🤖 machine learning

NeuroMambaLLM: Dynamic Graph Learning of fMRI Functional Connectivity in Autistic Brains Using Mamba and Language Model Reasoning

NeuroMambaLLM is an end-to-end framework that integrates dynamic latent graph learning and Mamba-based temporal modeling of fMRI data with a frozen LLM enhanced by LoRA, enabling both autism diagnostic classification and the generation of clinically meaningful textual reports through adaptive connectivity analysis.

Original authors: Yasaman Torabi, Parsa Razmara, Hamed Ajorlou, Bardia Baraeinejad

Published 2026-05-27

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Yasaman Torabi, Parsa Razmara, Hamed Ajorlou, Bardia Baraeinejad

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 as a bustling city with thousands of neighborhoods (brain regions) constantly talking to each other. In a healthy city, these conversations flow smoothly, changing quickly depending on what's happening. But in the "city" of an autistic brain, the paper suggests the traffic patterns and conversations might be different—sometimes too loud in one area, or the connections between neighborhoods might change too fast or too slow.

For a long time, scientists tried to map these conversations by taking a single, static "snapshot" of the city. They calculated the average friendship between neighborhoods over the whole day. The problem? This is like judging a whole movie by looking at just one frozen frame. You miss all the action, the plot twists, and the fleeting moments that actually tell the story.

Enter NeuroMambaLLM: The "Movie Director" AI

The researchers created a new tool called NeuroMambaLLM. Think of it as a three-part team working together to watch the brain's "movie" and write a review about it.

1. The Dynamic Detective (Dynamic Graph Learning)

Instead of taking a frozen snapshot, this part of the system watches the brain's activity second-by-second. It uses a special technique to figure out which neighborhoods are talking to each other right now.

The Analogy: Imagine a detective who doesn't just look at a phone book of who usually calls whom. Instead, they listen to the live phone lines to see who is actually talking at this exact moment. This allows the system to catch "transient" (short-lived) connections that static maps miss, like a sudden burst of conversation between two distant parts of the brain.

2. The Efficient Time-Traveler (Mamba)

Once the system sees these changing connections, it needs to remember the whole story without getting overwhelmed. Older AI models (like Transformers) try to remember everything at once, which is like trying to hold a library of books in your head—it gets heavy and slow.

The Analogy: This system uses Mamba, which is like a super-efficient time-traveler. It can scan a very long movie (a long brain scan) quickly, remembering the important plot points and ignoring the boring parts (like when the camera shakes or the actor moves too much). It does this in "linear time," meaning it stays fast and light even when the movie is very long.

3. The Storyteller (Large Language Model)

The final piece is a powerful language AI (based on LLaMA-3). Usually, these AIs are great at writing stories or answering questions but bad at reading brain scans.

The Analogy: The researchers taught this language AI to "speak brain." They didn't retrain the whole AI (which would be like teaching a human to read a new language from scratch). Instead, they gave it a few "translator notes" (called LoRA modules). Now, when the system sees the brain's dynamic patterns, it doesn't just spit out a "Yes/No" answer. It writes a short, readable report explaining why it thinks the brain looks the way it does.

What Did They Find?

The team tested this on a large collection of brain scans from people with Autism Spectrum Disorder (ASD) and those without (the ABIDE dataset).

The Score: The system correctly identified autism cases about 72% of the time. This is comparable to the best current methods, but with a major bonus: it explains its reasoning.
The "Why": When the system looked at the brains, it found specific patterns that match what scientists already know about autism. For example, it noticed that certain areas involved in social perception and language (like the "superior temporal sulcus") were acting differently. It also saw that the connections between the front and back of the brain were often less integrated in the autistic group.
The Output: Instead of just saying "Autism," the AI generated a sentence like: "The model identifies atypical temporal interactions involving frontal and temporal regions, with reduced long-range integration..." This makes the result much more useful for doctors who need to understand the mechanism, not just the label.

The Bottom Line

NeuroMambaLLM is a new way to look at the brain that treats it like a dynamic, moving movie rather than a static photo. By combining a fast memory system (Mamba) with a smart language writer (LLM), it can not only diagnose autism with decent accuracy but also tell a clear story about what is happening inside the brain, bridging the gap between complex data and human understanding.

Technical Summary: NeuroMambaLLM

Problem Statement
Current neuroimaging analysis for Autism Spectrum Disorder (ASD) faces two primary limitations. First, most methods rely on Static Functional Connectivity (sFC), which aggregates fMRI data into a single correlation matrix. This approach obscures transient neural dynamics and time-varying interactions that are critical for understanding neurodevelopmental disorders. Second, while Large Language Models (LLMs) offer powerful semantic reasoning, their integration with brain connectivity models is nascent. Existing "Brain-to-Text" frameworks (e.g., FCN-LLM, BrainLLM) typically map static connectivity matrices to LLMs, limiting reasoning to average biomarkers rather than specific temporal events. Furthermore, deep learning models for long fMRI sequences often suffer from the quadratic computational cost of Transformers or the forgetting issues of recurrent networks, making end-to-end modeling of long, noisy time series impractical.

Methodology
The paper proposes NeuroMambaLLM, an end-to-end framework that shifts the paradigm from static graph-to-text modeling to Dynamic Latent Signal-to-Text modeling. The architecture consists of three coupled stages:

Dynamic Latent Graph Inference (LGI):
Instead of using precomputed correlation matrices or fixed structural priors, the model learns time-resolved functional connectivity directly from raw Blood-Oxygen-Level-Dependent (BOLD) time series. A differentiable module maps instantaneous BOLD signals into a latent feature space using 1D convolutions and self-attention. Pairwise similarity between these latent node embeddings generates a time-varying adjacency matrix ( $G_t$ ), capturing instantaneous functional coupling and non-linear dependencies without heuristic sliding windows.
Selective State-Space Temporal Modelling (Mamba):
The graph-filtered brain states are processed by a Mamba architecture, a Selective State-Space Model (SSM). Unlike Transformers, Mamba processes signals in linear time with respect to sequence length, avoiding quadratic memory growth. It employs a hardware-aware selective scan mechanism where state transition and input projection matrices are generated as deterministic functions of the current input. This allows the model to adaptively select relevant temporal information, effectively suppressing motion-related artifacts and physiological noise while capturing long-range dependencies in fMRI data.
Cross-Modal Alignment with LLM:
The resulting sequence of latent temporal states is compressed into a fixed set of "brain-summary tokens." These tokens are projected into the embedding space of a frozen LLaMA-3-8B language model. Alignment is achieved via Low-Rank Adaptation (LoRA) modules inserted into the LLM's attention layers, keeping the base model frozen for parameter efficiency. The LLM receives these brain tokens concatenated with diagnostic prompts, enabling it to perform diagnostic classification and generate clinically meaningful textual reports grounded in the specific temporal patterns of the input.

Key Contributions

Dynamic Functional Graph Learning: Introduction of an end-to-end encoder that infers adaptive, time-resolved functional connectivity directly from raw BOLD signals, capturing transient neural interactions missed by static analyses.
Efficient Temporal Modelling: Application of the Mamba selective state-space architecture to model long-range fMRI dependencies with linear computational complexity, adaptively suppressing noise without the cost of Transformers.
LLM Reasoning Framework: A novel alignment strategy using LoRA to project dynamic brain graphs into a frozen LLM's embedding space. This enables joint diagnostic classification and language-based reasoning, producing interpretable textual explanations rather than black-box predictions.

Experimental Results
Experiments were conducted on the Autism Brain Imaging Data Exchange (ABIDE I) dataset, utilizing resting-state fMRI data from 39 Regions of Interest (ROIs) truncated to 100 time points.

Performance: NeuroMambaLLM achieved a test accuracy of 72.12%, outperforming recent state-of-the-art methods including BoLT (71.28%), SwinT (68.59%), and static graph-based approaches like BrainGNN (69.31%).
Metrics: The model demonstrated a precision of 0.8022 and an F1-score of 0.6931.
Ablation Studies:
- Replacing the dynamic graph with a static formulation ("Static MambaLLM") reduced accuracy to 66.50%, confirming the necessity of adaptive graph inference.
- Freezing the LLM parameters without LoRA adaptation ("Frozen MambaLLM") resulted in lower performance (68.20%) compared to the LoRA-tuned version, validating the value of lightweight fine-tuning.
- Comparisons of temporal backbones (GRU, TCN, Transformers, S4) showed Mamba provided the strongest performance.

Significance and Claims
The paper claims that NeuroMambaLLM bridges the gap between dynamic neurobiological processes and computational feasibility. Its primary significance lies in moving beyond "black-box" predictions toward clinically meaningful explanations. By grounding LLM outputs in causally relevant, temporally specific neural patterns, the framework provides interpretable reports that identify atypical interactions (e.g., in the left superior temporal sulcus and frontotemporal regions) consistent with known ASD neurobiology.

The authors emphasize that the language model is not introduced solely to maximize classification accuracy but to support structured interpretation of dynamic brain patterns for domain experts. The model successfully demonstrates that dynamic latent graph learning combined with efficient state-space modeling and LLM reasoning can yield competitive diagnostic performance while offering a level of interpretability (via generated text summaries and saliency maps) that static models lack. Future work is noted to focus on generalization across additional datasets and refining prompt constraints to mitigate potential hallucinations.