A three-tier AI solution for equitable glaucoma diagnosis across China's hierarchical healthcare system

Yi Zhou, Haitao Nie, Xinyu Gong, Minhui Dai, Zhaohong Guo, Xiaoling Deng, Mengyang Li, Yong Liu, Lingyu Sun, Xiangyi Tang, Ling Zhou, Zhiyao Tang, Ziqing Xia, Lemeng Feng, Wulong Zhang, Qingqing Yi, Xiaobo Xia, Bin Xie, Weitao SongNpj Digit. Medicine 2025 · DOI

TL;DR

This paper introduces "Multi-Glau," a novel three-tier AI system designed to address glaucoma diagnosis inequities in China's hierarchical healthcare system. By tailoring AI modules to the specific needs and data limitations of primary, secondary, and tertiary hospitals, Multi-Glau achieves high diagnostic accuracy and significantly improves ophthalmologists' performance, paving the way for more equitable glaucoma care.

What Problem Does This Paper Address?

Equitable access to specialized healthcare, like glaucoma diagnosis, is a significant challenge in countries with vast populations and uneven resource distribution, such as China. The country's three-tiered healthcare system (primary, secondary, tertiary) aims to address this, but implementation gaps persist, particularly in fields requiring specialized expertise and equipment. Current AI solutions for glaucoma often focus on single hospital tiers or rely on complete, high-quality data, making them unsuitable for resource-limited primary and secondary hospitals. This creates a diagnostic gap, exacerbating existing health disparities. The paper tackles the problem of adapting AI for glaucoma diagnosis to the realities of a tiered healthcare system with varying data availability and diagnostic capabilities.

Methods

The core of the solution is the "Multi-Glau" system, a three-tier AI framework:

  • Tier 1 (Primary Hospitals): Screening Module (XGBoost): This module uses readily available clinical features like age, gender, intraocular pressure (IOP), best-corrected visual acuity (BCVA), and cup-to-disk ratio (CDR). It deliberately avoids imaging data to be practical for primary care settings where such equipment might be unavailable.
  • Tier 2 (Secondary Hospitals): Pre-diagnosis Module (Freeze-Missing): This module is designed to handle incomplete data, a common issue in secondary hospitals. It uses a novel approach that can process missing modalities (e.g., missing visual field tests or OCT scans) and broadly categorizes glaucoma severity into early and serious stages.
  • Tier 3 (Tertiary Hospitals): Definitive Diagnosis Module (M³-VF): This module leverages multimodal data (including fundus photographs, OCT scans, visual field data, and clinical parameters) with a multi-perspective, multi-modal, and multi-stage attention mechanism to provide a precise four-stage glaucoma severity diagnosis (early, moderate, advanced, severe).

The system incorporates interpretability features (GradCAM++, Guided Backprop) to help physicians understand the AI's decisions.

Datasets

The study utilized a substantial dataset collected from the electronic medical records of 27,255 glaucoma patients at Xiangya Hospital Central South University (2017-2024). After data cleaning and exclusion of confounding factors (e.g., high myopia, secondary glaucomas), a dataset of 8,358 unique cases was used for development. For multicenter validation, data from Yiyang Central Hospital (99 cases) and Taojiang County People's Hospital (45 cases) were included.

Key Insights and Ideas

  • Tiered AI for Healthcare Equity: The most significant insight is the successful adaptation of AI to a hierarchical healthcare structure. Instead of a one-size-fits-all approach, Multi-Glau offers specialized AI tools for each tier, acknowledging and addressing their unique constraints.
  • Handling Data Scarcity is Crucial: The "Freeze-Missing" module's ability to perform pre-diagnosis with incomplete data is a major breakthrough. This directly tackles a critical bottleneck in real-world healthcare settings, where complete datasets are rare.
  • Simplicity Can Be Powerful: The screening module's reliance on basic clinical features, eschewing complex imaging, demonstrates that effective AI can be built even with limited resources, making it highly scalable.
  • AI as a Physician's Assistant, Not Replacement: The study highlights how AI, particularly with interpretability features, can significantly enhance ophthalmologists' diagnostic accuracy and decision-making, rather than simply automating the process.

Key Results

  • High Performance Across Tiers: Multi-Glau achieved impressive performance metrics:
  • Screening Module (XGBoost): AUC of 0.9254.
  • Pre-diagnosis Module (Freeze-Missing): AUC of 0.8650, even with up to 40% missing data.
  • Definitive Diagnosis Module (M³-VF): AUC of 0.9516.
  • Outperforms State-of-the-Art: Multi-Glau demonstrated superior performance compared to existing models, especially in handling missing data and providing precise glaucoma severity diagnosis.
  • Improved Physician Performance: In a study with ophthalmology experts, AI-assisted diagnosis led to significant improvements in diagnostic accuracy, particularly for severe cases. For instance, in the definitive diagnosis task, physician accuracy increased from 84% to 92% with AI assistance.
  • Streamlined Referral Process: Simulations showed that Multi-Glau could correct a substantial number of misdiagnoses during the referral process between hospital tiers.

Limitations and Future Work

  • Focus on Glaucoma: The current system is specific to glaucoma. Future work could extend the tiered AI framework to other eye diseases facing similar health equity challenges, such as diabetic retinopathy.
  • Data Completeness: While Freeze-Missing handles missing data, the study acknowledges the need for more comprehensive data collection, especially for risk factors like refractive errors and systemic diseases, which were sometimes incomplete.
  • Global Applicability: The current validation is primarily on East Asian populations. Further validation in diverse global populations is needed to ensure broader applicability and address potential geographic and diagnostic biases.
  • Inter-tier Knowledge Transfer: The authors plan to explore methods for pre-training feature extractors in lower-tier institutions using high-quality data from tertiary hospitals, further enhancing knowledge sharing across the healthcare system.

Summary updated

BibTeX
@article{zhou2025three,
  author = {Yi Zhou and Haitao Nie and Xinyu Gong and Minhui Dai and Zhaohong Guo and Xiaoling Deng and Mengyang Li and Yong Liu and Lingyu Sun and Xiangyi Tang and Ling Zhou and Zhiyao Tang and Ziqing Xia and Lemeng Feng and Wulong Zhang and Qingqing Yi and Xiaobo Xia and Bin Xie and Weitao Song},
  title = {{A three-tier AI solution for equitable glaucoma diagnosis across China's hierarchical healthcare system}},
  journal = {Npj Digit. Medicine},
  year = {2025},
  doi = {10.1038/s41746-025-01835-4},
}
Figure