Diagnostic accuracy of AI-assisted point-of-care ultrasound for abdominal free fluid detection in FAST trauma assessment: a systematic review and meta-analysis

Abstract

Study objective: Point-of-care ultrasound (PoCUS) is widely used in trauma care through the Focused Assessment with Sonography for Trauma (FAST) protocol, but its accuracy is highly operator-dependent. Artificial intelligence (AI) may reduce variability and improve reliability. This systematic review and meta-analysis evaluated the diagnostic accuracy of AI-assisted PoCUS for detecting free fluid in trauma patients. Methods: We searched PubMed, Scopus, and Web of Science up to April 2025, following PRISMA-DTA guidelines (PROSPERO ID: CRD420250615096). Eligible studies assessed AI-assisted PoCUS using the FAST or FAST-equivalent views and provided sufficient data for diagnostic accuracy analysis. Because one eligible study evaluated an ascites cohort rather than trauma patients, we included it in the broader abdominal free-fluid analysis but excluded it from the trauma-only sensitivity analysis. Pooled estimates were calculated with a bivariate random-effects model. Study quality was assessed using QUADAS-AI tool. Results: Seven retrospective studies (n = 2,332 patients, >34,000 images/videos) were included. The pooled analyses were based on the diagnostic units reported by the original studies, which varied across image-, frame-, video/clip-, FAST-view-, and examination-level data. In the trauma-only analysis, pooled sensitivity for detecting abdominal free fluid was 91.1% (95% CI: 77.9–96.8%) and specificity was 97.5% (95% CI: 95.3–98.7%), with negligible heterogeneity (I² < 1.0) and an AUC of 0.98. In the broader abdominal free-fluid analysis, which included one non-trauma ascites cohort, pooled sensitivity was 91.4% (95% CI: 81.6–96.3%) and specificity was 96.8% (95% CI: 86.5–99.3%), with an AUC of 0.97. CNN-based models showed similar performance (sensitivity 92.2%, specificity 95.4%, AUC 0.97). Narrative review highlighted substantial variability across models: high-performing frameworks such as YOLOv3 and ResNet50-V2 demonstrated sensitivities of 0.90–0.99, whereas others (e.g., VGG11_bn, MaskRCNN) were markedly less accurate. Evidence for pericardial effusion detection was limited to a single retrospective study and should be interpreted cautiously. No dedicated pelvic-view diagnostic accuracy data were available. Conclusion: AI-assisted PoCUS shows promising retrospective diagnostic performance for detecting abdominal free fluid during FAST assessment in trauma settings. The strongest evidence currently supports abdominal applications, whereas data for the cardiac/pericardial component of the FAST examination remain insufficient. Prospective multicenter studies with real-time workflow integration are needed before routine clinical implementation can be recommended.