Hypothesis Testing in Unsupervised Domain Adaptation with Applications in Alzheimer's Disease

Area of Science:

• Biostatistics
• Machine Learning
• Medical Imaging Analysis

Background:

• Combining clinical and imaging biomarkers from multiple sources is challenging due to data transformations and batch effects.
• Existing methods for comparing datasets with unknown transformations are often complex or computationally expensive.

Purpose of the Study:

• To develop a statistical hypothesis testing framework for comparing probability distributions from different data sources with unknown transformations.
• To address the common impediment of data heterogeneity in large-scale biomedical studies, such as Alzheimer's disease research.

Main Methods:

• Utilized hypothesis testing on transformed measurements, estimating data distortions concurrently with the statistical test.
• Derived a novel algorithm with detailed analysis of its convergence and consistency properties.
• Incorporated lower-bound strategies from continuous optimization.

Main Results:

• Developed a computationally efficient algorithm for testing equality of distributions under unknown transformations.
• Demonstrated the framework's effectiveness on a dataset of individuals at risk for Alzheimer's disease.
• Showcased competitive performance against more expensive and operationally challenging alternative methods.

Conclusions:

• The proposed method provides a feasible and efficient solution for comparing heterogeneous data sources in biomedical research.
• This approach facilitates the integration of multi-site and multi-batch data, enabling larger and more powerful analyses.
• The framework shows promise for advancing research in neurodegenerative diseases by improving data integration techniques.