Research analysis · Training methods

Per-neuron calcium dynamics that evolution discards

A single-author study evolves spiking seizure detectors whose neurons each carry their own activity-dependent calcium current. The per-neuron version wins on held-out patients, the shared versions do nothing, and an ablation then reveals the uncomfortable part: the finished networks score identically with the calcium removed, so the gain came from enriching the evolutionary search, not from the dynamics themselves.

Source: Evolved Spiking Neural Networks for EEG Seizure Detection: Calcium Dynamics and Delay Mechanics, arXiv:2611.03841 (q-bio.NC), 2 November 2026. Primary source. Read: the full arXiv HTML, including the neuron models, the evolution setup, all results tables, supplementary tables S1 to S3, and the ablation analysis.

What the work claims

Tuba Aksoy asks a precise question about neuroevolved spiking networks: does giving integrate-and-fire neurons a slow, activity-dependent calcium current, of the kind used in biophysical models of intrinsic excitability, improve networks that the EONS framework (Evolutionary Optimization of Neuromorphic Systems) designs for EEG seizure detection?4 The answer is yes, but only under one specific design choice. When the calcium current's two parameters are evolved separately for every hidden neuron (Ca-IFneu), record-level MCC on 10 CHB-MIT patients held out from training rose from 0.592 to 0.726 over 20 independent runs (paired Wilcoxon p = 0.008, Bonferroni-adjusted p = 0.025), and the false-alarm rate fell from 0.17 to 0.10. When the same parameters were fixed by hand or evolved once per network, nothing significant happened (p = 0.952 and p = 0.430 respectively).1

This is a primary results paper, in silico, from a single author, with no external funding declared. Its statistical protocol is better than the field's norm: all test patients are disjoint from training, validation, and ensemble selection; every comparison runs 20 seeded replicates paired by run; p-values are exact two-sided Wilcoxon signed-rank tests with Bonferroni adjustment within each family of arms. The work's real payload, however, is not the headline gain. It is the ablation.

How it works

The calcium model is borrowed from earlier biophysical work on active intrinsic conductances in recurrent networks.2 Each hidden neuron carries a calcium variable that decays with a time constant of 4 ticks (100 ms at 25 ms per tick) and increments by a per-neuron amount rho whenever the neuron spikes. A Hill function with exponent m = 2 and half-activation at calcium level 2 converts the calcium concentration into a conductance, and that conductance drives the membrane potential toward a reversal level of 0.8, mimicking the calcium-activated nonspecific cation current that makes biological neurons adapt after firing. The three variants differ only in where the parameters live: fixed values for all neurons (gmax 0.05, rho 0.09), one evolved pair per network, or one evolved pair per neuron, with gmax allowed in [0, 0.07] and rho in [0, 0.20].1

The rest of the pipeline deliberately follows a published edge-classification pipeline for evolved spiking networks.3 Scalp EEG from CHB-MIT (256 Hz, 18 bipolar channels) is normalized per patient by the 5th and 95th percentiles and encoded into 0.5-second windows of 20 ticks, two spike bins per channel, giving 36 input neurons. Networks start as 30 hidden neurons with 30 synapses; a population of 100 is evolved for 30 generations with fitness equal to the window-level Matthews correlation on 10 training patients; the top 10 percent by detection-threshold-tuned MCC are kept, and the best single network, pair, or trio on the training patients is scored once on the 10 held-out test patients, whose 34 records include 14 with seizures. Tested ensembles are tiny, about 24 to 25 synapses and 34 hidden neurons.1

Two control families bracket the calcium result. Evolved synaptic delays of 1 to 3 ticks (25 to 75 ms) per synapse gave a similar gain (MCC 0.703, +0.111 over baseline, adjusted p = 0.024), while longer delays, fixed delays, and per-neuron axonal delays gave none. And longer calcium time constants eroded the effect: 16 ticks (400 ms) survived only before Bonferroni correction (adjusted p = 0.15), and 32, 64, and 100 ticks were flat.1

Then the twist. Re-scoring the finished Ca-IFneu networks with the calcium current removed, or with the per-neuron calcium parameters shuffled among neurons, left MCC identical in all 20 runs. Removing the evolved synaptic delays from the delay-arm networks changed MCC by less than 0.01. Both kinds of modified network still beat the plain integrate-and-fire baseline. The author's interpretation, offered explicitly as one possible explanation, is that the extra per-neuron parameters add heritable variation that helps evolution reach better topologies, weights, and thresholds, even though the solutions they help find do not use the parameters afterwards.1

Where a skeptic should push

The load-bearing assumption is that the ablation diagnosis is right. Correlation between richer genomes and better search is not proof of causation: an alternative reading is that per-neuron parameters simply lengthen the genome and slow effective convergence in a way that happens to regularize this particular pipeline, or that any per-neuron degree of freedom would have done the same. The paper does not run the control that would discriminate these stories, for example evolving a per-neuron inert parameter with no dynamical effect. The honest statement is narrower than the title: something about per-neuron parameterization helps this search.

Second, the clinical picture is weaker than the CHB-MIT headline suggests. On Siena, with identical settings and no cohort-specific tuning, the gain shrank to +0.059 MCC (p = 0.052, not significant after correction) in one split and +0.096 (adjusted p = 0.021) in the other, and the false-alarm rates on Siena were 0.74 and 0.65 in the two splits, which no clinician would deploy. Direct transfer of CHB-MIT-trained ensembles to all 14 Siena patients gave a nonsignificant +0.053. The author is upfront: software only, modest cohorts, one training-set choice per split, record-level metrics rather than detection latency or false alarms per hour, and no neuromorphic hardware run yet. The work motivates larger evaluation; it does not establish clinical utility.1

Third, the scale. Networks of about three dozen neurons do not scare anyone as models of cortex, and seizure detection on 34 records is a niche where a two-layer heuristic can be competitive. The finding that matters is about the search dynamics of evolution, and that is where it should be stress-tested, on larger networks and tasks where evolved solutions cannot be trivially enumerated.

Cellular heterogeneity as a training lever

For organoid intelligence, the non-obvious implication cuts against the field's favorite assumption. The tempting reading of this paper for wetware is that because biological neurons carry real calcium-activated adaptation currents, tissue should compute better with them. The data argue something subtler: the measurable benefit of per-neuron adaptation was in shaping what training could find, not in what the trained system did at inference. Living neural tissue is saturated with exactly this kind of per-cell heterogeneity, intrinsic excitability, adaptation, ion-channel complements that differ cell to cell. This paper suggests the value of that heterogeneity may be mostly upstream: it enlarges and smooths the space of networks that a training process can reach. If so, an organoid computer's biological richness is not its runtime feature; it is its search affordance, and the engineering question becomes how to exploit heterogeneous dynamics during training rather than how to preserve them at readout.1

The practical transfer is the training loop itself. EONS evolves topology, weights, thresholds, and delays with no gradients at all, on networks small enough to map onto constrained hardware. That is precisely the shape of any trainer for organoid tissue, where backpropagation is biologically implausible and closed-loop stimulation experiments are the only training signal. The calcium result then becomes a design rule for closed-loop electrophysiology: when searching stimulation policies or readout decoders for a culture, give the search per-electrode, per-neuron tunable dynamics rather than one shared configuration. The paper's own speculation, that evolution could tune such parameters per network and possibly per patient, translates directly into per-culture and per-plate personalization, which is how organoid systems will have to work anyway, since no two cultures are alike.

The threats are calibration and cost. On calibration: if the lesson is "extra searchable parameters help the search," the field could waste years interpreting biological mechanisms as computations when they are scaffolding, and the failure mode is invisible until someone runs the ablation. On cost: the paper spent 100 networks times 30 generations, with fitness evaluated on windowed EEG, trivial in software. For a living substrate, one fitness evaluation is a day of cell culture and an hour of electrophysiology, so brute-force neuroevolution is a non-starter at tissue scale; the search budget has to be moved onto surrogates and models of the tissue, which reintroduces every sim-to-real gap this paper avoided by staying in simulation. There is also a quieter dual-use note: the same evolutionary machinery that tunes a seizure detector could be pointed at optimizing evoked responses in ways that harm the culture, and a search that treats tissue as a black-box fitness oracle will happily find stimulation regimes that injure it if injuries correlate with the metric.

The bottom line

Established: on 10 CHB-MIT patients never seen in training, per-neuron evolved calcium parameters raised record-level MCC from 0.592 to 0.726 (adjusted p = 0.025) and cut the false-alarm rate from 0.17 to 0.10, while shared or fixed calcium parameters did nothing, and the finished networks retained their advantage with the calcium removed or shuffled. Established with weaker confidence: the same effect on Siena, significant in only one of two splits with very high false-alarm rates, and no cohort transfer. Asserted but unverified: that the benefit operates through the evolutionary search, that it generalizes to larger networks and harder tasks, and that it survives contact with neuromorphic hardware. What would confirm the search explanation: a control with per-neuron inert parameters, and replication at larger scale. What would break it: evidence that the gain tracks genome length or mutation rate rather than the dynamical meaning of the parameters. For organoid intelligence the durable lesson is a question worth funding: in living tissue, is biological heterogeneity the computation, or the thing that makes the computation trainable? This paper is the cleanest small-scale evidence yet that the second answer deserves equal billing.

Frequently asked questions

What did the calcium current actually do in these neurons?

Each neuron kept a calcium variable that decayed over about 100 ms and jumped by a per-neuron amount whenever the neuron spiked. A Hill function turned that calcium level into a conductance that pushed the membrane toward a fixed reversal value, mimicking the calcium-activated cation current behind spike-frequency adaptation in real neurons.

Why did per-neuron parameters work when per-network ones did not?

That is the paper's central puzzle. Evolving one parameter pair per neuron gave a significant gain; evolving one pair shared across the network did not. The ablation, where removing or shuffling the calcium after training changed nothing, points to the parameters helping the evolutionary search find better networks rather than the finished networks depending on calcium dynamics.

How were patients split between training and testing?

On CHB-MIT, 10 pediatric patients were used for training and 10 entirely separate patients for testing, with 34 test records of which 14 contained seizures. Ensemble selection used training patients only, so no test data leaked into any design decision. Siena was handled with two patient-disjoint splits of 14 adult patients.

Is this ready for clinical seizure detection?

No. Everything ran in software on modest cohorts, the Siena false-alarm rates reached 0.65 to 0.74, and the metrics were record-level, not the event-level latency and false-alarms-per-hour measures clinical evaluation uses. The author frames the work as motivation for larger studies and hardware deployment, not as a deployable detector.

What does this have to do with organoid intelligence?

Two things. First, it is gradient-free training of spiking networks, the only realistic training paradigm for living tissue. Second, it suggests that per-cell biological heterogeneity, like the adaptation currents real neurons carry, may mainly help a training process find good solutions rather than power the finished computation, which reframes what an organoid computer's biological richness is for.

Can evolution like this train an actual organoid?

Not directly at anything like this budget. The study used 100 networks over 30 generations with cheap in-silico fitness evaluation. One fitness evaluation on living tissue costs far more, so the search would need to run on surrogate models of the culture first, with all the sim-to-real risk that entails. The design rule that survives transfer is per-channel, per-neuron tunability in the search space.

References

  1. Aksoy T. Evolved Spiking Neural Networks for EEG Seizure Detection: Calcium Dynamics and Delay Mechanics. arXiv:2611.03841 (q-bio.NC). 2026. https://arxiv.org/abs/2611.03841. Accessed 2026-11-04.
  2. Aksoy T, Shouval HZ. Active intrinsic conductances in recurrent networks allow for long-lasting transients and sustained activity with realistic firing rates as well as robust plasticity. Journal of Computational Neuroscience 50(1): 121-132. 2022. https://doi.org/10.1007/s10827-021-00797-2. Accessed 2026-11-04.
  3. Ghawaly J Jr, Nicholson A, Schuman C, Diez D, Young A, Witherspoon B. Exploring spiking neural networks for binary classification in multivariate time series at the edge. arXiv:2510.20997. 2025. https://arxiv.org/abs/2510.20997. Accessed 2026-11-04.
  4. Schuman CD, Mitchell JP, Patton RM, Potok TE, Plank JS. Evolutionary optimization for neuromorphic systems. Proceedings of the Neuro-inspired Computational Elements Workshop (NICE 2020), ACM, pp. 1-9. 2020. Accessed 2026-11-04.