Challenges and Opportunities with Causal Discovery Algorithms: Application to Alzheimer's Pathophysiology

Area of Science:

• Computational Biology
• Neuroscience
• Data Science

Background:

• Traditional association-based methods struggle to identify causal relationships.
• Causal Structure Discovery (CSD) offers advanced computational approaches.
• Alzheimer's disease (AD) provides a well-defined causal graph for evaluating CSD methods.

Purpose of the Study:

• To assess if CSD methods can uncover known causal links from observational clinical data.
• To provide guidance for accurate causal relationship discovery.
• To evaluate CSD performance using Alzheimer's disease as a model.

Main Methods:

• Evaluated Fast Causal Inference (FCI) and Fast Greedy Equivalence Search (FGES) algorithms.
• Utilized data from the Alzheimer's Disease Neuroimaging Initiative (ADNI).
• Compared CSD results against a literature-derived 'gold standard' causal graph under varying background knowledge scenarios.

Main Results:

• Dedicated CSD methods successfully discovered causal graphs closely matching the gold standard.
• The performance of CSD methods improved with increased background knowledge.
• Structural equation models, not designed for CSD, served as a control.

Conclusions:

• CSD methods are effective for identifying causal relationships in complex diseases like Alzheimer's.
• Optimal results are achieved when CSD algorithms are applied to longitudinal data with substantial prior knowledge.
• This study offers guidance for leveraging CSD in clinical data analysis.