A Bayesian transfer learning model for mixed-effects longitudinal data analysis

Area of Science:

• Statistics
• Biostatistics
• Machine Learning

Background:

• Longitudinal data analysis with high-dimensional measurements and small sample sizes presents challenges like variable estimates and low power.
• Borrowing information from similar datasets can improve inference but risks negative transfer and bias.

Purpose of the Study:

• To develop a novel Bayesian transfer learning model for longitudinal data (BTLL).
• To enhance estimation accuracy and enable data-adaptive information borrowing.
• To minimize negative transfer and reduce bias in target population inference.

Main Methods:

• Developed a Bayesian transfer learning model for longitudinal data (BTLL).
• Utilized mixture models for discrepancies between target and source study outcome model parameters.
• Applied BTLL to predict dementia risk using mild cognitive impairment datasets and non-invasive biomarkers.

Main Results:

• BTLL substantially improves the precision of parameter estimates in the target study compared to alternatives.
• BTLL reduces estimation bias in heterogeneous settings where target and source datasets deviate.
• The BTLL prediction model achieved the lowest prediction error for identifying individuals at high risk for dementia.

Conclusions:

• BTLL offers a robust approach for analyzing longitudinal data with limited sample sizes.
• The model effectively borrows information across studies while mitigating negative transfer.
• BTLL demonstrates significant potential for early dementia risk prediction using non-invasive biomarkers.