An alternative classification to mixture modeling for longitudinal counts or binary measures.

Area of Science:

• Biostatistics
• Machine Learning in Healthcare
• Clinical Research Methodology

Background:

• Trajectory classification is crucial in clinical research for analyzing patient longitudinal data.
• Traditional k-means clustering is suitable for continuous data but struggles with count and binary data due to mean-dependent variance.
• Existing mixture modeling approaches may prioritize data modeling over classification accuracy.

Purpose of the Study:

• To extend the k-means algorithm for trajectory classification of count and binary data.
• To address the challenge of mean-dependent variance in these data types.
• To introduce a novel distance metric for improved classification accuracy.

Main Methods:

• Developed an extension of the k-means algorithm incorporating a deviance distance metric.
• Justified the deviance distance approach through its analogy with classification likelihood.
• Applied the modified algorithm to two datasets: one with binary data and one with count data.

Main Results:

• Demonstrated significant differences in patient classification when using deviance distance compared to Euclidean distance.
• The proposed method effectively handles the heterogeneity of within-group variability inherent in count and binary data.
• Showcased the practical utility of the extended k-means algorithm in real-world clinical data scenarios.

Conclusions:

• The extended k-means algorithm with deviance distance offers a more appropriate method for trajectory classification of count and binary data.
• This approach enhances the accuracy and reliability of patient subgroup identification in clinical research.
• The findings suggest a valuable alternative to traditional methods when dealing with non-normally distributed longitudinal data.