Deep Feature Selection and Causal Analysis of Alzheimer’s Disease

Yuanyuan Liu, Zhouxuan Li, Qiyang Ge, Nan Lin, Momiao XiongFrontiers in Neuroscience 2019 · DOI

TL;DR

This paper tackles the "black box" problem of deep learning in Alzheimer's Disease (AD) diagnosis by developing a novel framework that combines deep learning with feature selection and causal inference. This approach not only improves prediction accuracy but also identifies key brain regions and potential causal pathways from genetics to AD, offering a more interpretable understanding of the disease.

What Problem Does This Paper Address?

Alzheimer's Disease (AD) is a devastating neurodegenerative disease affecting millions worldwide, with diagnosis and prediction being challenging, especially in early stages. While deep learning, particularly deep convolutional neural networks (DCNNs), has shown promise in analyzing brain imaging data for AD diagnosis, these models are often opaque ("black boxes"). This lack of interpretability hinders our ability to understand the underlying mechanisms of AD and identify specific brain regions or genetic factors contributing to its development. The paper aims to bridge this gap by creating a more transparent and insightful AI platform for AD analysis.

Methods

The researchers propose a novel framework integrating several key components:

  • Deep Learning for Classification: They employ a 3D Convolutional Neural Network (CNN), specifically a Visual Geometry Group (VGG) architecture, to classify AD status from Diffusion Tensor Imaging (DTI) data. This involves processing 3D brain images to capture complex structural information.
  • Feature Selection and Visualization: To understand which parts of the DTI images are most important for prediction, they use a "prediction difference analysis." This technique quantifies the impact of removing information from specific image regions on the classification accuracy, effectively highlighting critical brain areas.
  • Causal Inference: To move beyond mere association and identify causal relationships, they utilize Conditional Generative Adversarial Networks (CGANs) and classifier two-sample tests. This allows them to investigate whether specific brain regions or genetic variants cause AD, rather than just being correlated with it. This is a significant step towards understanding disease mechanisms.
  • Genetic-Imaging Data Analysis: The framework integrates genetic data with brain imaging to explore causal pathways from genetic variants to AD, mediated by brain structures.

Datasets

The study utilizes the Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset. This dataset contains DTI images from 151 subjects (51 AD patients and 100 non-AD controls) collected at four time points: baseline, 6 months, 12 months, and 24 months. Genetic data was also processed for these individuals.

Key Insights and Ideas

  • Beyond Correlation to Causation: The most significant insight is the move from simply identifying associations between brain regions and AD to establishing potential causal links. This is crucial for understanding disease progression and developing targeted interventions.
  • Interpretable Deep Learning: By integrating feature selection and visualization techniques, the "black box" of deep learning is opened up, revealing which brain regions are most influential in predicting AD.
  • Temporal Progression of AD: The study highlights how the number and significance of brain regions causally linked to AD increase over time, reflecting the progressive nature of the disease.
  • Key Brain Regions Identified: The temporal lobes (including the hippocampus) and the ventricular system are consistently identified as important regions for AD prediction and causation.
  • Genetic-AD Pathways: The framework allows for the exploration of how genetic variations might influence brain structures, which in turn contribute to AD.

Key Results

  • Prediction Accuracy: The VGG CNN model achieved promising prediction accuracies for AD, with AUC values around 0.85 at baseline and decreasing slightly over time, indicating its effectiveness in classifying AD status.
  • Important Brain Regions: Feature selection identified key brain regions such as the temporal lobe (left, medial, and right), ventricles, enlarged ventricles, occipital lobe, and prefrontal area as contributing significantly to AD prediction.
  • Causal Relationships: Through CGAN analysis, the study found significant causal relationships between specific brain regions and AD. Notably, the ventricles and enlarged ventricle, as well as the left and right temporal lobes, showed strong causal links across all time points.
  • Genetic-Causal Links: The study identified potential causal relationships between specific genes (e.g., CD33, APP) and brain regions implicated in AD, suggesting pathways through which genetic factors might influence the disease.

Limitations and Future Work

  • Sample Size: The authors acknowledge that the sample size (51 AD patients) is relatively small and imbalanced, which may limit the sensitivity of the prediction models and the generalizability of the findings. They emphasize the need for larger datasets and further analysis.
  • Preliminary Results: The study's results are described as preliminary, suggesting that more extensive validation and replication are necessary.
  • Future Directions: The paper calls for increased sample sizes and the analysis of additional datasets to validate and improve the robustness of the proposed framework. The authors also aim to develop more robust deep learning platforms that can integrate multiple imaging modalities and uncover causal pathways more effectively.

Summary updated

Abstract

Deep convolutional neural networks (DCNNs) have achieved great success for image classification in medical research. Deep learning with brain imaging is the imaging method of choice for the diagnosis and prediction of Alzheimer's disease (AD). However, it is also well known that DCNNs are "black boxes" owing to their low interpretability to humans. The lack of transparency of deep learning compromises its application to the prediction and mechanism investigation in AD. To overcome this limitation, we develop a novel general framework that integrates deep leaning, feature selection, causal inference, and genetic-imaging data analysis for predicting and understanding AD. The proposed algorithm not only improves the prediction accuracy but also identifies the brain regions underlying the development of AD and causal paths from genetic variants to AD via image mediation. The proposed algorithm is applied to the Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset with diffusion tensor imaging (DTI) in 151 subjects (51 AD and 100 non-AD) who were measured at four time points of baseline, 6 months, 12 months, and 24 months. The algorithm identified brain regions underlying AD consisting of the temporal lobes (including the hippocampus) and the ventricular system.
BibTeX
@inproceedings{liu2019deep,
  author = {Yuanyuan Liu and Zhouxuan Li and Qiyang Ge and Nan Lin and Momiao Xiong},
  title = {{Deep Feature Selection and Causal Analysis of Alzheimer’s Disease}},
  booktitle = {Frontiers in Neuroscience},
  year = {2019},
  doi = {10.3389/fnins.2019.01198},
}
Figure