Artificial intelligence manages congenital cataract with individualized prediction and telehealth computing

A medical revolution driven by artificial intelligence (AI) has emerged in the near future. Individualized prediction using machine learning promises to be transformative and indispensable for complex medical situations, including nephrology, cardiology, and ophthalmology. Moreover, intelligent telehealth computing has emerged as a powerful potential enabler of electronic applications of medical AI. Although the promise of these technologies is broad, the full integration of prediction and telehealth has not been achieved, and the actual benefits of AI regarding healthcare applications remain to be validated.

Chronic diseases and conditions are the leading causes of death and disability worldwide. Follow-up management of chronic disease patients remains one of the most intractable healthcare problems, and solutions are urgently needed. Presently, conventional follow-up plans scheduled by clinicians are one-size-fits-all and are primarily based on personal experience or limited clinical findings. Furthermore, among developing countries, only a few specialized care centers are capable of effective examinations and accurate interventions. The sparse distribution of these centers poses significant difficulty and economic pressure for patients, resulting in low priority for follow-ups. Therefore, there is a great need for precise effective follow-up management leveraging AI as highly desirable.

The study focuses on the follow-up management of patients with congenital cataract (CC), a typical chronic condition characterized by high long-term risk of two main complications: high intraocular pressure (IOP) and visual axis opacification (VAO). Rigorous follow-up care and timely intervention are necessary to prevent the irreversible and permanent loss of vision caused by these complications. Therefore, CC is a clear testing case for exploring follow-up management strategies for chronic conditions.

To explore the feasibility of applying AI to improve the follow-up care, we applied Bayesian and deep-learning algorithms to create CC-Guardian, an intelligent agent that identifies potential high-risk CC patients who are likely to suffer complications, (ii) a dispatching module that schedules individual follow-up based on the prediction results, and (iii) a telehealth module that makes intervention decisions in each follow-up examination. We verified CC-Guardian's performance using internal and multi-resource validation. We implemented a web-based smartphone app and proposed an operational

Study design and data: Multidimensional clinical records and postoperative images from congenital cataract (CC) patients enrolled in the Childhood Cataract Program of the Chinese Ministry of Health (CCPMH/CCPMOH) were collected. Training data comprised clinical records of 594 CC patients (pre- and post-surgery) and 4,881 postoperative retro-illumination slit-lamp images (2,615 follow-up, 2,266 intervention) acquired January 2011–December 2016. Eligibility criteria included a CC diagnosis, informed consent, and complete baseline, lesion, comorbidity, surgical, and 1-year complication records. Eligible images covered the posterior lens capsule via retro-illumination. Two validation sets were prepared: (a) internal validation with 142 patients (61 VAO, 81 non-VAO; 79 high IOP, 63 normal) and 1,220 images; and (b) a multi-resource dataset for external validation.

Prediction module (Naive Bayes): Twelve routinely obtainable variables were selected based on reported associations with postoperative complications: baseline (gender, age at surgery, laterality), lesion (area [extensive vs limited], density [dense vs non-dense], location [central vs peripheral]), comorbidities (nystagmus, microphthalmia, microcornea, persistent hyperplastic primary vitreous), and surgery (extraction type: UA, UA+PCCC, UA+PCCC+AV-IO; IOL implantation: primary vs secondary). Outcomes were binary: occurrence of VAO within 1 year and occurrence of high IOP (IOP > 21 mmHg) within 1 year. Algorithm selection compared naive Bayes, random forest, and a neural network using stratified k-fold cross-validation on the 594-patient set. Naive Bayes was chosen for its accuracy and computational simplicity, and retrained on all 594 patients for subsequent validations. Variable contribution analysis indicated non-dominant roles for age at surgery in predicting both VAO and high IOP.

Dispatching module (follow-up scheduling): Based on prediction outputs, individualized schedules were generated. If no complications were predicted, seven standard follow-up time points were scheduled: 1 week, 1 month, 2 months, 3 months, 6 months, 9 months, and 12 months post-surgery. If VAO risk was predicted, two additional visits at 7.5 and 10.5 months were added (reflecting later-year VAO tendency). If high IOP risk was predicted, three visits at 2.5 weeks, 2.5 months, and 4.5 months were added.

Telehealth module (deep residual network): A stacked deep residual network was trained to classify postoperative images into intervention versus continued follow-up. The architecture integrated convolution, max pooling, batch normalization, ReLU activations, data augmentation, and residual blocks to mitigate degradation and accelerate convergence. The network comprised 33 residual blocks (3 convolutions each; 99 layers total). Backpropagation addressed vanishing/exploding gradients. Transfer learning was used with ImageNet pretraining; outer layers were fine-tuned without freezing additional weights. Implementation used Caffe on Ubuntu 14.04 with CUDA 8.0.

Cloud platform and smartphone app: A prediction-telehealth cloud platform interfaced with a web-based smartphone app. At registration, patient metrics were uploaded for immediate complication risk prediction. The dispatching module generated individualized schedules and sent SMS reminders. Telehealth follow-ups were performed at primary care sites with image uploads to the cloud. If the telehealth model suggested intervention, a fast-track alert notified CCPMOH specialists for confirmation and expedited care.

Retrospective self-controlled test: Longitudinal records from 141 CC patients at CCPMOH (January 2017–May 2018) were retrospectively analyzed to assess real-world efficiency before versus after system use. Pre-system data included 987 in-person follow-up visits, with 93 patients developing VAO and 105 developing high IOP. For each patient, earlier detection potential was calculated as the reduction in “risk days” between conventional schedules and AI-driven schedules (e.g., additional AI-scheduled visits enabling detection at 10.5 months vs 12 months yields 45 days earlier). Telehealth benefits assumed all AI-scheduled follow-ups occurred via telehealth; travel distances were computed using AMap (Alibaba Group) geocoding and quickest-route round trips; time savings used standard speeds (highways 65 mph/104 kmh; main roads 30 mph; local roads/services 20 mph). Costs were estimated accordingly.

Statistical analysis: Due to skewed distributions, Wilcoxon paired tests compared distance, time, and cost before versus after AI use. Optimal operating thresholds used Youden’s index. Sensitivity and specificity 95% CIs employed binomial proportion intervals. Two-tailed tests with P < 0.05 were considered significant. Analyses were performed in R 3.2.4.

Ethics, data, and code: The study was IRB-approved (Sun Yat-sen University) and adhered to the Declaration of Helsinki. Data were HIPAA Safe Harbor–anonymized. Main data are in the paper/Supplement; additional datasets available upon reasonable request. Source code is available at https://github.com/longevitygc-guardian.

• Algorithm selection: In 5-fold cross-validation on 594 patients, naive Bayes achieved high performance and was comparable to random forest and neural network while being simpler computationally. Representative metrics (mean [95% CI]): VAO—accuracy 0.966 (0.916–0.991), sensitivity 0.964 (0.875–0.996), specificity 0.969 (0.827–0.996); High IOP—accuracy 0.975 (0.928–0.995), sensitivity 0.972 (0.902–0.997), specificity 0.979 (0.889–0.999). Random forest and neural network yielded broadly similar but not superior results.
• Internal validation (prediction module, n=142 patients): VAO prediction sensitivity 96.7% and specificity 97.5% (TP 59, TN 79, FP 2, FN 2). High IOP prediction sensitivity 96.2% and specificity 95.2% (TP 76, TN 60, FP 3, FN 3).
• Internal validation (telehealth model, n=1,220 images): AUCs: VAO 0.991; high IOP 0.979; intervention recommendation 0.996. Intervention classification counts: TP 544, TN 667, FP 5, FN 4.
• Multi-resource validation: VAO prediction sensitivity 94.0% and specificity 93.5%; high IOP prediction sensitivity 96.4% and specificity 94.1%. Telehealth intervention sensitivity 95.9% and specificity 94.5%. AUCs: VAO 0.944; high IOP 0.961; intervention 0.981.
• Retrospective self-controlled test (n=141 patients; pre-system 987 visits; 93 VAO, 105 high IOP): After applying the system, 90/93 VAO cases (96.8%) and 101/105 high IOP cases (96.2%) would have been successfully predicted. Earlier detection: VAO—73/90 (81.1%) benefited, total 2,759 risk days saved, mean 37.8 days/person (median 45 days). High IOP—89/101 (88.1%) benefited, total 1,709 risk days saved, mean 19.2 days/person (median 15 days).
• Telehealth benefits: Under an AI-telehealth paradigm, 1,579 telehealth visits would replace 987 in-person visits, saving per family on average 928.6 miles/year travel, 24.9 hours/year time, and $1,324.1/year in expenditure (all P < 0.001).

By integrating individualized risk prediction with an AI-driven telehealth workflow, CC-Guardian addresses two key challenges in chronic postoperative management of congenital cataract: timely identification of complications (VAO and high IOP) and reduction of follow-up burden. High sensitivities/specificities and AUCs in both internal and multi-resource validations indicate robust predictive and triage performance. When translated into a real-world retrospective self-controlled setting, these accuracies yield clinically meaningful benefits—earlier detection by weeks to months, which is critical for preventing amblyopia from VAO and irreversible optic nerve damage from high IOP, and substantial reductions in travel distance, time, and cost, improving adherence and access. The cloud-based platform operationalizes the workflow by enabling primary care telehealth imaging, automated scheduling with reminders, and fast-track alerts for specialist intervention, thereby reallocating specialist resources toward higher-need cases and first-visit patients. Collectively, the findings support that a combined prediction-telehealth approach can improve quality, timeliness, and efficiency of chronic condition management.

This study presents CC-Guardian, an AI agent that combines naive Bayes–based individualized complication prediction, risk-adaptive follow-up scheduling, and a deep residual network telehealth module within a cloud platform and smartphone app. The system achieves high diagnostic performance, enables earlier detection of VAO and high IOP, and reduces patient travel, time, and costs in a retrospective real-world analysis. These results demonstrate a viable path for translating AI into tangible clinical and socioeconomic benefits for chronic disease management. Future work should expand external validation across diverse populations and care settings, generalize the paradigm to other chronic conditions (e.g., hypertension, diabetes), refine telehealth imaging workflows for infants and uncooperative patients, and conduct longer-term, prospective evaluations of clinical outcomes and cost-effectiveness.

• External validation size and diversity: The multi-resource dataset was relatively small and limited to the Chinese population, constraining generalizability; validation in other ethnicities and countries is needed.
• Scope limited to congenital cataract: The system was developed and tested for CC patients; applications to other chronic conditions remain to be demonstrated.
• Image acquisition constraints: Telehealth imaging for infants may require assistance (e.g., straps) due to cooperation challenges.
• Short retrospective test duration: The self-controlled test period was relatively short (months), with incomplete post–May 2018 1-year records at submission, potentially limiting outcome assessment.
• Complication prevalence differences: Observed prevalence was higher than in some prior reports, possibly due to measurement/definition differences or “severe patient effects” (patients with complications more likely to complete follow-ups); such differences may affect comparability.
• Minor textual/data inconsistencies: Some reported counts and timings include typographical inconsistencies, though overall trends and significance are consistent across analyses.