Development of a Multimodal Large Language Model for Early Warning and Diagnosis of Dry Eye Disease

Abstract Background This study aimed to develop and externally validate an instruction-tuned multimodal large language model (MLLM) for early warning and early diagnosis of dry eye disease (DED) using routinely obtainable clinical data. Methods A total of 874 participants were enrolled, including an internal development cohort (n = 743) from Peking University Third Hospital (October 2024 to August 2025) and an independent external validation cohort (n = 131) from three additional centers. Participants included both DED and non-DED individuals, diagnosed according to the Chinese Expert Consensus (2020) and TFOS DEWS II criteria. We developed DED-MLLM on the basis of Qwen2-VL and constructed a domain-specific instruction-tuning dataset integrating Ocular Surface Disease Index (OSDI), Standard Patient Evaluation of Eye Dryness (SPEED), self-reported questionnaires, ophthalmic imaging, and clinical examination data. Model performance was evaluated for risk stratification and diagnosis under both diagnostic frameworks, with ablation analyses used to examine the relative contributions of visual and structured clinical modalities. Results Based on the international diagnostic criteria for DED, the model achieved an accuracy of 0.8232 for early warning (specificity: 0.6873, sensitivity: 0.9558) and an accuracy of 0.9203 for diagnosis (specificity: 0.8776, sensitivity: 0.9720). According to the Chinese diagnostic criteria for DED, the model achieved an accuracy of 0.9584 for early warning (specificity: 0.9112, sensitivity: 0.9927) and an accuracy of 0.9327 for diagnosis (specificity: 0.8742, sensitivity: 0.9851). External validation showed AUCs of 0.910 and 0.957 for early warning and 0.960 and 0.893 for diagnosis. Ablation analyses indicated that structured clinical information contributed most strongly to diagnostic performance, whereas image-based information provided complementary, interpretable ocular-surface evidence. Conclusion The proposed DED-MLLM demonstrates excellent performance in both early warning and diagnosis of DED.