Preventing Vision Impairments in Children with AI

GoCheck Kids is an FDA-registered pediatric photoscreening platform. The smartphone app helps prevent vision impairments and blindness in children aged 1 to 18. Healthcare practitioners use it to screen for amblyopia and other vision disorders at the point of care. The company serves over 6,500 pediatric teams across the US and Europe.

01 The Challenge

A screening app that worked well, with one gap that created manual work and delays

Vision impairment is the most common disabling childhood condition in the United States. Amblyopia alone affects 1-6% of children. Early detection is treatable in most cases. Yet only about 40% of three-year-olds receive a vision screening. GoCheck Kids was built to close that gap by putting a photoscreener on every pediatrician’s phone.

The app captures a photo of a child’s eyes, evaluates the image, and returns a result. In a small percentage of screenings, the result came back inconclusive. The most common cause: the child was not looking at the camera. When that happened, a trained review team had to flag the image and notify the provider to rescreen. That manual step consumed up to 20% of the review team’s capacity.

GoCheck Kids saw the opportunity to catch those cases in real time. If the app could detect a misaligned gaze at capture, it could alert the provider immediately. Retake the photo on the spot. No waiting. No manual review. No second visit. The barrier was the classification model: it caught misaligned gaze about 25% of the time.

GoCheck Kids partnered with Provectus, an AI-first systems integrator and solutions provider, to build the ML infrastructure.

02 The Approach

Audit the existing models, then build an experimentation engine that lets three engineers do the work of a lab

Provectus started by reviewing GoCheck Kids’ image classification software, prior modeling approaches, and data pipelines. The team assessed the dataset: over one million images and more than 150,000 eye screens. They evaluated labeling quality, data structure, and pipeline support for iterative training.

The baseline was clear. The models worked. The infrastructure for improving them did not. Running a single experiment required manual setup. Comparing results across experiments meant tracking artifacts by hand. The team needed a system where experiments ran at scale and results tracked automatically.

Provectus built that system on AWS. Amazon SageMaker handles model training. A pipeline coordination layer manages experiments. The infrastructure includes experiment tracking, model versioning, and automated data relabeling.

03 The Build

Automated ML pipelines, experiment tracking, and real-time model integration into the screening app

The build delivered three major components.

First, a secure, auditable environment for model training and experimentation. Engineers run experiments against the full dataset. Results are logged with full reproducibility: metrics, predictions, model artifacts, and data versions.

Second, automated pipelines for data relabeling and retraining. As new labeled data and user feedback arrive, the pipeline reprocesses and retrains without manual intervention. The dataset improves with every screening cycle.

Third, integration of the improved models into the GoCheck Kids mobile app. The gaze detection model now runs at the point of capture. If a child is not looking at the camera, the app alerts the provider in real time.

The infrastructure aligns with GoCheck Kids’ product roadmap. New models for additional screening conditions can be trained on the same foundation.

04 The Results

From 25% gaze detection to 91%, with 100+ experiments in three weeks

The new ML infrastructure changed both the product’s accuracy and the speed at which the team improves it. In three weeks, three engineers completed over 100 large-scale experiments across the full dataset.

3X

Detection accuracy improvement

From 25% to 91% gaze detection

The original model caught misaligned gaze about one time in four. The new model catches it nine times out of ten. Precision on other classifications held steady. For practitioners, that means fewer inconclusive results and fewer callbacks.

The operational impact was immediate. Manual image review dropped as the app began catching quality issues at capture. The review team that spent 20% of its time on rescreening requests now focuses on other work. Providers get a definitive result in the exam room instead of a follow-up days later.

With automated pipelines handling data preparation, 95% of engineering time goes to experimentation. The team iterates on models faster. Each improvement reaches the app sooner.

05 What’s Next

An ML foundation that grows with every screening

GoCheck Kids now has the infrastructure to improve screening accuracy with every new dataset. The automated pipelines support new model development for additional vision conditions. Provectus works with GoCheck Kids on extending detection capabilities as the company expands.