Excitation-inhibition imbalance in depression and the limits of observability: cohort-level inference from resting-state EEG with an amortised three-region Wilson-Cowan model

Depression is widely thought to involve an imbalance between neuronal excitation and inhibition, but the idea has rarely been tested quantitatively at the level of macroscopic brain networks, because dynamical models are hard to fit to EEG spectra and fitting them person by person is prohibitively slow. We present a complete, reproducible pipeline that makes such a test possible, and we report where it succeeds and where it does not. Using eyes-closed resting-state EEG from 117 participants (22 with lifetime major depressive disorder, 95 without), we fitted a three-region excitatory-inhibitory network model with an amortised Bayesian method, in which a single shared library of simulations serves every participant and reduces inference to milliseconds per person. Depressed participants showed a consistent shift towards excitation in three core networks, together with a stronger inhibitory loop gain and a lower operating-point drive. The model's default-mode resonance frequency rose with symptom severity, whereas the conventional peak-frequency estimate computed from the same recordings did not. We then asked what the model actually buys. For distinguishing groups it did not outperform simple spectral measures, and its per-person low-frequency predictions were poor, so we restrict our claims accordingly. A real-time state observer detected sudden shifts in the network's operating point within milliseconds, yet could not reconstruct the moment-to-moment state, a limit set by the model's intrinsic randomness rather than by our algorithm. Finally, a dose-matched control showed that closed-loop stimulation is no more accurate than a fixed stimulus at the same dose, so we report the real-time feasibility of the loop without claiming an accuracy advantage. The contribution of the paper is therefore a reusable, calibrated inference pipeline together with an explicit map of its limits, including negative findings that we deliberately kept as results.