Computational neuroimaging strategies for single patient predictions

Area of Science:

• Neuroscience
• Computational Psychiatry
• Medical Imaging

Background:

• Machine learning (ML) is widely used in neuroimaging for single-subject predictions.
• ML establishes links between observed data features and clinical variables.
• An alternative approach focuses on inferring underlying mechanisms.

Purpose of the Study:

• Discuss the rationale for computational models in neuroimaging-based single-subject inference.
• Highlight the potential for characterizing individual disease mechanisms.
• Connect these characterizations to clinical predictions.

Main Methods:

• Explores computational modeling for inferring (patho)physiological and cognitive mechanisms.
• Reviews Bayesian model selection and generative embedding for linking models to predictions.
• Examines how these methods handle heterogeneity in disorders.

Main Results:

• Computational models offer a mechanistic alternative to ML in neuroimaging.
• Methods discussed can characterize individual disease mechanisms.
• These approaches link mechanistic insights to clinical predictions.

Conclusions:

• Computational models provide a mechanistic framework for neuroimaging-based inference.
• This approach aids in understanding individual disease heterogeneity.
• It bridges mechanistic understanding with statistical ML perspectives.