Adolescent Mood Prediction From fMRI Network Maps

ISEF Category: Behavioral and Social Sciences

Subcategory: Behavioral Neuroscience  ·  Difficulty: Advanced  ·  Setup: Home Setup  ·  Time: Full Year

The Hook

A single mood question can hide a lot. The brain data behind it can act like a map, if you know how to read it. You can use public fMRI summaries and a decoding tool to see which affective networks line up with different self-report mood items. That makes this a real research project, not just a coding exercise.

What Is It?

This project asks whether self-report mood items in teens line up with brain activation patterns that already have known labels, like reward, threat, attention, or control. Think of each mood item as a short clue. Your job is to see which brain network patterns best match that clue.

You are not scanning brains yourself. You are working with public summary maps from OpenNeuro and meta-analytic decoding from Neurosynth. Summary maps are cleaned-up group-level brain images. Meta-analytic decoding means using many published studies to estimate what a brain pattern usually relates to. The goal is to build an interpretable classifier, which means a model that can explain its own choices instead of acting like a black box.

Why This Is a Good Topic

This is a strong science fair topic because it is measurable, computer-based, and full of real scientific tradeoffs. You can test model performance, compare feature sets, and check whether the classifier generalizes across mood items or study datasets. The topic connects to mental health research, adolescent development, and brain-behavior links, so the real-world angle is clear. You can learn data cleaning, feature engineering, validation, and basic machine learning without needing a wet lab.

Research Questions

• How does using canonical affective-network features change classification accuracy compared with using whole-brain summary features? ?
• What is the effect of different mood item wordings on how well the classifier predicts network labels? ?
• Does a model trained on one public dataset generalize to a separate adolescent dataset? ?
• To what extent do reward, threat, and cognitive control patterns separate the mood items from each other? ?
• Which feature selection method gives the most interpretable and stable classifier across cross-validation folds? ?
• What is the effect of adding demographic covariates such as age or sex on model performance? ?

Basic Materials

• Laptop or desktop computer with at least 16 GB RAM.
• Internet access for downloading public datasets and documentation.
• Python installed with Jupyter Notebook or VS Code.
• Public OpenNeuro dataset downloads.
• Neurosynth term decoding outputs.
• Spreadsheet software for tracking samples and labels.
• External storage for large neuroimaging files.

Advanced Materials

• Workstation with 32 GB RAM or more.
• Python neuroimaging stack, including Nilearn and NiBabel.
• Statistical package for model comparison and permutation testing.
• FreeSurfer or fMRIPrep outputs if you want to work from processed images.
• High-capacity storage for multiple OpenNeuro datasets.
• GPU access if you test a neural network baseline.
• Version control repository for code and analysis logs.

Software & Tools

• Python: Runs data cleaning, feature extraction, modeling, and evaluation.
• Jupyter Notebook: Keeps your analysis steps readable and easy to rerun.
• Nilearn: Loads fMRI maps and turns them into machine-learning features.
• NiBabel: Reads and writes neuroimaging file formats.
• scikit-learn: Builds interpretable classifiers and cross-validation pipelines.

Experiment Steps

1. Define the exact mood items and brain labels you want to compare.
2. Choose one public dataset as your training set and a separate one as your test set.
3. Decide how you will turn brain maps into features, such as regions, networks, or decoded terms.
4. Build a baseline model first, then compare it with your interpretable classifier.
5. Plan cross-validation and holdout testing so you can check generalization, not just fit.
6. Set up controls that test whether the model is learning true affective structure instead of dataset quirks.

Common Pitfalls

• Mixing datasets with different preprocessing pipelines, which can make the classifier learn scanner differences instead of mood patterns.
• Using too many features for too few samples, which inflates accuracy and hides overfitting.
• Skipping a true holdout set, which makes cross-validation look better than real-world performance.
• Treating Neurosynth terms as ground truth, which can blur the difference between decoded associations and actual psychological meaning.
• Ignoring class imbalance across mood items, which can make one dominant label look stronger than it really is.

What Makes This Competitive

A competitive version of this project goes beyond basic accuracy. You would compare several feature representations, test whether results hold across independent datasets, and report uncertainty with permutation tests or confidence intervals. A strong entry would also explain why the model makes each prediction, not just how often it gets the label right. That kind of analysis shows real judgment, not just coding speed.

Project Variations

• Use anxiety-related mood items instead of general mood items to see whether threat and control networks separate more cleanly.
• Compare adolescent data with adult public datasets to test whether the same classifier works across age groups.
• Swap the feature set from network summary maps to region-of-interest averages and see which version is easier to interpret.
• OpenNeuro: A public dataset repository where you can find processed fMRI studies and download subject-level or summary data.
• Neurosynth: A free platform for meta-analytic brain decoding and term-based association maps.
• NIH PubMed: Search for review articles on adolescent affect, brain networks, and neuroimaging methods.
• NCBI Bookshelf: Find free textbook chapters on neuroscience, statistics, and machine learning basics.
• MIT OpenCourseWare: Look for free courses in statistics, data analysis, and machine learning.

Learn More

• OpenNeuro: A public dataset repository where you can find processed fMRI studies and download subject-level or summary data.
• Neurosynth: A free platform for meta-analytic brain decoding and term-based association maps.
• NIH PubMed: Search for review articles on adolescent affect, brain networks, and neuroimaging methods.
• NCBI Bookshelf: Find free textbook chapters on neuroscience, statistics, and machine learning basics.
• MIT OpenCourseWare: Look for free courses in statistics, data analysis, and machine learning.