Predicting Depression Treatment Response From Brain Scans

ISEF Category: Cellular and Molecular Biology

Subcategory: Neurobiology  ·  Difficulty: Advanced  ·  Setup: University Lab  ·  Time: Full Year

The Hook

A brain scan can hide a pattern that does not show up on a regular image. In depression research, those patterns may help predict who gets better after treatment. You can test that idea with public data, even if you have never worked with neuroscience before. The hard part is not collecting scans. It is finding a signal that still works on new data.

What Is It?

Resting-state functional connectivity means how different brain regions seem to “talk” to each other when a person is not doing a task. Researchers measure this from fMRI scans, then turn the brain into a network, where regions are nodes and connections are edges. Your project asks whether that network contains a pattern that predicts treatment response in adolescent depression.

Think of it like city traffic. One road tells you little. The whole map shows where flow gets stuck, where shortcuts form, and where movement changes under stress. A graph-isomorphism network is a kind of machine learning model that reads those brain maps as graphs, then learns which network patterns matter most. You are not trying to diagnose depression. You are trying to see whether baseline brain connectivity can help predict which patients improve after treatment.

This kind of project sits at the border of neuroscience, data science, and medical prediction. The public datasets give you a real research question and real clinical labels. Your job is to clean the data, define the network carefully, and test whether your model predicts better than chance on unseen participants.

Why This Is a Good Topic

This is a strong science fair topic because it starts with a real biomedical question and gives you a clear yes-or-no outcome. You can test whether a brain connectivity pattern predicts treatment response, then compare that model against simpler baselines. The project connects to mental health care, which makes the stakes easy to explain. You can also learn a lot of transferable skills, like dataset cleaning, graph building, model validation, and statistical testing.

Research Questions

• How does the choice of brain parcellation affect prediction of treatment response in adolescent depression?
• What is the effect of using different functional-connectivity thresholds on model accuracy?
• Does a graph-isomorphism network predict treatment response better than logistic regression on the same dataset?
• To what extent do sex, age, or medication status change the predictive value of resting-state connectivity features?
• Which brain network communities contribute most to correct response prediction?
• How does training on one OpenNeuro cohort and testing on another affect generalization?
• What is the effect of removing noisy scans or motion-heavy participants on model performance?

Basic Materials

• Laptop or desktop computer with at least 16 GB RAM.
• Reliable internet access for downloading public datasets.
• External hard drive or cloud storage for large MRI files.
• Python installed with scientific libraries.
• Jupyter Notebook or another notebook editor.
• Spreadsheet software for tracking participants, labels, and model results.
• OpenNeuro account or access to OpenNeuro public datasets.
• Basic headset or speakers for tutorial videos and meetings.

Advanced Materials

• Workstation with a dedicated GPU for model training.
• Access to a university or school computing cluster.
• Neuroimaging software such as FSL, AFNI, or nilearn.
• Graph neural network library such as PyTorch Geometric or DGL.
• Container software such as Docker or Singularity for reproducible runs.
• Statistical software for permutation testing and mixed-effects analysis.
• Version-controlled code repository.
• Brain atlas files for parcellation and network construction.

Software & Tools

• Python: Handles data cleaning, feature building, model training, and statistical analysis.
• nilearn: Loads neuroimaging data, builds connectivity matrices, and supports brain-mapping workflows.
• PyTorch: Trains the graph-isomorphism network and compares it with baseline models.
• PyTorch Geometric: Provides graph neural network layers and data structures for connectome graphs.
• Jupyter Notebook: Lets you document each analysis step and keep code, notes, and figures together.

Experiment Steps

1. Define the prediction target, the brain scan type, and the exact treatment-response label you will use.
2. Choose one public dataset, then confirm that its age range, diagnosis criteria, and outcome labels fit your question.
3. Decide how you will turn each participant's scan into a graph, including the atlas, node definition, and edge rule.
4. Build a baseline model first, so you can tell whether the graph network adds real value.
5. Plan your validation strategy before training, including how you will separate training, validation, and test participants.
6. Choose the main metrics, then decide which plots and statistical tests will prove the model works better than chance.

Common Pitfalls

• Using the same participants for tuning and final testing, which makes the accuracy look better than it really is.
• Mixing datasets with different scan protocols without correcting for site effects, which can create fake patterns.
• Keeping motion-heavy scans in the analysis, which can make head motion look like disease signal.
• Building the graph with an arbitrary atlas and never checking whether the result changes under another parcellation.
• Reporting only the best model run, which hides how unstable the prediction is across random seeds.

What Makes This Competitive

A competitive version of this project does more than fit a model. It tests whether the prediction survives stricter validation, such as a held-out dataset, repeated random splits, or permutation testing. It also compares your graph model against simpler baselines and checks whether the signal holds across subgroups. The strongest projects explain not just that a model works, but why it works and where it fails.

Project Variations

• Use a different public cohort of adolescent or young adult depression scans and test whether the same connectivity signature generalizes.
• Replace the graph-isomorphism network with a simpler graph convolution model and compare performance on the same data.
• Focus on one brain network, such as the default mode network, and test whether within-network connectivity predicts response better than whole-brain features.

Learn More

• OpenNeuro: Search for public resting-state fMRI datasets and download participant-level imaging files and metadata.
• NIH National Library of Medicine PubMed: Search for review articles on resting-state connectivity, adolescent depression, and treatment response.
• NIH NIMH Data Archive: Find mental health imaging datasets and study documentation for clinical neuroscience projects.
• MIT OpenCourseWare: Search for introductory courses on machine learning, graph theory, or computational neuroscience.
• nilearn documentation: Learn how to load neuroimaging data, build connectivity matrices, and make brain plots.
• NeuroImage: Search the journal for papers on connectome-based prediction, graph neural networks, and resting-state fMRI.