Introduction
Alzheimer’s disease (AD) is a progressive neurodegenerative disorder typically monitored through longitudinal clinical visits. These visits generate diverse data types, including structural MRI scans, genetic profiles (Single Nucleotide Polymorphisms, or SNPs), and cognitive test scores. Despite the richness of this multi-modal, time-series data, most existing machine learning approaches for AD biomarker discovery analyze data from only a single time-point. Furthermore, these methods often treat clinical diagnosis (e.g., Healthy vs. AD) and cognitive score prediction as independent tasks, failing to exploit the inherent correlations between a patient's cognitive decline and their clinical status.
Methodology
To address these limitations, this research introduces a unified regularized joint learning model that utilizes tensors to represent multi-dimensional longitudinal data. The input features are organized into a tensor $\mathcal{X}$ (samples $\times$ features $\times$ time), while the targets are represented by an output tensor $\mathcal{Y}$ (scores/diagnoses $\times$ time).
The model's objective function combines multivariate regression for cognitive scores with multi-class Support Vector Machines (SVM) for clinical diagnosis. To ensure robustness and biological relevance, three key regularizations are applied:
- $\ell_{2,1}$-norm: Promotes longitudinal consistency by selecting features that are predictive across all time points.
- Group $\ell_1$-norm: Manages the relative importance of different data modalities (e.g., MRI vs. SNP).
- Trace norm (Nuclear norm): Captures inter-modal relationships, reflecting how different cognitive measures often degrade in tandem.
Given the complexity of the resulting non-convex objective function, the authors developed a custom optimization algorithm based on the Multi-Block Alternating Direction Method of Multipliers (ADMM) to iteratively solve the coupled sub-problems.
Data and Evaluation
The model was validated using data from the Alzheimer’s Disease Neuroimaging Initiative (ADNI-1). The study cohort included 412 subjects with complete data across four time points: Baseline, Month 6, Month 12, and Month 24.
- Input Features: 90 regional gray matter measures extracted from MRI scans via FreeSurfer and SNP genotypes.
- Targets: Rey’s Auditory Verbal Learning Test (RAVLT) scores (regression) and clinical diagnoses (AD, MCI, or Healthy Control) (classification).
Results and Discussion
The joint learning approach significantly outperformed standard baselines in both tasks. In the regression task, the model achieved a Root Mean Squared Error (RMSE) of 12.7 $\pm$ 1.05 for the RAVLT_TOT metric, substantially better than Ridge regression (18.9) or Lasso (19.4). In classification, the model reached a superior overall F1 score of 0.459 $\pm$ 0.0125, with particularly strong performance in identifying AD (F1: 0.496) and Healthy Controls (F1: 0.477).
Beyond predictive accuracy, the model demonstrated high clinical interpretability. The $\ell_{2,1}$-norm successfully identified biomarkers that remained stable over the 24-month period. The model independently highlighted known AD-related regions such as the hippocampus, entorhinal cortex, and amygdala. Furthermore, it correctly identified the APOE-$\epsilon$4 linked SNP (rs429358) as the most significant genetic predictor of the disease.
Conclusion
While the model excels at identifying clear AD and Healthy cases, it faced challenges with the intermediate Mild Cognitive Impairment (MCI) category, where standard Logistic Regression showed a slight advantage. This is likely due to the lack of normalization between the one-vs-all multi-class SVMs within the tensor framework. Future work will focus on refining these classification components and providing a more detailed mathematical breakdown of the optimization process.
Summary updated