Excitatory feedback, phase bistability, and the closed-loop MEA

What the work claims

Machado, Silva, Brito, Pena, and Matias model two reciprocally connected cortical-like populations and ask whether anticipated synchronization (AS) and phase bistability survive once the sender population receives excitatory feedback from the receiver.¹ Their claim is that both phenomena are robust to bidirectionality, and that the feedback conductance acts as a control knob. Depending on the local inhibitory strength in the receiver population, increasing feedback can route the system from AS to delayed synchronization (DS) either continuously through zero-lag (ZL) synchronization or discontinuously through a bistable regime in which AS and DS coexist.

The authors also report that external noise in the receiver population biases the phase relation. Lower noise makes the receiver more likely to lead; higher noise makes it more likely to follow. This is a computational paper: every result comes from numerical integration of Izhikevich neuron populations, not from a biological recording.

How it works

Each population contains 500 Izhikevich neurons, 80% excitatory and 20% inhibitory, with parameters drawn from distributions that produce regular-spiking, intrinsically bursting, chattering, fast-spiking, and low-threshold-spiking behavior.¹ Every neuron receives 50 recurrent synapses from its own population and 20 excitatory synapses from the other population. The forward coupling from population A to B has conductance g_fw = 0.5 nS; the feedback coupling from B to A, g_fb, is varied.

The neurons are driven by independent Poisson excitatory synapses at 2400 Hz, modeled as a conductance g_P = 0.5 nS. AMPA and GABA_A synaptic decays are fixed at 5.26 ms and 5.6 ms, respectively. The authors average the membrane potential across all neurons in a population and treat that mean as a proxy for the local field potential, then extract peak times cycle by cycle.

The phase relation is quantified by the per-cycle time delay τ_i = t_i^B - t_i^A. A negative mean τ means B peaks before A (anticipated synchronization), a positive mean means B peaks after A (delayed synchronization), and values near zero mean zero-lag synchronization. The authors classify regimes using thresholds of ±0.5 ms.

When the receiver population has strong local inhibition (g_I^B = 2.0 nS), raising g_fb drives a smooth AS-to-DS transition through ZL. When local inhibition is weaker (g_I^B = 0.6 nS), the same sweep produces a bimodal delay distribution: the system jumps between AS and DS. For very weak inhibition (g_I^B = 0.1 nS), bistability is lost once g_fb reaches about 10% of g_fw. Lowering the Poisson noise in B extends the AS regime; raising it favors DS.

Where a skeptic should push

The single most load-bearing assumption is that the mean membrane potential of a population is a faithful proxy for the local field potential and, by extension, for the phase signals a cortical or organoid MEA would actually record. Real extracellular fields are shaped by dendritic orientation, volume conduction, electrode impedance, and the spatial extent of the population; they are not simply population averages of somatic voltage. A skeptic should ask whether the reported phase bistability would survive a more realistic forward model of the field.

The model is also deliberately minimal. It lacks spike-frequency adaptation, dendritic nonlinearities, multiple synaptic timescales, and short-term plasticity. The ±0.5 ms classification thresholds are pragmatic but arbitrary; the qualitative result does not depend on the exact numbers, but the boundaries of each regime do. The authors note these limitations and frame the work as a proof of mechanism rather than a direct biological measurement.

Finally, the link to experimental observations of negative phase lags in EEG and MEG is suggestive, not causal. The paper shows that bidirectional coupling can produce such phase relations; it does not demonstrate that it does in vivo.

Why closed-loop MEA controllers must treat phase as a state variable

Most closed-loop MEA systems today trigger stimulation on threshold crossings, burst rates, or band-limited power.¹ This paper says that for two coupled populations, the same coupling strength can support two stable phase relations. A stimulus that only checks whether neurons are firing enough may push the network from the AS attractor to the DS attractor without warning, or leave it stranded in a bistable mixture where some cycles lead and others lag.

The non-obvious implication is that phase lag is not a readout; it is a state variable. A closed-loop controller needs a phase-locked-loop-like front end: track the cycle-by-cycle phase of the local field, measure the delay to the partner population or to the stimulus artifact, and gate stimulation within a narrow window. Simply knowing that a burst occurred is insufficient; the controller must know when in the cycle it occurred.

The opportunity is large. If the mechanism holds in organoids or slices, a phase-aware MEA could entrain a culture to a desired rhythm, suppress pathological oscillations by exploiting AS, or synchronize multiple organoids by timing stimuli to their relative phase. The acquisition chain would shift from spike sorting and rate estimation to real-time phase estimation at sub-millisecond resolution.

The threat is equally real. Bidirectional MEA-tissue coupling is exactly the feedback loop studied here: the MEA reads activity, computes a response, injects current, and the tissue changes. If the controller latency is more than a millisecond or two, the phase window closes. Worse, a stimulus artifact can be mistaken for the target phase, creating positive feedback that drives the system into an unintended attractor. The paper warns that history matters: the same waveform can produce opposite phase relations depending on initial conditions and noise.

There is also a measurement artifact angle. The authors use a population mean as an LFP proxy. A real MEA records spatially weighted mixtures of many sources. If two electrodes see different phase mixtures, one electrode may classify the system as AS while another classifies it as DS. Phase bistability in the tissue could be mistaken for electrode crosstalk or sorting error unless the acquisition system samples densely enough to resolve the spatial phase profile.

The bottom line

The paper establishes, in a simplified but carefully specified model, that excitatory feedback between two neuronal populations can preserve anticipated synchronization and create phase bistability. It is a computational mechanism, not an empirical measurement of cortex or organoid tissue. What would confirm it is a paired recording or optogenetic feedback experiment in which the same inter-population coupling strength is shown to switch the system between AS and DS depending on initial conditions. What would break it is evidence that real tissue never operates in the parameter regime where bistability occurs, or that volume conduction smooths away the phase signal faster than the controller can track it.

For MEA hardware, the takeaway is practical: closed-loop systems that only regulate firing rate are flying blind with respect to phase. The acquisition chain needs to deliver cycle-resolved phase, and the controller needs to treat that phase as a state with memory, not as a real-time feature.