Deep Learning-Based Electrocardiogram Analysis for Pulmonary Embolism Diagnosis: A Systematic Review and...

Although computed tomography pulmonary angiography (CTPA) is the diagnostic gold standard for confirming pulmonary embolism (PE), it has limitations such as the risks associated with contrast media and logistical delays. Electrocardiography (ECG) is a widely accessible and safe alternative; however, it also has a low diagnostic sensitivity.

This meta-analysis aimed to evaluate the diagnostic performance of artificial intelligence (AI)-based deep learning models using ECG data for PE detection and identify their strengths, limitations, and future directions.

PubMed, Embase, and the Cochrane Library were comprehensively searched to identify studies evaluating AI models using ECG for PE detection. Studies were included if CTPA was used as the reference standard.

Five studies involving 467,017 patients were analyzed. The pooled sensitivity and specificity of the AI-based ECG models for PE detection were 70% (95% confidence interval [CI]: 68-72%) and 67% (95% CI: 66-67%), respectively. The pooled diagnostic odds ratio was 3.68 (95% CI: 3.31-4.09), and the summary receiver operating characteristic analysis yielded an area under the summary receiver operating characteristic curve of 0.754, indicating moderate diagnostic accuracy. Significant heterogeneity was observed across the studies (I² > 90%). Compared with ECG-only approaches, multimodal models integrating ECG with clinical data demonstrated superior diagnostic performance.

Current AI-based deep learning models for ECG analysis have moderate accuracy in detecting acute PE, with promising potential for clinical screening. However, the significant heterogeneity and limited external validation highlight the need for standardized methodologies and larger multicenter studies.