Built the connectome
Converted signals from 200 brain regions into 19,900 unique pairwise connectivity features.
← Back to portfolio
Regeneron ISEF Finalist · Computational Biology and Bioinformatics
Machine Learning Approaches to Brain Connectivity in Autism
Using resting-state fMRI and interpretable machine learning, I investigated subtle autism-related patterns in functional brain connectivity and the research-site effects that challenge real-world generalization.
Influential connections were distributed across the entire brain rather than concentrated in one “autism region.”
My questions began with someone close to me.
This project began with my younger brother. During my sophomore year, he experienced a profound regression, yet his clinical MRI showed no structural abnormalities.
I began wondering what differences might exist not in the appearance of the brain, but in how its regions communicate. Resting-state fMRI offered a way to explore that question through functional connectivity, the synchronization of activity between brain regions while a person is at rest.
From brain signals to a rigorous multisite machine-learning test
I built a complete analysis pipeline using public, de-identified neuroimaging data from ABIDE.
Explored the data
Used PCA and UMAP to examine whether patterns reflected diagnosis or research site.
Tested real-world generalization
Trained logistic-regression models and tested each one on an entirely unseen hospital or scanner site.
Challenged the results
Used permutation testing, confidence intervals, site-balanced analysis, and model interpretation to evaluate reliability.
The signal was important, but the limitations mattered just as much.
Subtle and distributed differences
Autism-related connectivity differences appeared across many brain-network interactions rather than in one isolated brain region.
Site effects weren't negligible
Scanner and research-site differences were substantially stronger than diagnosis effects, making strict multisite validation essential.
Above chance, but not diagnostic
The model performed significantly above chance on unseen sites, but its accuracy was not sufficient for standalone clinical diagnosis.
Biologically meaningful network patterns
The most informative and reproducible connections were centered on interactions between limbic, attention, control, sensory-motor, and default-mode networks, consistent with commonly reported autism-related differences in emotion regulation, sensory processing, and attentional shifting.
What I learned
A model’s accuracy is not enough. You need to ask what information the model might be exploiting, whether the evaluation reflects the real setting, and how uncertainty should shape claims made.
Awards and recognition
Recognized at regional, state, and international levels
International
ISEF
Regeneron International Science & Engineering Fair
- ISEF Finalist
State
GSEF
Georgia Science & Engineering Fair
- Regeneron ISEF Award - Top 4
- Top Ten Grand Award
- Best in Category - Computational Biology & Bioinformatics
- First Honors Award
Special Award
CITADEL Securities Innovation Prize
Recognized for ingenuity and data-driven decision-makingCounty
Regional
Cobb-Paulding Regional Science Fair
- First Place Winner
Explore the project
Interested in the complete research?
The official ISEF page contains the full project presentation and results. The GitHub repository contains the analysis code, reproducibility materials, figures, and technical documentation.
Official Project Page
Regeneron ISEF
View complete project details ↗Source Code and Documentation
GitHub Repository
Explore the analysis pipeline ↗Machine LearningAutismResting-State fMRIFunctional ConnectivityComputational NeuroscienceMultisite GeneralizationInterpretable AI