A multiple hold-out framework for Sparse Partial Least Squares.

Area of Science:

• Neuroscience
• Machine Learning
• Medical Imaging

Background:

• Supervised learning faces challenges in heterogeneous brain disease populations due to unreliable labels.
• Exploratory methods like Sparse Partial Least Squares (SPLS) can uncover brain mechanisms by linking neuroimaging and clinical data.
• Identifying these relationships aids in understanding disease mechanisms, refining assessments, and patient stratification.

Purpose of the Study:

• To introduce a novel SPLS framework for identifying and validating associations between neuroimaging and clinical variables.
• To apply this framework to dementia research, specifically examining grey matter probability maps and Mini-Mental State Examination (MMSE) items.
• To assess the reliability of identified associative effects using data splitting techniques.

Main Methods:

• Developed a novel SPLS framework incorporating variable selection and reliability testing through data splitting.
• Applied the framework to analyze associations between brain grey matter and MMSE performance in a dementia cohort.
• Compared SPLS with non-sparse Partial Least Squares (PLS) and evaluated projection deflation versus classical PLS deflation.

Main Results:

• The SPLS framework identified two statistically significant associative effects.
• These effects linked specific subsets of brain voxels with subsets of MMSE questions/tasks.
• The framework demonstrated the ability to find reliable associations in complex datasets.

Conclusions:

• SPLS demonstrated superior performance compared to PLS, yielding statistically significant effects and higher correlations on hold-out data.
• Projection deflation proved more effective than classical PLS deflation in both SPLS and PLS.
• The proposed SPLS framework offers a robust method for exploring complex relationships in neuroimaging and clinical data, particularly in the context of brain diseases.