Spatial Computing and Machine Learning for Mobile Visual Field Triage: A Retinal Detachment Early Warning Prototype

Overview

Retinal detachment is a severe ocular condition in which the retina separates from its underlying tissue, potentially causing progressive and irreversible vision loss if left undetected. Traditional visual field tests (perimetry) require large, stationary clinical equipment and rigid patient compliance, limiting accessibility for early screening. The purpose of this engineering project was to develop a preliminary, mobile diagnostic prototype utilizing spatial computing and machine learning, executed in two distinct phases: Visual Field Mapping and Diagnostic Prediction. Phase 1: Visual Field Mapping. The first phase focused on securely gathering and visualizing peripheral vision data. Using an Apple Vision Pro, an immersive virtual reality visual field test was developed. To ensure data integrity, native eye tracking technology enforced strict central fixation, automatically pausing the test if the user's gaze deviated. A secure pipeline captured five metrics per stimulus: eccentricity, polar angle, brightness, reaction time, and hit/miss status. A Logistic Regression algorithm was then applied to interpolate untested spatial gaps, generating a continuous topographical heatmap estimating the user's overall field of view. Phase 2: Diagnostic Prediction. While visual heatmaps are useful, they are susceptible to human error; a user who blinks during the test produces a blind spot on the map that mimics a mild detachment. To advance the prototype from a mapping tool into a diagnostic engine, a training dataset of 200 simulated patient profiles was generated, including Normal baselines alongside Mild, Medium, and Severe detachments simulated by algorithmically projecting "dark curtain" scotomas. Crucially, synthetic human anomalies—such as random misses mimicking blinks or fatigue—were injected into this data. A Random Forest Classifier was trained on this noisy dataset to leverage spatial clustering features to distinguish between scattered user errors and clustered physiological vision loss. In preliminary testing on unseen synthetic data, the model consistently filtered through injected artifacts, outputting a triage recommendation on whether a clinical visit is warranted. Conclusion. This project demonstrates that combining eye-tracking spatial computing with anomaly-trained machine learning can create an accessible early-warning triage system. While the current model is trained on synthetic data as a proof of concept, the architecture is designed to scale to clinical datasets for future validation.