Abstract
Complex diseases like Parkinson’s Disease (PD) are driven by intricate interactions across biological scales, from epigenetic markers to structural brain changes. Traditional deep learning approaches often fail to capture these cross-modal dependencies effectively or provide biological interpretability. This paper introduces a Multi-Head Co-Attention (MHCA) architecture that learns a joint similarity distribution between 3D SPECT brain scans and DNA methylation data. Our model achieves state-of-the-art accuracy (98.35%) with 53% fewer parameters than standard late-fusion models. Furthermore, the co-attention mechanism enables the discovery of novel biological links, specifically highlighting the interferon-gamma pathway's role in PD via DNA methylation.
Introduction
Diagnosing and understanding complex pathologies like Parkinson’s Disease (PD) requires synthesizing data from diverse sources, such as 3D medical imaging and high-dimensional genomic sequences. Standard "late fusion" approaches process these modalities independently and concatenate features at the final layer. This often results in inefficient learning and a "black box" model that lacks clinical interpretability. To address these challenges, we propose a co-attentive framework that actively weights features from different modalities based on their mutual context.
Methods: Multi-Head Co-Attention (MHCA)
The core of our approach is the Multi-Head Co-Attention (MHCA) mechanism, inspired by Transformer architectures in natural language processing. Unlike self-attention, co-attention computes a similarity matrix between two distinct data streams: 3D spatial features from SPECT images and 1D epigenetic features.
The process involves several key stages:
- Feature Extraction: A 3D CNN encodes SPECT images to extract spatial features. Simultaneously, epigenetic data (DNA methylation) undergoes targeted feature selection via XGBoost and Recursive Feature Elimination to align its dimensionality with the image latent space.
- Bidirectional Weighting: The model calculates two sets of attention coefficients ($\alpha_q$ and $\alpha_v$), determining how much each modality should inform the other. This bidirectional approach reflects the interdependent nature of medical data.
- Integration: Weighted features are combined using residual connections and layer normalization before passing to a final linear layer for prediction.
Data and Experimental Setup
We utilized data from the Parkinson’s Progression Markers Initiative (PPMI), including 121 healthy controls and 287 PD patients.
- SPECT Images: 3D scans modeling dopamine-transporter functionality.
- DNA Methylation (DNA-m): Epigenetic markers filtered from ~765,000 features down to 441 predictive markers. This upfront selection ensures that attention weights remain directly mappable to specific biological markers.
Results and Discussion
The MHCA model significantly outperforms existing benchmarks. It achieved an accuracy of 98.35% and an AUROC of 0.976, surpassing the previous state-of-the-art 3D-CNN (96.00%) and a baseline concatenation model (97.12%). Notably, the model is highly efficient, utilizing only ~2 million parameters—a 53% reduction compared to the concatenation baseline.
Interpretability and Biological Discovery
A key advantage of MHCA is its inherent interpretability. By visualizing the attention masks, we confirmed the model focuses on the putamen, the region most associated with dopaminergic cell loss in PD.
Beyond imaging, the model identified DNA methylation markers linked to the interferon-gamma (IFN-$\gamma$) pathway. While previously linked to PD through gene expression, our findings suggest a novel epigenetic regulatory mechanism for this pathway in PD progression, potentially influenced by environmental factors.
Conclusion
The MHCA framework demonstrates that co-attention is a powerful tool for multimodal medical data integration. By moving beyond simple concatenation, we achieve higher accuracy, greater parameter efficiency, and direct biological interpretability. Future work will focus on scaling the attention mechanism to handle higher-dimensional data natively and exploring how perturbations in one modality affect the attention weights of another.
Summary updated