Artificial intelligence manages congenital cataract with individualized prediction and telehealth computing

Abstract

A challenge of chronic diseases that remains to be solved is how to liberate patients and medical resources from the burdens of long-term monitoring and periodic visits. Precise management based on artificial intelligence (AI) holds great promise; however, a clinical application that fully integrates prediction and telehealth computing has not been achieved, and further efforts are required to validate its real-world benefits. Taking congenital cataract as a representative, we used Bayesian and deep-learning algorithms to create CC-Guardian, an AI agent that incorporates individualized prediction and scheduling, and intelligent telehealth follow-up computing. Our agent exhibits high sensitivity and specificity in both internal and multi-resource validation. We integrate our agent with a web-based smartphone app and prototype a prediction-telehealth cloud platform to support our intelligent follow-up system. We then conduct a retrospective self-controlled test validating that our system not only accurately detects and addresses complications at earlier stages, but also reduces the socioeconomic burdens compared to conventional methods. This study represents a pioneering step in applying AI to achieve real medical benefits and demonstrates a novel strategy for the effective management of chronic diseases.

Keywords: Computer science; Health care economics; Lens diseases; Translational research.

Fig. 1. Study pipeline for agent training, validation, application, and testing.

a Multidimensional clinical records of 594 congenital cataract patients were collected for prediction module training (279 VAO: 315 non-VAO; 341 high IOP: 253 normal). For telehealth module training, a total of 4881 postoperative retro-illumination images were obtained (2615 follow-up: 2266 intervention). Each image was independently described and labeled by an expert panel. b Two datasets were used for the validation of the trained agent, including internal validation dataset (clinical records of 142 patients, and 1220 follow-up images) and multi-resource dataset (clinical records of 79 patients, 214 follow-up images). c A web-based smartphone app was implemented, and a prediction-telehealth cloud platform was prototyped for the clinical application of our intelligent follow-up system. d A retrospective self-controlled test was conducted to investigate the real-world efficiency of our follow-up system in complication prediction, telehealth detection, and cost-effect benefit. VAO visual axis opacification, IOP intraocular pressure.

Fig. 2. Functional architecture of CC-Guardian.

a The prediction module is trained to identify high-risk patients likely to suffer complications (the occurrence of VAO, happen or not and the occurrence of high IOP, happen or not). b The dispatching module is responsible for scheduling individual follow-up based on the prediction module results. c In the telehealth module, a clinical decision regarding further treatment (intervention or continued follow-up) is made after each telehealth examination based on follow-up images and IOP values. VAO visual axis opacification, IOP intraocular pressure, m months.

Fig. 3. Highly accurate performance of CC-Guardian in internal validation.

a Using the prediction module, our agent predicted VAO with 96.7% sensitivity and 97.5% specificity, and high IOP with 96.2% sensitivity and 95.2% specificity in internal validation. Using the telehealth module, our agent provided intervention suggestions with 99.1% sensitivity and 99.4% specificity. b Confusion matrices for agent validation. c Our agent had AUCs of 0.991, 0.979, and 0.996 for detecting VAO, high IOP, and intervention, respectively, in internal validation. AUC area under curve, TP true positive, TN true negative, FP false positive, FN false negative.

Fig. 4. Comparative performance of CC-Guardian in multi-resource validation.

a Using the prediction module, our agent predicted VAO with 94.0% sensitivity and 93.5% specificity, and high IOP with 96.4% sensitivity and 94.1% specificity for the multi-resource dataset. Using the telehealth module, our agent provided intervention suggestions with 95.9% sensitivity and 94.5% specificity for the multi-resource dataset. b Confusion matrices for agent validation. c Our agent had AUCs of 0.944, 0.961, and 0.981 for detecting VAO, high IOP, and intervention, respectively, for the multi-resource dataset. AUC area under curve, TP true positive, TN true negative, FP false positive, FN false negative.

Fig. 5. The prediction-telehealth cloud platform.

When potential patients register at the specialized care center (CCPMOH), their clinical metrics (valuable for the prediction module) are collected with their permission and immediately uploaded to the CC-Guardian cloud platform for complication prediction. Based on the prediction results, the dispatching module designs an individualized follow-up schedule and sends a short message to notify each corresponding patient in a timely manner. Patients can complete their regular follow-up in primary care hospitals and upload their examination results to the web-based telehealth module. If the telehealth module recommends intervention, the fast-track notification system is triggered, and an emergency notification is sent to doctors at the specialized care center (CCPMOH) for immediate confirmation. These patients are informed that they should undergo intervention to manage complications after confirmation.

Fig. 6. Real-world efficiency in retrospective self-controlled test.

Error bars presented the standard deviation. a, b The longitudinal follow-up records of 141 CC patients (987 follow-up visits to CCPMOH) were retrospectively used for real-world testing. Before using CC-Guardian, a total of 93 patients underwent VAO and 105 patients suffered from High IOP. After using our system, 90 cases of VAO (90/93, 96.8%) and 101 cases of high IOP (101/105, 96.2%) are successfully predicted. c–e Each family would have saved an average travel distance of 928.6 miles (1185.4 vs. 256.8 miles, P < 0.001), an average time of 24.9 h (33.8 vs. 8.9 h, P < 0.001), and an average expenditure of $1324.1 ($1791.4 vs. $467.3, P < 0.001) per year. VAO visual axis opacification, IOP intraocular pressure.