How it started

What if a simple three-minute browser assessment could help researchers better understand how someone pays attention to the world around them?

That was the question behind NeuroVision, the project my team and I built at HackPrinceton 2026.

Spatial neglect is a neurological condition that can occur after a stroke or brain injury. Someone experiencing it may fail to notice objects or events on one side of their surroundings, even when their eyes are functioning.

Imagine eating food from only one side of your plate or overlooking objects on one side of a room without realizing anything is missing.

We wanted to explore whether interactive browser-based assessments could make spatial attention easier to measure and monitor.

That idea became NeuroVision.

What we built

NeuroVision is a browser-based research prototype designed to measure patterns of spatial attention through three interactive assessments.

Our goal was to make the experience accessible without requiring specialized equipment, software installation, or a complicated setup.

The three assessments take approximately three minutes in total.

1. Starry Night: Measuring visual detection

In this assessment, green dots appear and disappear across the screen while red targets flash at different positions.

The user responds whenever they notice a red target.

Behind the scenes, NeuroVision records reaction times, successful detections, missed targets, and differences between the left and right sides of the screen.

This helps us explore whether attention is distributed differently across the visual field.

1. Hollow Rings: Understanding visual search

Users are presented with hollow rings mixed among solid circles.

Their task is to identify all the hollow rings.

Although the interaction looks simple, the application records much more than the final score.

It tracks which objects were selected, which were missed, the order of selections, and how long the search took.

We can then examine whether someone consistently overlooks targets on one side.

1. Line Bisection: Measuring spatial judgment

Users see horizontal lines and are asked to mark their midpoint.

NeuroVision calculates how far each selected point is from the actual midpoint.

By examining the direction and magnitude of these errors across multiple lines, we can study potential spatial biases.

The technology behind NeuroVision

We built NeuroVision using a modern full-stack architecture.

Our technology stack includes:

React and TypeScript for the interactive frontend

Vite for frontend development and builds

Node.js and Fastify for backend services

PostgreSQL for storing assessment records

Custom statistical scoring and probabilistic modeling

Vercel for deployment

One of the most interesting engineering challenges was figuring out how to combine the results of three different assessments.

A missed target, a delayed reaction, and an incorrectly marked midpoint are not directly comparable.

Instead of simply adding scores together, we implemented a probabilistic evidence model.

Each assessment produces a measure of left-right asymmetry and an estimate of uncertainty. The model standardizes these measurements against reference values and combines them to estimate whether the observed pattern leans left, leans right, or appears balanced.

The model can also report inconclusive results when the evidence is insufficient.

This distinction was important to us. We didn't want a single missed target to automatically become a definitive conclusion.

An engineering decision: Why we didn't rely on eye tracking

During development, we also explored webcam-based eye tracking.

Our repository includes experimental gaze-estimation and calibration components.

However, we ultimately chose not to make eye tracking part of our primary three-test assessment.

Webcam-based gaze estimation introduces additional challenges, including calibration, lighting conditions, camera differences, and head movement.

We wanted NeuroVision's main assessments to work through ordinary browser interactions without requiring camera access.

This experience reminded me that engineering isn't just about adding more features. Sometimes, choosing which features not to include is equally important.

Testing our prototype at HackPrinceton

One of the most meaningful parts of the project was moving beyond the interface and collecting preliminary data.

On November 14, 2026, we collected assessment results from 20 healthy adult volunteers.

We used the data to explore reference distributions for the Hollow Rings and Line Bisection assessments.

Eighteen of the twenty participants successfully identified every hollow ring, which also revealed a limitation: the task may be too easy for many healthy users.

Our preliminary analysis did not classify any of these volunteers as having strong evidence of a leftward or rightward attention bias.

However, these results are only an initial feasibility exercise. The reference values were fitted using the same participants, and we have not yet evaluated the system with stroke survivors.

That means we cannot claim NeuroVision accurately detects spatial neglect.

The pilot helped us identify what needs improvement and what future research would require.

What I learned

Hackathons have taught me that building software is only one part of engineering.

With NeuroVision, I learned how important it is to understand the assumptions behind a model, question whether a feature actually improves the experience, and communicate limitations rather than hiding them.

I also gained a deeper appreciation for the intersection of computer science, mathematics, and healthcare research.

What excited me most wasn't just seeing the interface work. It was seeing how interactive tasks could produce structured measurements that we could analyze mathematically.

What's next?

NeuroVision is still a research prototype, and there's plenty we want to explore.

Future work includes collecting more reference data, improving the sensitivity of our assessments, evaluating test-retest reliability, and eventually collaborating with stroke clinicians and researchers.

We would also like to investigate how repeated assessments could help visualize changes in spatial attention over time.

Our long-term goal is to contribute to more accessible ways of studying neurological function, while keeping scientific validation at the center of development.

Explore NeuroVision

Live application: https://neurovision-app-kappa.vercel.app/

Built at HackPrinceton 2026 by Sebastian Villalba, Jonathan Zheng, Ekagra Jain, and Pronita Ghimire.

I'm grateful to my teammates for the collaboration, late-night debugging, and countless decisions that brought NeuroVision together.

HackPrinceton gave us the opportunity to take an idea, challenge our assumptions, and turn it into something we could actually test.

And that's one of my favorite things about hackathons: you arrive with an idea, and leave with a project, new knowledge, and even more questions worth exploring.

mlhacks #hackathon #webdev #healthtech