🤖 machine learning

A foundation-model approach to pediatric headache classification from rs-fMRI

This study demonstrates that a foundation-model approach (NeuroSTORM) applied to resting-state fMRI data outperforms traditional functional-connectivity methods in classifying pediatric headaches and distinguishing chronic migraine from other subtypes, offering a promising proof of concept for fMRI-based individualized treatment strategies.

Original authors: Guilherme S. Imai Aldeia, Clara Moon, Julie Shulman, Navil Sethna, Allison Smith, Alyssa Lebel, William G. La Cava, Scott Holmes

Published 2026-08-10

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Guilherme S. Imai Aldeia, Clara Moon, Julie Shulman, Navil Sethna, Allison Smith, Alyssa Lebel, William G. La Cava, Scott Holmes

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine your brain is a bustling city where billions of tiny messengers (neurons) are constantly sending text messages to one another. Even when you are just sitting still, daydreaming, or staring at a wall, these messengers are busy chatting in specific patterns. Scientists can listen to these conversations using a special camera called an MRI machine, which takes pictures of the brain's activity without needing you to do anything. Usually, to understand these chats, researchers try to map out who is talking to whom, creating a giant "friendship chart" of the brain. This is helpful, but it's like trying to understand a whole city by only looking at a list of phone numbers; you miss the tone of voice, the speed of the conversation, and the hidden meanings.

Now, imagine a super-smart robot that has read millions of these brain "chat logs" from people of all ages. This robot, known as a "foundation model," has learned to recognize the unique rhythm and style of a healthy brain versus a brain that is in pain, just by looking at the raw video of the activity. This is the exciting corner of science where artificial intelligence meets medicine: teaching computers to spot invisible patterns in our brains that human doctors might miss. Why does this matter? Because headaches are the most common reason kids go to the doctor for a neurological problem, but telling one type of headache from another is often like guessing the flavor of ice cream with your eyes closed. If a computer could listen to the brain's chatter and say, "Ah, this specific rhythm means a migraine," it could help doctors give the right treatment much faster.

The Big Idea: Listening to the Brain's "Raw Feed"

In this study, a team of researchers from Boston Children's Hospital asked a simple but tricky question: Can we teach a computer to tell the difference between a kid with a headache and a kid without one, just by looking at their brain scans? They wanted to see if a new, super-smart AI tool could do a better job than the old, standard way of doing things.

The team gathered data from 110 kids and teenagers, aged 8 to 22. Some were healthy and didn't have headaches, while others were there because they suffered from different kinds of headaches, with chronic migraine being the most common. In total, they looked at 189 brain scans.

The Old Way vs. The New Way

To solve the puzzle, the researchers tried two different strategies, like using two different maps to find a hidden treasure.

The Old Map (Functional Connectivity): The traditional method involves taking the brain scan and turning it into a "friendship chart" (called a functional connectivity matrix). This chart lists how strongly different parts of the brain are talking to each other. It's like summarizing a whole movie into a list of who sat next to whom. The researchers trained standard computer programs to read these charts and guess if the person had a headache.
The New Map (The Foundation Model): The new method used a powerful AI called NeuroSTORM. Instead of turning the brain scan into a summary chart, they fed the raw brain video directly into this AI. Think of NeuroSTORM as a student who has already watched millions of brain movies and learned to recognize the "vibe" of a healthy brain versus a painful one. It doesn't need a summary; it understands the whole story at once.

What They Found

The results were a clear win for the new approach.

The Old Map Struggled: When the computer tried to use the "friendship charts" to tell if a kid had a headache, it did a pretty poor job. It got the answer right about 67% of the time (a score called AUROC of 0.67). That's barely better than flipping a coin. It was like trying to identify a song just by reading the names of the instruments; the computer just couldn't hear the melody.
The New Map Succeeded: When the team used the NeuroSTORM AI to look at the raw brain data, the results were much better. It correctly identified kids with headaches about 82% of the time (AUROC of 0.82) and was even more impressive at spotting the specific signs of pain, with a score of 0.93. This suggests that the AI was able to catch subtle patterns in the brain's activity that the old method completely missed.

Can It Tell Different Headaches Apart?

The researchers then tried a harder challenge: Could the AI tell the difference between a healthy kid, a kid with chronic migraines, and a kid with other types of headaches (like headaches after a virus or a bump on the head)?

Here, the AI did okay, but it wasn't perfect. It was quite good at spotting the kids with chronic migraines (the most common type), but it struggled to tell the difference between migraines and other types of headaches. The overall score for this tricky three-way guess was about 0.69. The authors suggest this is because the other types of headaches are harder to define and might look very similar to each other in the brain, making them a tough puzzle even for a smart AI.

Why This Matters

This study suggests that using these powerful, pre-trained AI models is a promising new way to look at pediatric headaches. It shows that we don't always need to simplify the brain's complex data into simple charts; sometimes, letting the AI listen to the full, raw conversation works much better.

However, the authors are careful to say this is just the beginning. They found that the AI works well for spotting the general presence of a headache, but it still needs more practice to perfectly sort out the specific types. They also noted that the AI is a "black box," meaning we know it works, but we don't fully understand how it sees the patterns yet.

In short, this research lights a path forward. It suggests that with the right AI tools, we might one day be able to use brain scans to help doctors diagnose headaches more accurately and quickly, moving away from just guessing based on how a child describes their pain. But for now, it's a proof-of-concept—a very strong hint that this technology has real potential, not a finished cure.

Technical Summary: A Foundation-Model Approach to Pediatric Headache Classification from Resting-State fMRI

Problem Statement
Headache is the most prevalent neurological disorder in children, significantly impacting quality of life and daily functioning. However, diagnosis relies heavily on subjective self-reporting and semi-structured interviews, which can be challenging for pediatric populations. While machine learning (ML) has been applied to neuroimaging to identify patterns associated with neurological disorders, existing approaches face significant hurdles: many studies rely on small sample sizes, lack standardized pipelines, and struggle with generalizability. Furthermore, prior work has predominantly focused on adult populations and utilized functional connectivity (FC) matrices—derived features representing pairwise correlations between brain regions—as inputs. These derived features often require manual feature selection and may lose complex spatiotemporal information present in raw fMRI data. There is a critical need to evaluate whether modern foundation models, trained on massive datasets, can effectively encode resting-state fMRI (rs-fMRI) data to classify pediatric headache conditions in data-scarce settings without relying on hand-crafted FC features.

Methodology
The study utilized a dataset of 189 rs-fMRI scans from 110 individuals (aged 8–22) recruited from Boston Children's Hospital, comprising healthy controls ( $n=45$ ) and patients with various headache conditions, most notably chronic migraine ( $n=68$ ). The data was collected across two visits.

The authors compared two distinct representation strategies for machine learning classification:

Functional Connectivity (FC) Representation: Raw fMRI data was preprocessed using the CONN toolbox pipeline (including denoising, normalization to MNI152 space, and bandpass filtering). The Harvard-Oxford atlas was used to parcellate the brain into 48 cortical regions of interest (ROIs). Pairwise correlations between these ROIs were computed to generate symmetric $48 \times 48$ FC matrices, resulting in 1,128 unique features.
Foundation Model Embedding Representation: The same preprocessed rs-fMRI volumes were fed directly into NeuroSTORM, a recently proposed foundation model trained on over 28 million fMRI frames. NeuroSTORM utilizes a state-space model (Mamba) architecture instead of standard attention mechanisms. The encoder of the pre-trained model generated a 288-dimensional embedding vector for each scan.

Experimental Design:

Binary Classification: Distinguishing healthy controls from any headache patient.
Multiclass Classification: Distinguishing healthy controls, chronic migraine, and non-chronic headaches (e.g., post-traumatic, post-viral, new daily persistent headache) using a one-vs-all scheme.
Validation Strategy: Due to the limited sample size, the authors employed a leave-one-out (LOO) cross-validation strategy where all visits from a held-out participant were excluded from the training set. This procedure was repeated 100 times with different random seeds to estimate uncertainty.
Classifiers: A range of shallow ML models were tested, including Decision Trees (DT), Logistic Regression (LR with L1/L2 regularization), Random Forests (RF), and Gradient Boosting (GB). For the embedding representation, Logistic Regression with L1 regularization was prioritized based on the assumption that foundation models produce linearly separable representations.

Key Results

Binary Classification Performance: The NeuroSTORM-based approach significantly outperformed traditional FC-based methods.
- NeuroSTORM: Achieved an Area Under the Receiver Operating Characteristic Curve (AUROC) of 0.82 (95% CI: 0.82–0.82) and an Area Under the Precision-Recall Curve (AUPRC) of 0.93 (95% CI: 0.93–0.94).
- FC-Based Models: The best-performing FC model (Logistic Regression with L1) achieved an AUROC of 0.67 and an AUPRC of 0.85. Other FC models (DT, RF, GB) performed near random chance or worse.
- Demographic Baseline: A logistic regression model using only age and sex achieved an AUROC of 0.61, indicating that neuroimaging features provided substantial additional predictive value.
Multiclass Classification: When tasked with distinguishing subtypes (Healthy, Chronic Migraine, Non-Chronic Headache), NeuroSTORM achieved a macro-AUROC of 0.69. Performance was highest for distinguishing healthy controls and chronic migraine but struggled to differentiate non-chronic headache subtypes from chronic migraine.
Robustness: NeuroSTORM maintained consistent performance across sex subgroups, whereas FC-based models showed greater performance degradation when stratified by sex.

Key Contributions

Proof-of-Concept for Foundation Models in Pediatrics: This work presents the first demonstration that a pre-trained foundation model (NeuroSTORM) can effectively encode pediatric rs-fMRI data to classify headache conditions, outperforming traditional FC-based approaches in a limited-data setting.
Direct Encoding vs. Derived Features: The study validates that direct processing of fMRI volumes via foundation models can capture latent representations relevant to headache pathology that are missed by standard FC matrix approaches, which often rely on dimensionality reduction and manual feature engineering.
Subtype Differentiation: The paper extends the scope of ML headache research beyond binary classification to include subtype identification, highlighting both the potential and the current limitations of distinguishing between complex headache phenotypes like chronic migraine and other non-chronic types.

Significance and Claims
The authors claim this work provides a "proof-of-concept" for using foundation models to predict pediatric headache from fMRI without relying on functional connectivity features. They argue that large-scale, pre-trained neuroimaging models offer a practical starting point for clinical ML applications where site-specific datasets are too small to train high-capacity models from scratch.

The paper modestly acknowledges that while the approach shows strong discriminative power for binary classification (headache vs. non-headache) and chronic migraine identification, it currently struggles to differentiate other headache subtypes. The authors attribute this to the inherent complexity of headache phenotypes, potential labeling noise in non-chronic diagnoses, and the fact that the foundation model was pre-trained on datasets dominated by healthy controls and other psychiatric conditions (e.g., ADHD) rather than headache-specific cohorts.

Ultimately, the paper suggests that while these tools are not yet ready for clinical deployment due to the need for interpretability and prospective validation, they represent a significant step toward improving diagnosis based on brain activity and tailoring treatment strategies for pediatric patients.

More like this