Unsupervised learning for medical data: A review of probabilistic factorization methods.

Area of Science:

• Computational Biology
• Data Science
• Medical Informatics

Background:

• High-dimensional data analysis is crucial in fields like genomics, medical imaging, and biobanks.
• Unsupervised learning methods are widely used but often treated in isolation.
• Understanding the underlying principles of these methods is key for effective application.

Purpose of the Study:

• To unify and clarify commonly used unsupervised learning methods for high-dimensional data analysis.
• To highlight the similarities and differences between methods like PCA, K-means, NMF, and LDA.
• To guide applied medical researchers in selecting appropriate methods for their specific health data applications.

Main Methods:

• Review and formulation of four popular unsupervised learning methods: principal component analysis (PCA), K-means clustering, nonnegative matrix factorization (NMF), and latent Dirichlet allocation (LDA).
• Demonstration that these methods can be represented as probabilistic models based on low-rank matrix factorization.
• Discussion of the assumptions, restrictions, inference, and model selection aspects relevant to health data.

Main Results:

• The four reviewed methods share a common foundation in low-rank matrix factorization when viewed as probabilistic models.
• This unified perspective clarifies the distinct assumptions and limitations inherent in each method.
• Provides a framework for applied researchers to better understand and choose between these analytical techniques.

Conclusions:

• A unified probabilistic model framework simplifies the understanding of diverse unsupervised learning techniques.
• Clarifying method-specific assumptions aids in the appropriate selection and application in health data analysis.
• This work facilitates more rigorous and informed use of unsupervised learning in medical research and biobanking.