Learning effective connectivity from fMRI using autoregressive hidden Markov model with missing data.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Biomedical Engineering

Background:

• Functional-integration analysis using fMRI provides insights into neuronal group interactions.
• fMRI signals suffer from low temporal resolution and indirect measurement of neuronal activity, limiting effective connectivity (EC) estimates.

Purpose of the Study:

• To address the limitations of low temporal resolution and indirect measurement in fMRI for more reliable EC estimates.
• To propose and evaluate a novel framework for learning EC using multiple fMRI time series.

Main Methods:

• Utilized an autoregressive hidden Markov model with missing data (AR-HMM-md) within a dynamically multi-linked (DML) framework.
• Modeled unobserved neuronal activity and missing data as variables, estimating their values through joint optimization.
• Employed Bayesian model selection for performance evaluation.

Main Results:

• Demonstrated effectiveness in learning EC through simulated experiments, including studies on sampling and noise effects.
• Achieved over 94% cross-validation accuracy in classifying Attention-Deficit/Hyperactivity Disorder subjects versus controls using distinctive network EC.
• Showed that the proposed DML-AR-HMM-md yields EC estimates more consistent with literature than traditional Granger causality.

Conclusions:

• The proposed DML-AR-HMM-md framework offers more reliable EC estimates compared to existing latent models.
• Overcame the disadvantage of low temporal resolution inherent in fMRI data.
• Significantly improved cross-validation accuracy by incorporating missing data variables and an autoregressive process.