Neuroimaging Data Informed Mood and Psychosis Diagnosis Using an Ensemble Deep Multimodal Framework.

Area of Science:

• Neuroimaging and computational psychiatry.
• Application of artificial intelligence in mental health research.

Background:

• Current mental illness diagnoses rely heavily on symptoms and self-reports, lacking biological validation.
• This limits neurobiological insights and precise categorization of disorders.
• Previous work integrated structural neuroimaging, but this study advances the approach.

Purpose of the Study:

• To identify brain-based markers for mental illnesses using multimodal neuroimaging (fMRI and structural MRI).
• To enhance diagnostic accuracy by creating more biologically homogeneous patient categories.
• To integrate neuroimaging data with symptom-based diagnoses using advanced computational methods.

Main Methods:

• Utilized multimodal neuroimaging data (fMRI and structural MRI).
• Employed ensemble methods, deep learning (deep convolutional framework), and data fusion techniques.
• Integrated symptom-based categories with biologically derived information to identify distinct patient subgroups.

Main Results:

• Multimodal neuroimaging frameworks outperformed unimodal approaches.
• Ensemble deep learning models showed superior diagnostic classification compared to individual models.
• Identified discrepancies between brain imaging features and symptom-based categories, revealing potential for improved classification and sample heterogeneity mitigation.

Conclusions:

• Combining symptom-based categorization with multimodal neuroimaging and advanced data-driven methods can significantly improve mental illness classification.
• This approach aids in identifying potential biomarkers and biologically homogeneous groups.
• The findings highlight the potential for more precise and biologically informed diagnostic categories in psychiatry.