Clustering and variable selection in the presence of mixed variable types and missing data.

Area of Science:

• Statistics
• Machine Learning
• Biostatistics

Background:

• Clustering patients with autism spectrum disorder (ASD) requires handling complex, mixed-type, correlated, and missing data.
• Existing methods may struggle with the high dimensionality and mixed variable types common in clinical datasets.

Purpose of the Study:

• To develop a model-based clustering approach for patients with potential autism spectrum disorder (ASD).
• To identify influential cognitive and behavioral test scores for accurate patient stratification.
• To enable more efficient and informative diagnostic testing for ASD.

Main Methods:

• Utilized a latent continuous variable approach for discrete variables and a Dirichlet process mixture model for an unknown number of clusters.
• Incorporated variable selection to identify key predictors of cluster membership.
• Applied the method to a dataset of 486 patients with 55 cognitive/behavioral test scores, many with missing or discrete values.

Main Results:

• The proposed method effectively clusters patients and identifies informative variables, outperforming other methods in simulations.
• Analysis of autism spectrum disorder (ASD) data suggested three distinct patient clusters.
• Identified a sparse subset of four key test scores with high posterior probability of informing cluster membership.

Conclusions:

• The developed model-based clustering approach offers an efficient and informative method for patient stratification in complex health science problems.
• This technique can significantly reduce the number of tests required for diagnosis, leading to cost savings and improved patient experience.
• The methodology is broadly applicable to other disciplines facing similar challenges with high-dimensional, mixed-type, and incomplete data.