ĐÁNH GIÁ TÁC ĐỘNG CỦA TRÍ TUỆ NHÂN TẠO ĐA PHƯƠNG THỨC ĐỐI VỚI QUY TRÌNH CHẨN ĐOÁN UNG THƯ VÒM MŨI HỌNG TẠI VIỆT NAM

Abstract

Nasopharyngeal carcinoma in Vietnam is still commonly diagnosed at advanced stages because early symptoms are subtle and clinically fragmented information often exists across endoscopy, cross-sectional imaging, and biomarkers. This study evaluated the impact of multimodal artificial intelligence integrating endoscopy, CT/MRI, and EBV-DNA on diagnostic performance, clinical decision-making, and implementation feasibility. A mixed-method design was used, combining a practice survey, retrospective-prospective analysis of 256 patients, and comparison between conventional and AI-assisted diagnostic pathways. Multimodal AI improved early-stage detection from 28.9% to 46.5% (p<0.001), reduced missed small lesions from 18.7% to 6.2% (p<0.001), shortened diagnostic time from 28.4 to 16.7 days (p<0.001), and increased correct biopsy indication from 72.3% to 88.7%. The multimodal model achieved sensitivity of 96.5%, specificity of 90.2%, accuracy of 93.8%, and an AUC of 0.97. AI changed diagnostic decisions in 22.3% of cases and increased physician agreement for biopsy indication from 68.5% to 87.9%. Multimodal AI therefore improves both technical performance and real-world diagnostic workflow for nasopharyngeal carcinoma in Vietnam.