Brain age prediction in schizophrenia: Does the choice of machine learning algorithm matter?

Area of Science:

• Neuroscience
• Radiology
• Machine Learning

Background:

• Brain-predicted age difference (brainPAD) quantifies biological brain age deviation.
• Inter-study variability in brainPAD for schizophrenia is often attributed to sample differences.
• The influence of machine learning algorithms on brainPAD estimation remains under-evaluated.

Purpose of the Study:

• To systematically assess how different machine learning algorithms affect brain-age estimation.
• To evaluate the impact of algorithm choice on brainPAD variability in schizophrenia.

Main Methods:

• Six common regression algorithms were applied to identical brain structural data.
• Data included healthy individuals and schizophrenia patients from multiple cohorts.
• Reproducibility and performance similarity were assessed using correlation and hierarchical clustering.

Main Results:

• Ordinary least squares regression performed poorly compared to penalized algorithms.
• All other evaluated algorithms produced comparable but variable brain-age estimates.
• BrainPAD in schizophrenia varied significantly by algorithm, impacting effect size estimates.

Conclusions:

• Algorithm selection is a critical factor influencing brain-age estimation.
• Variability in brainPAD due to algorithms can confound schizophrenia research.
• Standardizing algorithms is crucial for reproducible brainPAD studies in schizophrenia.