Learning Multi-Modal Biomarker Representations via Globally Aligned Longitudinal Enrichments

Lyujian Lu, Saad Elbeleidy, Lauren Zoe Baker, Hua WangAAAI 2020 · DOI

Abstract

Predicting the progression of Alzheimer’s Disease (AD) is often complicated by the irregular and multi-modal nature of longitudinal clinical data. We present a novel framework for learning "enriched," fixed-length biomarker representations that align individual patient trajectories with global population trends. By utilizing a soft alignment constraint and structured sparsity, our method integrates heterogeneous data sources—including MRI, cortical thickness, and genetics—while remaining robust to missing time points. Experimental results on the ADNI dataset demonstrate that our enriched representations consistently improve prediction accuracy for various cognitive assessments across multiple regression models.

Introduction

Alzheimer’s Disease (AD) is a progressive neurodegenerative condition characterized by a gradual decline in cognitive function. Tracking this progression requires longitudinal studies, where patients undergo repeated brain imaging and cognitive testing over several years. However, analyzing such data presents three significant challenges:

  1. Missing Data: Irregular follow-up visits and patient attrition lead to "holes" in temporal records.
  2. Multi-Modal Complexity: Integrating diverse data types like Voxel-Based Morphometry (VBM), FreeSurfer cortical measures, and Single-Nucleotide Polymorphisms (SNPs) is non-trivial.
  3. Variable Input Lengths: Standard machine learning models typically require fixed-length inputs, making it difficult to process histories with varying numbers of visits.

To address these issues, we propose a method to learn Enriched Multi-Modal Biomarker Representations. This approach transforms a patient's entire longitudinal history into a single, fixed-length vector, facilitating the use of standard predictive models.

Methodology: Globally Aligned Enrichment

The core of our approach lies in projecting variable-length longitudinal data into a stable feature space. We learn a "global" projection ($W_0$) that captures the baseline structure of the entire population and "local" projections ($W_i$) for each individual patient to capture their specific disease progression.

Soft Alignment and Robustness

To ensure biological plausibility, we implement a soft alignment constraint that keeps local individual trends consistent with the global baseline. We employ a structured sparsity technique ($G_1$-norm) to automatically select the most relevant features across different modalities. Furthermore, to handle the inherent noise in clinical data, we use non-squared $\ell_2$-norm distances, which are more resilient to outliers than traditional squared-error metrics. The resulting non-smooth objective function is solved using a custom iterative algorithm based on the Alternating Direction Method of Multipliers (ADMM).

Experimental Setup

We evaluated our framework using data from the Alzheimer’s Disease Neuroimaging Initiative (ADNI), involving 456 subjects (77 AD, 171 MCI, and 208 Healthy Controls). The modalities included VBM, FreeSurfer, and SNP data. We targeted five primary cognitive assessments, such as the Mini-Mental State Examination (MMSE) and the Alzheimer's Disease Assessment Scale (ADAS).

Key Results and Discussion

The enriched representations were tested against baseline methods using four regression models: Ridge Regression (RR), Lasso, Support Vector Regression (SVR), and Convolutional Neural Networks (CNN).

Performance Gains

In every test case, the enriched representation led to a lower Root Mean Squared Error (RMSE) compared to standard baselines. For instance, MMSE prediction RMSE dropped from 0.4188 to 0.2928 using Ridge Regression. The model also demonstrated superior performance in handling missing data without the need for error-prone imputation.

Comprehensive RMSE results across various cognitive scores and models.

Clinical Validation

Beyond numerical accuracy, the model identified the hippocampus and amygdala as the most significant biomarkers for AD progression. These findings are highly consistent with established medical literature, which identifies these regions as primary sites of early atrophy.

Brain heatmaps identifying key biomarkers like the hippocampus and amygdala.

Conclusion

Our framework for globally aligned longitudinal enrichment provides a robust solution for multi-modal AD progression modeling. By effectively "summarizing" complex patient histories into fixed-length vectors, we enable more accurate and clinically relevant predictions. Future work will explore the integration of non-linear deep manifold learning to further capture the intricate biological relationships underlying neurodegeneration.

Summary updated

BibTeX
@inproceedings{lu2020learning,
  author = {Lyujian Lu and Saad Elbeleidy and Lauren Zoe Baker and Hua Wang},
  title = {{Learning Multi-Modal Biomarker Representations via Globally Aligned Longitudinal Enrichments}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2020},
  doi = {10.1609/aaai.v34i01.5426},
}
Figure