Neuropsychiatric Disorder Subtyping Via Clustered Deep Learning Classifier Explanations

Area of Science:

• Neuroscience
• Psychiatry
• Artificial Intelligence

Background:

• Neuropsychiatric disorders require better diagnostic and treatment strategies.
• Dynamic functional network connectivity (dFNC) from resting-state fMRI is key to studying brain network interactions over time.
• Current research often uses deep learning or clustering but rarely combines both.

Purpose of the Study:

• To develop a novel approach for subtyping individuals with neuropsychiatric disorders, specifically schizophrenia (SZ).
• To leverage explainable deep learning and cross-validation for robust subtype identification.
• To investigate the relationship between identified subtypes, brain network interactions, and symptom severity.

Main Methods:

• Trained an explainable deep learning classifier on dFNC data to distinguish individuals with SZ from controls.
• Achieved a 79% test accuracy in differentiating SZ from controls.
• Employed cross-validation to identify 5 distinct SZ subtypes based on network connectivity patterns and symptom severity.

Main Results:

• Identified 5 schizophrenia subtypes, each with unique dFNC differences compared to controls.
• Observed distinct alterations in network interactions, including visual, subcortical, sensorimotor, auditory, and cerebellar networks.
• Found statistically significant differences in negative symptom scores among the identified subtypes.

Conclusions:

• The novel subtyping approach offers a new way to understand and characterize schizophrenia.
• These subtypes, defined by distinct brain connectivity patterns, correlate with varying symptom severity.
• This work has the potential to advance personalized medicine for schizophrenia and other neuropsychiatric disorders.