Objective: To evaluate the effectiveness of an artificial intelligence system integrated with a medical knowledge graph for the early detection of nasopharyngeal carcinoma.
Subjects and Methods: This multicenter study included 1,200 cases, comprising 600 patients with nasopharyngeal carcinoma and 600 controls, collected from five healthcare institutions representing different levels of care in Vietnam. The artificial intelligence system was developed using a multimodal framework integrating medical imaging and clinical data, combined with a medical knowledge graph. Model performance was evaluated on an independent test set using accuracy, sensitivity, specificity, and area under the ROC curve. Results: The male-to-female ratio was 2.5:1 in the nasopharyngeal carcinoma group, compared with 1.2:1 in the control group (p < 0.05). On the independent test set of 180 cases, the system achieved an accuracy of 94.5% (95% CI: 92.1–96.8), sensitivity of 96.2%, specificity of 92.8%, and an AUC of 0.97. The correct detection rate for early-stage disease increased from 60.0% to 85.0%, while the miss rate for lesions smaller than 1 cm decreased from 42.0% to 12.0%. Compared with the image-only model, the multimodal model integrated with a medical knowledge graph improved diagnostic accuracy from 82.4% to 94.5%. Conclusion: The artificial intelligence system integrated with a medical knowledge graph enabled early detection of nasopharyngeal carcinoma, reduced missed small lesions, and improved diagnostic accuracy.