Evolved Spiking Neural Networks for EEG Seizure Detection: Calcium Dynamics and Delay Mechanics

arXiv:2611.03841 · q-bio.NC, cs.NE · Submitted 2026-11-02 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Genomics Radio. Generated commentary on the latest computational biology and genomics papers.

Ines: We're looking at the paper: "Evolved Spiking Neural Networks for EEG Seizure Detection".

Marcus: Here's the gist of the paper. Evolving calcium parameters per neuron in integrate-and-fire networks gave seizure detectors with higher MCC and fewer false alarms on held-out CHB-MIT patients, and smaller gains on Siena with unchanged settings Core Methodology The EONS framework evolves the topology, weights, thresholds and delays of neuromorphic spiking networks without gradients The study compares integrate-and-fire networks with those having a calcium current whose parameters are fixed (Ca-IFfix), evolved per network (Ca-IFnet), or evolved per neuron (Ca-IFneu) Synaptic delays were not evolved in the main experiments but fixed at the minimum transmission time of one tick The data sources used are the CHB-MIT Scalp EEG database and the Siena Scalp EEG database Each 0.5 s window of the 18 channels was encoded as a spike train of 20 ticks, giving 36 input neurons Performance Gains on CHB-MIT Evolving the calcium parameters per neuron (Ca-IFneu) raised MCC over IF by +0.134 (p = 0.008, adjusted p = 0.025) and lowered the false-alarm rate from 0.17 to 0.10 Ca-IFfix and Ca-IFnet gave no significant gain over IF The false-alarm reduction was consistent across patients, with Ca-IFneu lowering the FPR in 7 out of 10 held-out patients Role of Delays and Time Constants Short evolved synaptic delays, specifically those of 1–3 ticks (25–75 ms), gave a similar gain to Ca-IFneu on CHB-MIT (adjusted p = 0.024) In contrast, longer or fixed delays and axonal delays gave no significant gain The default calcium time constant of 4 ticks provided the largest gain over IF (p = 0.008, adjusted p = 0.025) Generalizability and Limitations On Siena, Ca-IFneu gained +0.059 over IF in one patient-disjoint split (p = 0.052), and +0.096 in the other split (p = 0.011) However, scoring CHB-MIT-trained networks on Siena without retraining resulted in lower MCC scores than on CHB-MIT, and the Ca-IFneu advantage was not significant A primary limitation of this study is the modest cohort size and the limited number of seizure records Evolving calcium parameters per neuron in integrate-and-fire networks gave seizure detectors with higher MCC and fewer false alarms on held-out CHB-MIT patients, and smaller gains on Siena with unchanged settings Both calcium and evolved delays may act by helping evolution find better networks, not by changing how the finished networks work These software results on small cohorts motivate larger evaluations and, next, neuromorphic hardware Declarations Data availability.

Ines: The CHB-MIT Scalp EEG Database

16, 17: and the Siena Scalp EEG Database

18, 19: are publicly available from PhysioNet

20: Code availability.

Marcus: Code is available from the author on reasonable request Funding.

Ines: This work received no external funding Competing interests.

Title and authors: Ines: So we're summarizing this paper now on "Evolved Spiking Neural Networks for EEG Seizure Detection: Calcium Dynamics and Delay Mechanics". Basically they showed that when you let the network evolve the calcium parameters per neuron you get a better seizure detector for patients they haven't seen before.

Yuki: From a population genetics angle I see this as evolution finding a way to encode biological variability into the network structure itself. It’s not just fitting data; it’s discovering how neurons are built to cope with those temporal patterns across different individuals.

Ines: Exactly they found that evolving both the calcium parameters and some of the synaptic delays between one and three ticks gave a similar boost over standard models on held-out patients. But they pointed out that axonal delays didn't help at all which tells us about how information flows in different directions in the brain.

Marcus: So it’s not just making one part better it’s finding a whole set of interconnected parameters the calcium dynamics and the synaptic timing that work together as a system to make a detector more accurate.

Yuki: That points to evolution searching for an architecture that reflects how neurons naturally organize themselves to handle these complex temporal patterns in EEG data. It suggests we can use evolution to discover new biological structures not just tune existing ones.

Ines: That's the big thing yeah the implication is that we might use these evolutionary methods not just to optimize a detection algorithm but to discover underlying principles of neural network organization itself.

Yuki: That means this specific structural discovery isn't universally guaranteed it’s tied to that particular cohort's characteristics and sample size and they need larger evaluations to see if this general structural discovery holds up everywhere.

The paper's summary: Ines: So we're looking at how these calcium parameter evolutions actually improve the network structure itself beyond just getting better statistics on held-out data.

Marcus: Right so they found that evolving those synaptic delays between one and three ticks gave a similar gain over the standard integrate-and-fire networks on CHB-MIT.

Yuki: That means temporal flexibility is a key thing here not just the intrinsic firing rate of the neurons. What does that suggest about how biological systems adapt their timing under stress?

Ines: It suggests that evolving those delays helps the network find better ways to handle temporal patterns which is important for capturing seizure activity accurately. But they also noted that axonal delays didn't help at all which tells us about how information flows in different directions in the brain.

Marcus: So it’s not just making one parameter better it’s about finding a set of interconnected parameters the calcium dynamics and the synaptic timing that work together as a system.

Yuki: That points to the idea that evolution is searching for an architecture that reflects how neurons naturally organize themselves to handle these complex temporal patterns in EEG data Does this mean we can use evolution to discover new biological structures?

Ines: It means the evolutionary process itself is finding better ways to incorporate both the timing of signals and the intrinsic activity dynamics of individual neurons which is a big thing.

Yuki: So it’s about guiding the search toward biologically plausible structures that moves us past just tuning numbers and towards discovering how networks are built in the first place.

Ines: Exactly The implication is that we might be able to use these evolutionary methods not just to optimize a detection algorithm but to discover underlying principles of neural network organization.

Marcus: But we have to remember the caveat they flagged scoring those CHB-MIT trained networks on Siena without retraining actually resulted in lower MCC scores than when they used the CHB-MIT data.

Yuki: That means the discovery is highly dependent on the training data you start with and it doesn't always transfer perfectly across different patient groups. The sample size and the limited number of seizure records are real constraints.

Ines: It’s a limitation they state plainly the method works best when applied to held-out data from that specific cohort and we need larger evaluations to see if this general structural discovery holds up universally.

The paper's improvements: Ines: So we’re wrapping up this look at "Evolved Spiking Neural Networks for EEG Seizure Detection: Calcium Dynamics and Delay Mechanics". Basically evolving calcium parameters per neuron really improved detection accuracy and lowered false alarms on held-out CHB-MIT patients.

Marcus: And they showed that evolving synaptic delays between one and three ticks gave a similar gain over the standard integrate-and-fire networks on CHB-MIT.

Yuki: It’s clear that incorporating activity dependent variables like calcium into these spiking models helps them capture more complex temporal structures in EEG data.

Ines: It sets a benchmark for how much structure we can discover through this evolutionary approach showing it can boost the Matthews correlation coefficient and lower that false alarm rate significantly.

Marcus: And while they saw gains with evolved synaptic delays too the most robust improvement came from the per-neuron calcium evolution.

Yuki: It’s a solid demonstration of how incorporating activity dependent variables like calcium into these spiking models can help them capture more complex temporal structures in EEG data.

Ines: Next up we'll talk about the implications of these findings for building even more complex neural models and the challenges ahead in hardware implementation.

Marcus: That’s what this paper on "Evolved Spiking Neural Networks for EEG Seizure Detection: Calcium Dynamics and Delay Mechanics" boils down to, with calcium dynamics per neuron giving better detection performance on held-out patients.

Yuki: And it shows that incorporating activity dependent variables like calcium into these spiking models helps them capture more complex temporal structures in EEG data.

Ines: We’re done for now.

Conclusion: Ines: So we're wrapping up our look at "Evolved Spiking Neural Networks for EEG Seizure Detection: Calcium Dynamics and Delay Mechanics". Basically, evolving calcium parameters per neuron really improved detection accuracy and lowered false alarms on held-out CHB-MIT patients.

Marcus: And we saw that evolving synaptic delays between one and three ticks gave a similar gain over the standard integrate-and-fire networks on CHB-MIT. It’s clear that incorporating activity dependent variables like calcium into these spiking models helps them capture more complex temporal structures in EEG data.

Yuki: From a population genetics angle, I see this as evolution finding a way to encode biological variability into the network structure itself. It’s not just fitting data; it’s discovering how neurons are built to cope with those temporal patterns across different individuals.

Ines: Exactly. They found that evolving both the calcium parameters and some of the synaptic delays between one and three ticks gave a similar boost over standard models on held-out patients. But they pointed out that axonal delays didn't help at all, which tells us about how information flows in different directions in the brain.

Marcus: So it’s not just making one part better; it’s finding a whole set of interconnected parameters—the calcium dynamics and the synaptic timing—that work together as a system to make a detector more accurate.

Yuki: That points to evolution searching for an architecture that reflects how neurons naturally organize themselves to handle those complex temporal patterns in EEG data. It suggests we can use evolution to discover new biological structures, not just tune existing ones.

Ines: That's the big thing, yeah. The implication is that we might use these evolutionary methods not just to optimize a detection algorithm, but to discover underlying principles of neural network organization itself.

Marcus: But we have to remember the caveat they flagged—scoring those networks on new patient data without retraining actually resulted in lower performance than when they used the CHB-MIT data. The discovery is highly dependent on the training data you start with.

Yuki: That means this specific structural discovery isn't universally guaranteed; it’s tied to that particular cohort's characteristics and sample size. The method works best when applied to held-out data from that specific group, and they need larger evaluations to see if this general structural discovery holds up everywhere.

Ines: So for someone listening, these models aren't just fancy filters; they are finding better ways to build the internal structure of a spiking network based on calcium activity and signal timing.

Marcus: That sets a new benchmark for what kind of complexity we can discover through this evolutionary approach before we move into trying to build these things on real hardware.

Yuki: It’s about guiding the search toward biologically plausible structures. That moves us past just tuning numbers and towards discovering how networks are built in the first place.

Ines: Exactly. The implication is that we might be able to use these evolutionary methods not just to optimize a detection algorithm, but to discover underlying principles of neural network organization.

Marcus: So that’s the gist of this paper on "Evolved Spiking Neural Networks for EEG Seizure Detection: Calcium Dynamics and Delay Mechanics". It shows how structural evolution helps build better detectors.

Ines: Exactly. And it opens up a lot of questions about whether we can use evolution to discover fundamental principles of neural network organization itself.

Yuki: It’s not just finding a good set of weights; it’s finding a good way to build the architecture. That is where the real interest lies for us in population genetics and modeling complex systems.

Ines: Right, so while this method works well on held-out patients from that specific cohort, transferring those learned structures directly to another cohort without retraining doesn't guarantee the same level of performance, which is a limitation they noted in the paper.

Marcus: So for someone listening, it means these models aren't just fancy filters; they are finding better ways to build the internal structure of a spiking network based on calcium activity and signal timing.

Yuki: And that’s what we need to keep watching as we try to model even more complex biological systems.

q-bio.NC, cs.NE

Submitted: 2026-11-02

Updated: 2026-11-02

Comments: 16 pages, 5 figures, 4 tables, plus three supplementary tables and one supplementary figure

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 80/100

The gist: Evolving calcium parameters per neuron in integrate-and-fire networks gave seizure detectors with higher MCC and fewer false alarms on held-out CHB-MIT patients…

Key concepts

Integrate-and-fire networks
These are spiking neural networks where neurons accumulate incoming signals (like EEG data) until they reach a certain threshold, at which point they fire an output spike. The study focused on how the internal calcium dynamics within these neurons affect seizure detection performance.
Calcium parameters per neuron (Ca-IFneu)
This method involves evolving the specific calcium dynamics—the way a neuron's internal calcium level changes over time—for each individual neuron in the network. This evolution was found to be effective at improving seizure detection metrics compared to fixed or evolved synaptic parameters.
Matthews Correlation Coefficient (MCC)
MCC is a metric used to evaluate the performance of a binary classifier, like a seizure detector. A higher MCC score indicates that the network is better at correctly identifying both positive and negative cases (seizures and non-seizures) simultaneously.
Synaptic delays
These represent the time lag between when one neuron sends a signal to another. The study found that evolving short synaptic delays (1–3 ticks) provided similar performance gains on CHB-MIT data, suggesting temporal flexibility is also beneficial for network evolution.

Terminology

Summary

Evolving calcium parameters per neuron in integrate-and-fire networks gave seizure detectors with higher MCC and fewer false alarms on held-out CHB-MIT patients, and smaller gains on Siena with unchanged settings

Core Methodology

The EONS framework evolves the topology, weights, thresholds and delays of neuromorphic spiking networks without gradients The study compares integrate-and-fire networks with those having a calcium current whose parameters are fixed (Ca-IFfix), evolved per network (Ca-IFnet), or evolved per neuron (Ca-IFneu) Synaptic delays were not evolved in the main experiments but fixed at the minimum transmission time of one tick The data sources used are the CHB-MIT Scalp EEG database and the Siena Scalp EEG database Each 0.5 s window of the 18 channels was encoded as a spike train of 20 ticks, giving 36 input neurons

Performance Gains on CHB-MIT

Evolving the calcium parameters per neuron (Ca-IFneu) raised MCC over IF by +0.134 (p = 0. p. 6008, adjusted p = 0. p. 6 p. 6025) and lowered the false-alarm rate from 0. p. 617 to 0.10 Ca-IFfix and Ca-IFnet gave no significant gain over IF The false-alarm reduction was consistent across patients, with Ca-IFneu lowering the FPR in 7 out of 10 held-out patients

Role of Delays and Time Constants

Short evolved synaptic delays, specifically those of 1–3 ticks (25–75 ms), gave a similar gain to Ca-IFneu on CHB-MIT (adjusted p = 0.024) In contrast, longer or fixed delays and axonal delays gave no significant gain The default calcium time constant of 4 ticks provided the largest gain over IF (p = 0. p. 9008, adjusted p = 0.025)

Generalizability and Limitations

On Siena, Ca-IFneu gained +0.059 over IF in one patient-disjoint split (p = 0. p. 0052), and +0. p. 8096 in the other split (p = 0. p. 0011) However, scoring CHB-MIT-trained networks on Siena without retraining resulted in lower MCC scores than on CHB-MIT, and the Ca-IFneu advantage was not significant A primary limitation of this study is the modest cohort size and the limited number of seizure records

Conclusion

Evolving calcium parameters per neuron in integrate-and-fire networks gave seizure detectors with higher MCC and fewer false alarms on held-out CHB-MIT patients, and smaller gains on Siena with unchanged settings Both calcium and evolved delays may act by helping evolution find better networks, not by changing how the finished networks work These software results on small cohorts motivate larger evaluations and, next, neuromorphic hardware

Declarations

Data availability. The CHB-MIT Scalp EEG Database [16, 17] and the Siena Scalp EEG Database [18, 19] are publicly available from PhysioNet [20] Code availability. p. 1 Code is available from the author on reasonable request Funding. p. 2 This work received no external funding Competing interests. p. 1

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Improve network design by evolving calcium parameters per neuron to yield higher Matthews correlation coefficient (MCC) and lower false-alarm rate (FPR) on held-out CHB-MIT patients, as reported by Ca-IFneu raised MCC over IF (+0.134, p = 0.008) and lowered the false-alarm rate from 0.17 to 0.10.

  2. Enhance temporal flexibility by evolving synaptic delays of 1 to 3 ticks per synapse in integrate-and-fire networks, as these evolved delays gave a similar gain over IF networks on CHB-MIT, whereas axonal delays did not.

  3. Increase robustness against parameter uncertainty by testing whether calcium and delays may make the evolutionary search more effective, suggesting that calcium and delays may make the evolutionary search more effective.

Sources

Related papers