Introduction
Alzheimer’s Disease (AD) is characterized by significant clinical and biological heterogeneity. Two patients presenting with similar cognitive impairment may follow vastly different trajectories—one progressing rapidly to dementia while the other remains stable for years. Traditional subtyping methods often employ "top-down" approaches, grouping patients based on visible patterns of brain atrophy or protein accumulation. However, these visual clusters do not always align with clinical outcomes.
This research presents a "bottom-up" computational framework that subtypes individuals with Mild Cognitive Impairment (MCI) based on their biological risk of progression. By integrating cerebrospinal fluid (CSF) biomarkers with deep learning analysis of structural MRI, the study identifies distinct risk-based subtypes that can be predicted using non-invasive imaging.
Methodology
The researchers developed a multi-stage pipeline to move from biological risk to structural signatures.
Risk Stratification
The study utilized Amyloid-β (Aβ) levels from CSF to categorize individuals into four risk quartiles: High-risk (H), Intermediate-High (IH), Intermediate-Low (IL), and Low-risk (L). This stratification was performed on a discovery cohort from the Alzheimer’s Disease Neuroimaging Initiative (ADNI, n=544) and validated using the National Alzheimer’s Coordinating Center (NACC, n=508) dataset.
Deep Learning Frameworks
To predict the conversion from MCI to AD, the team evaluated three distinct neural network architectures:
- Multi-layer Perceptron (MLP): Trained on pre-processed, parcellated gray matter volumes.
- Survival Convolutional Neural Network (SCNN): A 3D-CNN that processes volumetric MRI scans directly, utilizing a survival loss function to account for time-to-event data.
- Survival Vision Transformer (SViT): An adaptation of the Transformer architecture for 3D medical imaging.
Explainability and Subtyping
The researchers calculated "centroids" (average patterns) of brain atrophy for each risk group. To understand the models' decision-making, they employed SHAP (SHapley Additive exPlanations) values to identify the specific brain regions most influential in predicting risk for each subtype, referred to as "atrophy-related importance."
Key Results
Progression and Survival
The High-risk (H) subtype demonstrated a significantly greater risk of progressing to AD compared to all other subtypes. In the ADNI cohort, the H subtype showed higher progression rates at both 2- and 4-year marks (p < 0.05). These findings were largely mirrored in the NACC validation cohort, where the H group showed a trend toward lower survival compared to intermediate groups and significantly higher risk compared to the low-risk group.
Model Performance
The MLP model achieved the highest internal performance with a Concordance Index (CI) of 0.869. However, for external validation on the more heterogeneous NACC dataset, the SCNN (enhanced with transfer learning) proved most robust, achieving a CI of 0.681 and a low Brier Score (BS) of 0.095. Interestingly, the SCNN outperformed the SViT, suggesting that the inductive biases of convolutional layers remain highly effective for the spatial hierarchies found in 3D medical imaging at current dataset scales.
Distinct Anatomical Signatures
The study revealed that each risk subtype is associated with a unique anatomical "signature." The average pairwise Jaccard Similarity between the top regions of the subtypes was only 0.25 ± 0.05, indicating minimal overlap. While the High-risk group showed significant atrophy in the medial temporal lobe, other subtypes relied on different regional combinations to drive the model's predictions.
Discussion and Clinical Implications
A critical finding of this work is the stability of these subtypes. Patients identified as "High Risk" at their baseline visit typically maintained that specific atrophy pattern throughout the disease course. Furthermore, neuropathological data (n=31) confirmed that AD-related changes were present across all subtypes in comparable proportions (p > 0.05). This suggests that the subtypes represent different trajectories of the same underlying disease process rather than distinct pathologies.
The ability to map invasive CSF-based risk categories onto non-invasive MRI scans represents a significant step forward for clinical practice. MRI is widely accessible and far less burdensome than the lumbar punctures required for CSF analysis. By identifying a patient's specific risk subtype early, clinicians can better tailor care management and intervention strategies.
Limitations and Future Directions
The drop in model performance during external validation highlights the challenges of dataset heterogeneity; the NACC cohort was generally older and progressed faster than the ADNI group. Future research should focus on integrating longitudinal data—tracking changes across multiple scans—and incorporating multi-omic data (such as genomics and proteomics) to further refine these risk-based models.
Conclusion
This deep learning-driven framework provides an objective, risk-based approach to subtyping cognitive impairment. By shifting the focus from visual patterns to biological risk, the study offers a path toward more precise and individualized management of Alzheimer’s Disease.
Summary updated