Supervised, Multivariate, Whole-Brain Reduction Did Not Help to Achieve High Classification Performance in...

Area of Science:

• Neuroimaging
• Psychiatric Disorders
• Machine Learning in Medicine

Background:

• Schizophrenia diagnosis and research benefit from advanced analytical techniques.
• Supervised, multivariate, whole-brain reduction methods are explored for their utility.
• Magnetic resonance imaging (MRI) provides crucial data for brain structure analysis.

Purpose of the Study:

• To evaluate penalized linear discriminant analysis (PLDA) with resampling for schizophrenia diagnostics.
• To identify specific brain regions associated with first-episode schizophrenia.
• To compare PLDA performance against other dimensionality reduction techniques.

Main Methods:

• Utilized penalized linear discriminant analysis (PLDA) with resampling on MRI data.
• Analyzed brain images from 52 first-episode schizophrenia patients and 52 healthy controls.
• Compared PLDA with mass univariate selection (t-test) and principal component analysis (PCA).

Main Results:

• PLDA identified relevant brain areas including the prefrontal cortex, anterior cingulum, insula, thalamus, and hippocampus.
• Classification performance did not significantly improve compared to t-test or PCA.
• No significant influence of imaging features (deformations, gray matter volumes) or classification methods (LDA, linear SVM) was observed.
• Cross-validation settings significantly impacted results, leading to overestimation of classification performance.

Conclusions:

• Rigorous cross-validation across all analysis steps is critical to avoid optimistic bias in neuroimaging classification studies.
• The choice of feature type and classification method had minimal impact on results.
• Careful validation strategies are essential for reliable schizophrenia diagnostic research using neuroimaging.