Abstract
Managing congenital cataract (CC) post-surgery is a lifelong challenge due to the risk of complications like Visual Axis Opacification (VAO) and high Intraocular Pressure (IOP). This paper introduces CC-Guardian, an integrated AI system designed to personalize patient follow-up and facilitate remote telehealth diagnosis. By combining risk prediction with automated scheduling and remote image analysis, CC-Guardian detects sight-threatening complications significantly earlier than conventional methods while reducing the socioeconomic burden on families.
Introduction
Congenital cataract (CC) is a leading cause of treatable childhood blindness. While surgery is effective, the long-term management of postoperative complications remains a logistical hurdle. The current "one-size-fits-all" follow-up model often fails to catch complications that arise between scheduled visits and imposes significant travel and financial costs on families who must visit specialized centers for routine checkups. The authors sought to create a "closed-loop" AI agent that manages the patient's journey from risk prediction to scheduling and remote intervention.
Methods
The researchers developed CC-Guardian, a closed-loop AI agent comprising three functional modules:
- Prediction Module: Utilizes a Naive Bayes algorithm to analyze 12 clinical variables (e.g., age at surgery, lesion density), categorizing patients into high or low risk for VAO or high IOP. Naive Bayes was selected for its computational efficiency in real-world deployment.
- Dispatching Module: A rule-based system that generates individualized follow-up schedules, providing extra "buffer" visits for high-risk patients during peak complication windows.
- Telehealth Module: Employs a 101-layer Deep Residual Network (ResNet-101) to analyze postoperative retro-illumination images uploaded via a smartphone app to determine if immediate intervention is required.
The system was integrated into a cloud platform and smartphone app, allowing data upload from local primary care hospitals for review by specialists. It was trained on clinical records from 594 patients and 4,881 postoperative images. Validation was performed using internal datasets, multi-resource datasets from seven hospitals, and a retrospective longitudinal test of 141 patients to simulate real-world performance.
Results
CC-Guardian demonstrated exceptional diagnostic performance, with AUCs of 0.991 for VAO prediction and 0.979 for high IOP. The telehealth module achieved an AUC of 0.996 for intervention decisions. Performance remained robust in multi-resource validation, with AUCs between 0.944 and 0.981.
Crucially, the system enabled earlier detection of complications—37.8 days earlier for VAO and 19.2 days earlier for high IOP compared to conventional methods. Beyond clinical metrics, the system was estimated to save families an average of 928 miles of travel, 24.9 hours of time, and $1,324 USD annually.
Discussion
The success of CC-Guardian highlights a shift from simple AI classification to clinical orchestration. By prioritizing computational efficiency and focusing on "risk days saved," the system is optimized for real-world deployment. However, limitations remain, including population specificity (primarily Chinese datasets) and the current requirement for slit-lamp photography. Future work will focus on expanding hardware compatibility through smartphone-based adapters and conducting prospective randomized controlled trials to further validate these findings across diverse populations.
Summary updated