💻 computer science

Predicting Alzheimer's disease progression using rs-fMRI and a history-aware graph neural network

This paper proposes a history-aware graph neural network model that integrates recurrent neural networks and visit distance features to analyze rs-fMRI data, achieving robust accuracy in predicting the progression of Alzheimer's disease across cognitively normal, mild cognitive impairment, and AD stages, even with irregular or missing clinical visits.

Original authors: Mahdi Moghaddami, Mohammad-Reza Siadat, Austin Toma, Connor Laming, Huirong Fu

Published 2026-04-09

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Mahdi Moghaddami, Mohammad-Reza Siadat, Austin Toma, Connor Laming, Huirong Fu

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 with thousands of neighborhoods (brain regions) constantly sending messages to one another. In a healthy city, the traffic flows smoothly, and the neighborhoods talk to each other in a predictable rhythm. But in Alzheimer's disease, the roads start to crumble, the traffic lights get confused, and the neighborhoods stop talking to each other properly.

This paper is about building a super-smart detective that can look at the "traffic patterns" of this brain city to predict if the city is about to fall into chaos (worsen from mild confusion to full-blown dementia) before the disaster actually happens.

Here is the story of how they built this detective, explained in simple terms:

1. The Problem: The Foggy Crystal Ball

Alzheimer's is a sneaky disease. It starts slowly, often when a person is just "Mildly Cognitively Impaired" (MCI)—like having a little bit of fog in their mind. The big question for doctors is: "Will this person stay in the fog, or will the fog turn into a storm?"

Currently, it's very hard to tell. Doctors usually have to wait and see, which is like trying to predict a storm by only looking at the sky after the rain has started. We need a way to see the storm coming earlier.

2. The Clues: The Brain's "Phone Calls"

The researchers used a special type of brain scan called rs-fMRI. Think of this as a high-tech microphone that listens to the brain while a person is just resting (not doing any puzzles or tasks).

The Analogy: Imagine the brain regions are people in a room. Even when they aren't talking loudly, they are whispering to each other. The scan records these whispers.
The Map: The researchers turned these whispers into a map of connections (a graph). If Neighborhood A talks to Neighborhood B a lot, they draw a thick line between them. If they stop talking, the line gets thin or disappears.

3. The Detective: The "History-Aware" AI

Most computer programs look at just one picture of the brain and guess. But the human brain changes over time. A single photo isn't enough; you need a movie.

The researchers built a new type of AI called a History-Aware Graph Neural Network (HA-GNN). Here is how it works:

The Graph Part (The Map Reader): This part of the AI is great at looking at the "map" of brain connections. It knows that if the "memory neighborhood" stops talking to the "language neighborhood," that's a bad sign.
The History Part (The Time Traveler): This is the special sauce. Real life is messy. People don't always go to the doctor every 6 months exactly. Some come every 3 months, some every 18 months, and sometimes they miss an appointment.
- The Analogy: Imagine a detective trying to solve a case by looking at a suspect's diary. Some days the diary has entries; other days it's blank. The AI is smart enough to say, "Okay, there was a 6-month gap here, and a 2-year gap there. I need to weigh those gaps differently to understand the story."
The "Recurrent" Brain: The AI uses a special memory bank (called an RNN) to remember what the brain looked like in the past visits. It connects the dots between Visit 1, Visit 2, and Visit 3 to see the trend.

4. The Training: Learning from Mistakes

To teach this AI, they used data from 303 people who had multiple brain scans over time.

The "Pre-training" Trick: Before teaching the AI to predict the future, they first taught it to just identify what stage a person was in right now (Normal, Mild, or Severe). It's like teaching a student to read a single sentence before asking them to write a whole novel. This helped the AI understand the "vocabulary" of brain diseases first.
The Challenge: The data was unbalanced. Most people stayed stable (didn't get worse), and only a few got worse. It's like trying to teach a dog to find a specific rare bird when 90% of the birds in the forest are pigeons. The researchers used a special math trick (called "Focal Loss") to force the AI to pay extra attention to the rare "worsening" cases.

5. The Results: A Winning Prediction

The AI was put to the test, and the results were impressive:

Overall Accuracy: It correctly predicted whether a person would get worse or stay stable about 83% of the time.
The Hard Part: Predicting who would go from "Mildly Confused" to "Severe" is the hardest task. The AI got this right 68.8% of the time.
Why this matters: In the world of Alzheimer's research, being able to spot the transition from "Mild" to "Severe" is like spotting a crack in a dam before it bursts. If we can catch it early, doctors can start treatments to slow it down.

6. The Catch (Limitations)

The authors are honest about what their detective can't do yet:

Small Class Size: They only had a few "converter" cases (people who got worse), so the AI needs more data to become a master detective.
One Type of Clue: Right now, the AI only looks at the "whispers" (functional scans). It doesn't look at the "building structure" (structural scans) or the person's genetics. Combining all these clues might make it even smarter.
The "Black Box": The AI gives an answer, but it doesn't always explain why. Future versions need to point to the specific brain neighborhoods that caused the alarm.

The Bottom Line

This paper presents a new tool that acts like a time-traveling brain map reader. By combining a map of brain connections with a memory of a patient's history (even if the visits were irregular), it can predict the future of Alzheimer's with surprising accuracy.

It's not a cure yet, but it's a powerful flashlight that could help doctors catch the disease earlier, giving patients more time to slow it down and live better lives.

1. Problem Statement

Alzheimer's Disease (AD) is a progressive neurodegenerative disorder with no current cure. Early detection is critical for interventions that can slow progression. The core challenge addressed in this study is longitudinal prediction: determining whether a subject will transition to a more severe stage of cognitive impairment (e.g., from Cognitively Normal [CN] to Mild Cognitive Impairment [MCI], or MCI to AD) at their next clinical visit.

Key challenges in this domain include:

Data Complexity: Resting-state functional MRI (rs-fMRI) data is high-dimensional and forms complex network structures.
Longitudinal Irregularities: Clinical visits occur at irregular time intervals, and subjects have varying numbers of visits (some missing data).
Class Imbalance: There are significantly more "stable" subjects (no progression) than "converter" subjects (progression), making it difficult for models to learn rare transition patterns.
Limitations of Existing Models: Most Graph Neural Network (GNN) approaches focus on cross-sectional (single-timepoint) data and fail to utilize the temporal history of a patient's visits.

2. Methodology

The authors propose a History-Aware Graph Neural Network (HA-GNN) model that integrates spatial brain network analysis with temporal sequence modeling.

Data Source and Preprocessing

Dataset: Data was sourced from the Alzheimer's Disease Neuroimaging Initiative (ADNI), specifically the TADPOLE dataset.
Subjects: 303 participants with at least two rs-fMRI scans and one structural MRI (sMRI).
Classes:
- Stable: Diagnosis remains the same between the last two visits.
- Converter: Diagnosis transitions to a more severe stage (CN→MCI, CN→AD, or MCI→AD).
Preprocessing: Used the fMRIPrep pipeline (skull stripping, segmentation, normalization to MNI space, motion correction, slice-timing correction).
Feature Extraction:
- The brain was parcellated into 100 Regions of Interest (ROIs) using the Schaefer atlas.
- Time series were extracted for each ROI, and a Functional Connectivity (FC) matrix was computed using Pearson correlation, resulting in a $100 \times 100$ graph for each visit.

Model Architecture (HA-GNN)

The model combines a Graph Convolutional Network (GCN) for spatial feature extraction and a Recurrent Neural Network (RNN) for temporal modeling.

GCN Component (Spatial):
- Based loosely on the BrainGNN architecture.
- Consists of two convolutional blocks, each containing: GraphSAGE layer $\rightarrow$ Graph Normalization $\rightarrow$ Dropout $\rightarrow$ TopK Pooling.
- Pretraining Strategy: The GCN is first pretrained on a single-visit, 3-class classification task (CN, MCI, AD) to learn robust representations of functional connectivity patterns before being integrated into the full longitudinal model.
RNN Component (Temporal):
- Processes the sequence of FC matrices from a subject's visit history ( $v_1, v_2, \dots, v_n$ ).
- Tested architectures: LSTM, GRU, and standard RNN.
- Input Augmentation: To handle irregular time gaps, the time distance (in months) between consecutive visits is explicitly concatenated to the feature vector before being fed into the RNN.
Output: A binary classification predicting "Stable" vs. "Converter" for the next visit.

Training and Evaluation

Loss Function: Focal Loss ( $\alpha=0.9, \gamma=3$ ) was employed to address the severe class imbalance between stable and converter subjects.
Validation: 5-fold cross-validation.
Metrics: Accuracy, Balanced Accuracy (BA), and AUC-ROC.

3. Key Contributions

Novel Architecture: Introduction of the HA-GNN, a hybrid model that effectively fuses GNNs (for spatial brain topology) and RNNs (for temporal disease progression) to handle longitudinal rs-fMRI data.
Handling Irregularity: Explicit incorporation of visit time intervals as input features, allowing the model to function robustly despite irregular clinical visit schedules and missing data points.
Two-Stage Training: Implementation of a pretraining strategy where the GCN learns disease-state representations from single visits before the full model learns temporal transitions, improving convergence and performance.
Robustness to Imbalance: Successful application of Focal Loss to achieve competitive performance on a dataset with a 5:1 ratio of stable to converter subjects.

4. Results

The model was evaluated on the 303-subject dataset (1089 total scans).

Overall Performance:
- Accuracy: 82.9% (± 5.8%).
- Balanced Accuracy (BA): 77.1% (± 11.4%).
- AUC-ROC: 0.852 (± 0.065).
Specific Transition Performance:
- CN to MCI Conversion: 68.8% accuracy (a notoriously difficult task).
- MCI to AD Conversion: 67.6% accuracy.
Architecture Comparison:
- LSTM outperformed GRU and standard RNN, achieving the highest BA (77.1%) and overall accuracy. This suggests LSTM is superior at capturing long-term dependencies across irregular time gaps.
- Pretraining Impact: Removing the GCN pretraining step caused a significant drop in performance (Accuracy dropped to 72.3%, BA to 59.4%), validating the importance of the pretraining strategy.

5. Significance and Future Work

Clinical Impact: The model demonstrates that rs-fMRI functional connectivity patterns contain predictive information even in the earliest stages of cognitive decline (CN to MCI). This supports the viability of using rs-fMRI for early intervention strategies.
Methodological Advance: The study proves that combining graph-based spatial modeling with history-aware temporal modeling is a powerful approach for neurodegenerative disease progression, outperforming methods that rely solely on tabular features or cross-sectional data.
Limitations:
- Dataset Size: The dataset (303 subjects) is relatively small for deep learning, leading to high variance in Balanced Accuracy across folds.
- Unimodal: The model currently uses only rs-fMRI; future work should integrate sMRI, diffusion MRI, and cognitive/genetic data.
- Interpretability: The "black box" nature limits clinical trust; future iterations need attention mechanisms to identify specific brain regions driving predictions.
- Generalizability: Validation is currently limited to the ADNI dataset; external validation on diverse, real-world clinical populations is required.

In conclusion, this paper presents a robust framework for predicting AD progression by effectively modeling the complex spatiotemporal dynamics of the brain, offering a promising tool for early diagnosis and monitoring.