D28-02 The Relationship Between Ai-derived Embolus Volume And Clinical Severity In Acute Pulmonary Embolism: A Retrospective Cohort Study

Rationale

Quantitative clot burden on computed tomography pulmonary angiography (CTPA) is commonly estimated with visual obstruction scores, which are time-consuming and prone to interobserver variability. We evaluated whether deep learning-derived embolus volume provides an objective imaging metric associated with clinical severity and outcomes in acute pulmonary embolism (PE).

Methods

We performed a retrospective observational cohort study at a tertiary university hospital. Consecutive adults (18 years or older) with CTPA-confirmed acute PE between January 2021 and July 2025 were included. Patient selection is summarized in Figure 1. A convolutional neural network segmentation model quantified total intravascular embolus volume (mL) on CTPA. Clinical severity was categorized as non-massive, submassive, or massive PE using guideline-based criteria, and echocardiographic right ventricular (RV) dysfunction was recorded. Because embolus volume was right-skewed, nonparametric tests were used: Kruskal-Wallis for severity groups; Spearman correlation between log10(volume) and severity (coded 1-3); Mann-Whitney U for RV dysfunction; and Spearman correlations for biomarkers. Receiver operating characteristic (ROC) analysis evaluated discrimination for RV dysfunction and 90-day mortality; Pulmonary Embolism Severity Index (PESI) and simplified PESI (sPESI) underwent parallel ROC evaluation for mortality.

Results

The cohort included 105 patients; 23 (22%) died within 90 days. Embolus volume differed across severity categories (p = 0.028) and increased with greater severity (Spearman rho=0.2545; p = 0.0088). Embolus volume correlated with troponin (rho=0.31; p = 0.0016) but not B-type natriuretic peptide (BNP) (rho=0.02; p = 0.91). Patients with RV dysfunction had higher embolus volume than those without (11.06 vs 3.42 mL; p = 0.00042). Embolus volume moderately discriminated RV dysfunction (area under the curve [AUC] 0.70) and showed modest discrimination for 90-day mortality (AUC 0.617). In comparison, PESI showed good mortality discrimination (AUC 0.79), whereas sPESI performed poorly (AUC 0.55).

Conclusion

AI-derived embolus volume is associated with clinical severity and RV dysfunction in acute PE, supporting its potential role as an objective imaging marker of acute cardiopulmonary strain. Its prognostic performance for 90-day mortality was modest and inferior to PESI, suggesting volumetric clot burden alone is insufficient for longer-term risk prediction.

This abstract is funded by: None