Project Story
Inspiration
Early identification of Autism Spectrum Disorder (ASD) is often delayed because current diagnostic methods rely on subjective behavioral evaluations that require trained clinicians. These assessments can take several months to schedule and complete, are costly for families, and are not widely accessible in many regions. At the same time, research in developmental psychology has shown that eye-gaze behavior contains measurable markers of social attention and engagement. Many children with ASD demonstrate consistent patterns such as reduced fixation on faces or social stimuli. However, most existing eye-tracking studies rely on specialized hardware that is not practical for everyday clinical or educational settings. This gap motivated our goal: to determine whether it is possible to detect behavioral markers associated with ASD using standard webcams in a more scalable and accessible way.
What We Built
We developed NeuroGaze, a system that analyzes how a child's eye position changes over time while viewing a set of images and visual scenes. The aim is to characterize patterns in gaze movement that can help support early screening for ASD. The system does not provide a clinical diagnosis. Instead, it identifies whether gaze behavior aligns more closely with patterns commonly observed among children with ASD compared to those without. The overall goal is to assist clinicians, educators, and families in determining whether further professional evaluation may be beneficial.
How We Built It
Our data consists of time-series eye-movement measurements. Each data sample is a sequence of gaze coordinates, pupil measurements, and frame-by-frame attention shifts. Because each sample reflects changes over time, it is not sufficient to treat the data as static points. The temporal order of eye movements is meaningful and must be modeled directly.
We first trained a Long Short-Term Memory (LSTM) network to learn patterns in how gaze changes across short intervals. The LSTM identifies characteristics such as how smoothly the eyes move across regions of a scene, how consistently the viewer returns to areas containing human faces, and how gaze transitions differ between social and non-social elements. The LSTM compresses each sequence into a numerical feature vector that represents the participant's overall gaze behavior.
We then trained an XGBoost model using these feature vectors. XGBoost is well suited for learning structured decision rules from numeric features. In this stage, the model determines whether the patterns identified by the LSTM are more characteristic of ASD or non-ASD gaze behavior. The pipeline therefore uses the LSTM for pattern representation and XGBoost for classification.
Tech Stack
The backend and model training were implemented in Python using scikit-learn for traditional machine learning components and standard deep learning libraries for the LSTM. The interactive interface was built in TypeScript, allowing the system to run in a browser environment. The system is designed to work using standard webcam data, without requiring specialized eye-tracking devices.
Challenges
A primary challenge was dealing with variation in webcam quality, lighting conditions, and head positioning. Unlike laboratory eye-trackers, standard webcams produce noisier input. We addressed this by refining preprocessing steps, stabilizing coordinate estimates over time, and testing model robustness under different recording conditions. Another challenge was ensuring that the model generalized beyond the training dataset and did not rely on dataset-specific artifacts.
What We Learned
We learned that time-series modeling of eye movements can effectively capture patterns relevant to ASD screening, but reliability depends heavily on consistent data collection procedures. We also learned the importance of balancing performance with practicality. A tool intended for real-world early screening must not rely on ideal lab environments. Interdisciplinary collaboration between machine learning, behavioral science, and interface design was essential for making progress on a problem that involves both technical and human-centered considerations.
Impact
NeuroGaze supports early screening by analyzing behavior that may be difficult to recognize without trained expertise. Earlier recognition of atypical social attention patterns can guide families toward professional evaluation sooner, enabling earlier intervention and improved developmental support. The aim is not to replace clinicians but to increase access to screening and reduce delays in the evaluation process.
Log in or sign up for Devpost to join the conversation.