MIRAGE: Knowledge Graph-Guided Cross-Cohort MRI Synthesis for Alzheimer's Disease Prediction

MIRAGE is a novel framework that leverages a Biomedical Knowledge Graph and a frozen 3D U-Net decoder to distill EHR data into a latent diagnostic representation, enabling accurate Alzheimer's disease prediction in cohorts lacking MRI scans without performing computationally expensive 3D voxel reconstruction.

The Gist

The Big Problem: The "Missing Puzzle Piece"

Imagine you are trying to diagnose a patient with Alzheimer's disease. To get the full picture, doctors usually need two things:

1. The Medical Chart (EHR): A list of symptoms, age, memory test scores, and history. This is like reading a textbook summary of the patient's life.
2. The Brain Scan (MRI): A 3D photo of the brain showing if parts of it have shrunk. This is like looking at the actual house to see if the walls are crumbling.

The Catch: MRI scans are expensive, time-consuming, and not everyone has one. Many patients only have the "textbook summary" (the chart) but no "photo of the house" (the scan). Without the photo, it's hard to be sure about the diagnosis.

The Old Way vs. The MIRAGE Way

The Old Way (The "Bad Artist"):
Previously, scientists tried to use AI to draw a brand-new brain scan from scratch, just based on the text in the medical chart.

• The Analogy: Imagine asking an artist to paint a detailed, realistic portrait of a stranger based only on a written description like "tall, blue eyes, sad." The artist might draw a face, but it often looks weird, blurry, or completely wrong because they are guessing the details. In medicine, a "wrong" brain scan is dangerous.

The MIRAGE Way (The "Smart Detective"):
The MIRAGE team realized they didn't need to draw a perfect new brain from thin air. Instead, they needed to find the right existing brain that matches the patient's story and then tweak it slightly to fit the specific details.

They call this "Anatomy-Guided Latent Distillation." Let's break that down with a simpler story.

---

How MIRAGE Works: The Three-Step Recipe

1. The "Medical Library" (The Knowledge Graph)

First, MIRAGE builds a massive, super-connected library of medical knowledge.

• The Analogy: Think of this as a giant social network for diseases and symptoms. If a patient has "memory loss" and "confusion," the system knows these are linked to "Alzheimer's." It connects the patient's messy medical notes to a clean, organized map of medical facts. This helps the AI understand the patient's story even if the notes are messy or incomplete.

2. The "Ghost Architect" (The Frozen 3D U-Net)

The researchers took a pre-trained AI that is already an expert at looking at real brain scans. They "froze" it, meaning they locked its brain so it couldn't learn new things, but it could still act as a strict teacher.

• The Analogy: Imagine a master architect who knows exactly what a healthy house looks like. You can't change their rules, but you can ask them to check your blueprints. If your blueprint looks weird (like a door in the ceiling), the architect says, "No, that's not right."

3. The "Tweaking Process" (Cross-Cohort Synthesis)

This is the magic part. When MIRAGE meets a patient with only a medical chart:

1. It uses the Medical Library to find other patients who have similar stories.
2. It grabs the real brain scans of those similar patients.
3. It asks the Ghost Architect to check: "Does this patient's story match the shape of these real brains?"
4. The AI adjusts the patient's "digital brain" until it looks like a real brain that fits their specific symptoms.

• The Analogy: Instead of painting a new house from scratch, MIRAGE finds a house that looks 90% like the patient's needs, then uses the "Ghost Architect" to tweak the windows and doors so it matches the patient's specific medical history perfectly.

Why This is a Game-Changer

The paper tested this on thousands of patients. Here is what happened:

• Better Diagnosis: By using this "tweaked" brain scan, the AI could predict Alzheimer's 13% better than just looking at the medical chart alone.
• No Hallucinations: Because MIRAGE doesn't try to invent a brain from nothing, it doesn't create fake, scary-looking tumors or weird shapes. It creates a "plausible" brain that fits the biology.
• Speed: It skips the heavy, slow process of generating a full 3D image pixel-by-pixel. Instead, it creates a "diagnostic shortcut" (a compressed digital representation) that is fast and accurate.

The Bottom Line

MIRAGE is like a smart translator. It takes the messy, incomplete story of a patient's life (their medical records) and translates it into the language of brain anatomy (MRI scans) without ever needing to see the actual brain.

It doesn't guess; it infers. It uses the collective knowledge of thousands of other patients to fill in the missing pieces, ensuring that even if a patient can't afford or access an MRI, their doctor can still get a clear, reliable picture of what's happening inside their brain.

Technical Summary

1. Problem Statement

The paper addresses a critical bottleneck in Alzheimer's Disease (AD) diagnosis: modality missingness. While effective AD diagnosis relies on multimodal data combining structural Magnetic Resonance Imaging (MRI) and Electronic Health Records (EHR), MRI scans are expensive and often unavailable in large patient cohorts.

• The Challenge: Existing methods attempt to synthesize full 3D MRI volumes from sparse, high-dimensional tabular EHR data. This is an ill-posed problem because mapping low-dimensional clinical records to complex 3D anatomical structures often leads to "posterior collapse," anatomical hallucinations, and severe clinical risks.
• The Goal: Instead of generating visually perfect but potentially hallucinated 3D images, the authors aim to bridge the modality gap by transferring structural priors from MRI-rich cohorts to EHR-only cohorts. The objective is to generate a "diagnostic-surrogate" representation that captures biologically plausible pathological semantics without the computational cost of full voxel reconstruction during inference.

2. Methodology: MIRAGE Framework

MIRAGE reframes the missing-MRI problem as an anatomy-guided cross-modal latent distillation task. The framework operates in three main phases:

A. MRI Representation Learning (Pre-training)

• A 3D U-Net Autoencoder is trained on patients with real MRI data.
• The Encoder compresses 3D MRI volumes into a compact latent space ($z$) and extracts hierarchical spatial features (skip connections).
• The Decoder is trained to reconstruct the MRI. Once trained, the decoder weights are frozen to serve as a rigorous structural regularizer, ensuring that any generated latent vectors adhere to anatomical constraints.

B. Knowledge Graph (KG) Guided Embedding Propagation

To handle EHR sparsity and heterogeneity across different institutions:

• KG Construction: A Biomedical Knowledge Graph (KG) is built linking patients to medical concepts (symptoms, diagnoses).
• Patient-to-KG Linking: SapBERT generates embeddings for EHR concept descriptions, connecting patient nodes to concept nodes via cosine similarity.
• Graph Attention Network (GAT): A GAT propagates information across the graph.
  • Stage 1 (Supervised Regression): The GAT learns to predict the latent MRI embeddings ($z$) for training patients based on their EHR-derived graph positions.
  • Stage 2 (Decoder-Consistency): The GAT outputs are decoded by the frozen U-Net decoder. A loss function (MSE + SSIM) penalizes reconstruction errors, forcing the GAT to learn embeddings that are compatible with the anatomical manifold.

C. Adapter and Decoder Alignment (Inference)

For EHR-only patients (testing set), the framework synthesizes a latent representation without generating full 3D volumes:

• Whitening-Coloring (WC) Transformation: Aligns the graph-propagated embeddings with the pre-trained MRI latent space distribution to reduce domain shift.
• Adapter Network: A trainable module maps the aligned EHR embeddings into the decoder's latent space.
• Cohort-Aggregated Skip Feature Compensation: Since EHR-only patients lack the high-frequency spatial details (skip connections) required by the 3D decoder, the system retrieves Top-K (specifically $K=1$) clinically analogous patients from the training set. Their skip features are aggregated and used as "structural priors" to guide the decoder.
• Inference: The final step uses the distilled 1D latent vector (processed by a classification head) for AD prediction. The heavy 3D generative branch is discarded during inference, ensuring efficiency.

3. Key Contributions

1. Anatomy-Guided Latent Distillation: The first framework to treat missing-MRI as a latent distillation task using a frozen 3D autoencoder as a structural regularizer, avoiding pixel-level synthesis risks.
2. KG-Driven Cohort Alignment: Utilizes a Biomedical Knowledge Graph and GAT to map heterogeneous EHR schemas into a unified semantic space, effectively transferring structural knowledge across cohorts.
3. Auxiliary Decoding Regularization: Introduces a novel strategy using cohort-retrieved skip features to bridge the gap between macro-prognostic representations (EHR) and micro-texture constraints (3D MRI), acting as a rigorous structural penalty.
4. Clinical Validation: Demonstrates that the framework improves AD classification significantly while maintaining high inference efficiency.

4. Experimental Results

• Dataset: 1,175 patients from the ADNI dataset (70% training, 30% testing). Testing patients had their MRIs masked to simulate EHR-only scenarios.
• Performance:
  • MIRAGE improved the AD classification rate by 13% compared to the strongest unimodal EHR-only baselines (e.g., Logistic Regression, MLP).
  • Balanced Accuracy (BAcc): MIRAGE achieved 0.70 (with MLP classifier), outperforming all baselines (EHR-only MLP: 0.62; EHR→MRI generation: 0.65).
  • Sensitivity: MIRAGE achieved 0.63, a significant improvement over baselines which struggled to detect AD cases (e.g., EHR-only MLP: 0.42).
• Ablation Studies:
  • Removing the KG or GAT caused significant performance drops (BAcc dropped to ~0.54–0.58), proving the necessity of structured semantic mapping.
  • Removing the pre-trained Autoencoder (replacing with MLP) caused performance collapse, confirming the need for a pre-trained structural manifold.
  • Removing the Adapter showed that direct mapping fails, highlighting the need for cross-modal alignment.
• Human Evaluation: Two biomedical experts rated synthetic MRIs. They found synthetic scans to be nearly indistinguishable from real ones in terms of Anatomical Realism and Clinical Usability, with specific AD biomarkers (hippocampal atrophy, ventricular enlargement) correctly preserved.

5. Significance

• Clinical Impact: MIRAGE enables the deployment of high-accuracy multimodal AD diagnostic models in settings where MRI is unavailable, democratizing access to advanced neuroimaging-based diagnostics.
• Methodological Shift: It moves away from the risky paradigm of "generating images from scratch" toward "distilling structural knowledge." By using the 3D decoder only as a training-time regularizer and discarding it at inference, the model achieves high efficiency without sacrificing biological plausibility.
• Robustness: The use of Knowledge Graphs effectively handles the heterogeneity of real-world EHR data, making the solution scalable across different healthcare institutions.

In conclusion, MIRAGE successfully bridges the missing-modality gap by leveraging graph-based knowledge propagation and anatomical constraints to synthesize diagnostic-quality structural representations from tabular clinical data alone.