis (DVT) or elevated D-dimer levels (defined as >1 mg/L) but with negative AI predictions, the confirmation rates of APE were low (5.1% and 4.6%, respectively). In contrast, among those negative for DVT or normal D-dimer levels (≤1 mg/L) but with positive AI predictions, the APE confirmation rates were significantly higher (75.0% and 57.1%, respectively). These findings suggest that AI algorithms can help identify cases missed by clinical indicators and may reduce unnecessary imaging in high-risk patients without actual embolism. Logistic regression analysis identified DVT, elevated plasma fibrinogen levels (>4 g/L), and male gender as independent risk factors for APE. When AI predictions were combined with these clinical indicators, the diagnostic performance improved markedly, with an AUC of 0.981 (95% CI: 0.967-0.994; P<0.001), outperforming both AI algorithms or clinical indicators alone. Conclusions: Integrating AI algorithms with clinical indicators significantly enhances the accuracy of diagnosing APE, reduces the risk of misdiagnosis, and improves screening efficiency.