MAGIC

MAGIC performs multi-scale semi-supervised clustering of neuroimaging data to stratify patients into anatomically and clinically meaningful brain disease subtypes.

Key Features:

• Multi-Scale Analysis: Derives multi-scale and clinically interpretable feature representations that capture disease heterogeneity across scales rather than relying on fixed-scale atlas-based regions.
• Semi-Supervised Clustering (HYDRA): Implements a semi-supervised clustering approach building on HYDRA to integrate supervised and unsupervised signals for enhanced disease-specific pattern recognition.
• Double-Cyclic Optimization: Employs a double-cyclic optimization procedure to enforce consistency of subtype solutions across different scales.
• Extensive Semi-Simulated Validation: Validates performance using semi-simulated experiments on UK Biobank healthy controls (N=4403) assessing brain atrophy levels, heterogeneity degrees, overlapping subtypes, class imbalance, and sample size.
• Application to Real-World Cohorts: Applies the method to imaging cohorts including ADNI (N=1728) and the PHENOM schizophrenia cohort (N=1166) to investigate neuroanatomical heterogeneity.

Scientific Applications:

• Subtype discovery in neurodegenerative and psychiatric disorders: Identifies anatomically distinct subtypes in Alzheimer's disease and schizophrenia using ADNI and PHENOM imaging data.
• Assessment of clustering reliability: Evaluates conditions under which clustering recovers true disease-related heterogeneity via semi-simulated experiments.
• Interpretation of neuroanatomical variation: Provides multi-scale, clinically interpretable feature representations to aid interpretation of disease-specific neuroanatomical patterns.

Methodology:

Derives multi-scale feature representations; performs semi-supervised clustering based on HYDRA; applies a double-cyclic optimization to ensure inter-scale consistency; conducts semi-simulated experiments on UK Biobank (N=4403) varying brain atrophy, heterogeneity degree, subtype overlap, class imbalance, and sample size; and applies analyses to ADNI (N=1728) and PHENOM (N=1166) imaging datasets.