Kernels for longitudinal data with variable sequence length and sampling intervals.

Area of Science:

• Gerontology
• Machine Learning
• Biostatistics

Background:

• Longitudinal data analysis is crucial for understanding cognitive decline.
• Existing methods may not fully capture complex temporal patterns in elderly health data.
• Early detection of cognitive decline is vital for timely intervention and care.

Purpose of the Study:

• To develop and evaluate novel kernel methods for classifying longitudinal data.
• To apply these methods for the detection of cognitive decline in elderly individuals.
• To compare the performance of different kernel approaches and data types.

Main Methods:

• Development of mixed-effects models as hierarchical empirical Bayes generative models for time series.
• Creation of novel Fisher kernels based on mixture of mixed-effects models.
• Application of support vector machine classifiers with developed kernels.
• Exploration of nonparametric kernels based on reproducing kernel Hilbert space.
• Utilizing longitudinal clinical data from motor and neuropsychological tests.

Main Results:

• Mixed-effects models demonstrated utility in likelihood ratio classifiers, outperforming standard regression models.
• Hierarchical generative models effectively handled variations in sequence length and sampling intervals.
• Classifiers based on neuropsychological tests yielded superior performance compared to those using motor behavior data.
• Discriminant classifiers outperformed likelihood ratio classifiers for motor behavior tests.

Conclusions:

• Novel kernel methods, particularly Fisher kernels derived from mixed-effects models, show promise for detecting cognitive decline.
• Neuropsychological test data appears more informative than motor behavior data for this classification task.
• The developed methods offer a robust approach to analyzing complex longitudinal health data for early disease detection.