Evaluating State-Transition Models in Short Resting-State fMRI with Continuous-Dynamics Controls

In short resting-state fMRI, vector autoregressive models outperform hidden Markov models in continuous-dynamics simulations (67/72), while HMMs win unfiltered discrete cases (139/144).

• Why it matters: Apparent discrete state switches in fMRI may reflect smoothly changing activity and analysis choices, so biological validity depends on whether discrete and continuous dynamics can be distinguished.
• What they did: The study compared HMMs and VAR models that used the same prior observations to predict the next time point’s signals from the seven Yeo networks, scoring by log probability density; it also tested temporal filtering effects and analyzed ADNI data from 270 participants.
• The result: Temporal filtering during testing shifted support toward HMMs for continuous cases (from 2/72 with fixed filtering to 68/72), and in ADNI VAR beat HMM in 160/184 datasets with at least 180 time points, implying state-transition interpretations require continuous-model and filtering comparisons.

AI summary of the abstract · may contain errors · not medical advice · read the paper.