Diagnostic Test Accuracy of Artificial Intelligence in Large Vessel Occlusion: A Systematic Review and Meta-Analysis.
Authors
Affiliations (11)
Affiliations (11)
- Department of Neurology, Faculty of Medicine, Andalas University, Padang, Indonesia, unand.ac.id.
- Department of Medicine, Faculty of Medicine, Andalas University, Padang, Indonesia, unand.ac.id.
- Department of Medicine, Faculty of Medicine, Sebelas Maret University, Surakarta, Indonesia, uns.ac.id.
- Department of Medicine, Faculty of Medicine, Islamic University of Indonesia, Yogyakarta, Indonesia, uii.ac.id.
- Department of Medicine, Faculty of Medicine, Veteran National Development University (UPNVJ), Jakarta, Indonesia.
- Department of Medicine, Faculty of Medicine, Sumatra Utara University, Medan, Indonesia.
- Department of Medicine, Faculty of Medicine, Ganesha University of Education, Bali, Indonesia, undiksha.ac.id.
- Department of Medicine, Faculty of Medicine, Jenderal Soedirman University, Purwokerto, Indonesia, unsoed.ac.id.
- Department of Medicine, Faculty of Medicine, Hasanuddin University, Makassar, Indonesia, unhas.ac.id.
- Department of Medicine, Faculty of Medicine, Jember University, Jember, Indonesia, unej.ac.id.
- Department of Medicine, Faculty of Medicine, Public Health, and Nursing, Gadjah Mada University, Yogyakarta, Indonesia, ugm.ac.id.
Abstract
Large vessel occlusion (LVO) requires prompt detection, and CT angiography (CTA) is frequently used due to its short acquisition time and visibility of vessels. Artificial intelligence (AI), including Viz-LVO, CINA-LVO, RAPID-CTA and JLK, may be available as emerging tools for supporting timely and accurate diagnoses. This study aimed to examine and summarise the evidence of AI diagnostic performance in detecting LVO. Scopus, PubMed and ScienceDirect were utilised to search relevant articles before February 2, 2025. Studies were included in the primary outcomes analysis if they reported an overall confusion diagnostic matrix and were included in the secondary outcomes if they reported AI's diagnostic performance by occlusion site. Of the 878 records, 11 articles were included, and 10.937 patients were identified. The pooled sensitivity and specificity were 0.87 (95% CI: 0.76-0.93) and 0.95 (95% CI: 0.91-0.97). The positive likelihood ratio (PLR) showed statistical significance (9.55 (95% CI: 5.79-13.30; <i>p</i> < 0.001; <i>I</i> <sup>2</sup>: 99.9%)), whereas the negative likelihood ratio (NLR) was not significant with a pooled value of 0.14 (95% CI: 0.03-0.25; <i>p</i> < 0.624; <i>I</i> <sup>2</sup>: 0%). The pooled AUC and DOR were substantial, with a pooled value of 0.87 (95% CI: 0.83-0.92; <i>p</i> < 0.001; <i>I</i> <sup>2</sup>: 98.4%) and 4.69 (95% CI: 4.19-5.19; 0.001; <i>I</i> <sup>2</sup>: 98.5%), respectively. Three covariates were identified (type of AI, AI software and region). However, significant heterogeneity remains in pooled PLR, AUC and DOR. The anterior circulation occlusion performed was generally acceptable, demonstrating good performance for M1 and ICA-type <i>T</i> occlusion and moderate performance for M2 occlusion. However, poor performance was observed in ICA Type I and posterior circulation occlusion. In conclusion, AI has demonstrated excellent performance in sensitivity, specificity, PLR, AUC and DOR while showing limitations in NLR, suggesting that negative cases detected by AI require careful reevaluation through imaging review and assessment of patients' clinical profiles to ensure better diagnostic accuracy.