Longitudinal clinical score prediction in Alzheimer's disease with soft-split sparse regression based random forest

Area of Science:

• Neuroscience
• Machine Learning
• Biostatistics

Background:

• Alzheimer's disease (AD) is a progressive neurodegenerative disorder impacting global populations.
• Cognitive scores over time are crucial for clinical assessment of AD progression.
• Existing machine learning models for AD prediction face limitations like linearity assumptions and exclusion of subjects with missing data.

Purpose of the Study:

• To develop a nonlinear, supervised sparse regression-based random forest (RF) framework for predicting longitudinal Alzheimer's disease clinical scores.
• To introduce a soft-split technique for probabilistic path assignment in RF, enhancing prediction accuracy.
• To address the challenge of missing data in longitudinal studies by estimating scores rather than excluding subjects.

Main Methods:

• A nonlinear supervised sparse regression-based random forest (RF) framework was developed.
• A novel soft-split technique was proposed for probabilistic sample assignment within the RF.
• Missing cognitive scores were estimated using the proposed RF method before predicting subsequent time points.

Main Results:

• The proposed RF framework demonstrated superior performance compared to traditional RF models.
• The method outperformed other state-of-the-art regression models in predicting AD clinical scores.
• The soft-split technique improved the accuracy of longitudinal score predictions.

Conclusions:

• The developed nonlinear supervised sparse regression-based RF framework offers a robust approach for predicting Alzheimer's disease progression.
• The proposed method effectively handles missing data, a significant improvement over existing techniques.
• This framework has the potential for broader application in predicting clinical scores for various diseases.