Chronic disease management is burdened by long-term monitoring and frequent visits. Artificial intelligence (AI), specifically individualized prediction using machine learning and intelligent telehealth computing, offers a potential solution. However, fully integrated AI applications for predicting and managing chronic diseases are lacking, with limited real-world validation. Congenital cataract (CC), a chronic condition with high risks of high intraocular pressure (IOP) and visual axis opacification (VAO), serves as a suitable test case. Current follow-up plans are often generic and insufficient, particularly in developing countries where access to specialized care is limited. This study aims to address these shortcomings by developing an AI-powered system to improve the follow-up management of CC patients, improving the timeliness and efficiency of care.

The introduction cites several studies demonstrating the potential of AI in various medical fields, including nephrology, cardiology, and ophthalmology. It also acknowledges the potential of telehealth computing to enhance AI's effectiveness. However, it highlights the lack of fully integrated AI systems for prediction and telehealth in chronic disease management and the need for real-world validation of AI's benefits in healthcare.

The study developed CC-Guardian, an AI agent composed of three modules: a prediction module, a dispatching module, and a telehealth module. The prediction module, using a Naive Bayes algorithm, identifies high-risk patients based on 12 easily collected clinical variables. The dispatching module creates personalized follow-up schedules based on the prediction results. The telehealth module, employing a deep residual network (101-layer), analyzes postoperative retro-illumination images to detect complications and recommend interventions. The system was validated through internal and multi-resource validation datasets. A web-based smartphone app and a prediction-telehealth cloud platform were developed for clinical application. A retrospective self-controlled test evaluated the system's real-world effectiveness in terms of complication prediction, telehealth detection, and cost-effectiveness.

CC-Guardian achieved high accuracy in both internal and multi-resource validation. In internal validation, the prediction module showed 96.7% sensitivity and 97.5% specificity for VAO and 96.2% sensitivity and 95.2% specificity for high IOP. The telehealth module had AUCs of 0.991, 0.979, and 0.996 for detecting VAO, high IOP, and intervention, respectively. Multi-resource validation yielded similarly high performance. The retrospective self-controlled test showed that CC-Guardian accurately predicted 96.8% of VAO cases and 96.2% of high IOP cases. Early detection resulted in an average of 37.8 fewer risk days for VAO and 19.2 fewer risk days for high IOP. The system also significantly reduced travel distance (928.6 miles per year per family), time (24.9 hours per year per family), and cost ($1324.1 per year per family) compared to conventional methods.

CC-Guardian successfully integrates individualized prediction and telehealth computing, translating accurate predictions and detections into improved treatment timing and significant cost savings. Early diagnosis of VAO and high IOP is crucial for preventing permanent vision impairment. The telehealth component enhances patient compliance by reducing the burden of travel and time commitment. This approach also optimizes resource allocation by freeing specialists to focus on new patients. The findings demonstrate a significant advancement in patient care management for congenital cataracts and provide a model for managing other chronic diseases.

This study demonstrates the successful application of AI in improving the management of congenital cataracts. CC-Guardian, with its high accuracy, personalized scheduling, and telehealth integration, significantly improves early detection of complications, reduces socioeconomic burdens, and offers a scalable model for managing chronic diseases. Future research should focus on validating the system with larger, more diverse datasets and extending its application to other chronic conditions.

The study's limitations include a relatively small multi-resource validation dataset limited to a Chinese population, the system's current limitation to CC patients, the use of self-taught imagery for telehealth module training requiring assistive devices for infants, and a relatively short duration for the self-controlled test. Further research is needed to address these limitations and ensure broader generalizability.