Inspiration
Recently, I have a growing interest in understanding the brain and how this marvelous, complex biological system can be understood through modeling. A few months ago, I started researching neurodegenerative diseases and found work that framed Alzheimer’s disease as a network-level disorder instead of just isolated issues. However, there was little work on modeling disease progression over time, especially using graph-based and continuous-time methods. This is something that would push work further than standard Euclidean representations, but most importantly, something I can contribute towards. Overall, I am motivated by my interest in machine learning and neurological diseases, hoping to apply my computational skills to improve the quality of life for individuals, such as my close family friend and mentor, who are currently living with this condition.
What it does
To model Alzheimer’s disease as a progressive network disorder, this project portrays each subject as a longitudinal sequence of functional brain networks evolving. For each clinical session, resting-state fMRI (rs-fMRI) time series are extracted from regions of interest (ROIs). Pairwise Pearson correlations between ROI signals are computed to construct functional connectivity matrices at each time point. Each time point is represented as a weighted, undirected graph, where nodes correspond to ROIs and edges encode the strength of functional connectivity. The novelty is that they are organized into subject-specific temporal trajectories that preserve the evolution of network structure across clinical visits. This formulation explicitly captures progressive connectivity disruption rather than simply static representations.
To model continuous-time network evolution under irregular longitudinal sampling, the model uses graph neural ordinary differential equations (GNN-ODEs) in the learning process. This allows the latent trajectories to be modeled smoothly, accommodating missing visits and variable inter-scan intervals common in longitudinal neuroimaging.
Disease-relevant predictions are produced from time-dependent graph representations without collapsing temporal information into flat vectors. The model is trained end-to-end in a supervised setting using gradient-based optimization. The architectural choices are guided by stability, interpretability, and biological plausibility.
It is also built with various interpretability features; overall, this work represents a novel AND significant advancement into longitudinal modeling of Alzheimer's, whilst making the results actually clinically significant. View the Google Doc in the Future Work section for a more indepth methodology.
How I built it + What I learned
I utilized PyTorch and various other ML frameworks to create this model. I first did a bunch of research (which can be seen in the Google Document under Future Works) and utilized this information to create my model.
Through this project, I have developed a strong foundation in programming and machine learning, with the ability to implement and train deep learning models. I have independently designed and built a functional prototype that includes preprocessing pipelines for resting-state fMRI time series, construction of subject-specific functional connectivity graphs, and baseline graph neural network (GNN) models for predictive tasks. This work demonstrates my ability to carry out technically complex computational research independently, from data handling to model design and evaluation.
In addition to these technical skills, I have gained practical experience at the intersection of machine learning and neuroscience, learning how abstract graph-based models can be grounded in real neuroimaging data. I now understand the challenges involved in modeling brain connectivity, handling noisy longitudinal data, and translating domain-specific scientific questions into machine learning formulations.
Accomplishments and Challenges
Accomplishments: The project represents the independent design and implementation of an end-to-end computational system for the simulation of Alzheimer’s disease using resting-state fMRI images. This covers the implementation of the data processing pipeline for the fMRI time series dataset, the design of the functional graph for each subject, as well as the development of the basic graph neural network architecture for the prediction tasks. In all, it shows just how much I learned and what I've accomplished.
The project involved translating a scientific task from the field of neuroscience, specifically longitudinal degeneration of brain networks, into a machine learning task. This allowed me to gain practical experience in modeling brain graphs and dealing with neuroimage-related issues of noisy and high-dimensional data.
In addition to implementation, another aspect where my ability improved over time is in my ability to independently implement machine learning models. This has been achieved through understanding novel ideas and studying relevant literature. Furthermore, I've developed a strong computer science-related mathematical and theoretical framework across a variety of subjects (it was a lot more math than expected).
Challenges: There was a need for a LOT of debugging. It took forever, and I had to learn so much, and it took forever, and mumbles incomprehensibly about the struggles and continues to self-combust...
Future Work
All planned extensions are detailed in the accompanying design document and will be completed before major research competitions (ISEF and JSHS) and the conclusion of the Hack for Health mentorship: https://docs.google.com/document/d/1cy0_JPhTfHO2ONlpRYeG0yKOuKpWstInn_SPx_0A3I0/
Note: All mentions to the above Google Doc throughout this write-up are entirely valid; 80% of it has already been implemented, and there are only a few features that still need polishing. For the innovation/impact and research section, it holds for both my existing model and the improved one I will be working on.
Random Glaze
Thank you so much for organizing this event! This really taught me a lot and gave me an idea for a major project/research paper I plan to expand upon. Thank you once again!
Log in or sign up for Devpost to join the conversation.