The problem
A neural biocomputer is a culture of living neurons grown across a Multi-Electrode Array — a grid of electrodes that both stimulates the culture and records from it. The raw signal at each electrode is a continuous voltage trace. Buried in it are spikes: the action potentials that constitute the culture's actual computation.
Everything downstream depends on detecting those spikes correctly. Burst analysis, connectivity mapping, closed-loop stimulation protocols — all of it is built on top of a spike train. If detection is wrong, every conclusion above it is wrong too.
The conventional answer is a hard voltage threshold, usually some multiple of the median absolute deviation of the noise floor. It is simple and fast, and it fails in exactly the situations that matter: when the noise floor drifts, when electrode impedance varies across the array, and when one culture behaves nothing like the next.
The obvious alternative — train a deep classifier — trades one problem for another. It detects well and explains nothing. In a domain where the object of study is the biology, a model that cannot tell you which electrodes and which firing patterns drove its decision has answered the wrong question.
The system had to be accurate enough to trust and transparent enough to interrogate — simultaneously, and across chips whose biology differs enough that a model tuned to one may not transfer to another.
Approach
I built a fuzzy rule-based classifier: a system whose entire decision surface is a set of IF-THEN rules over linguistic variables. Rather than approximating interpretability after the fact with a post-hoc explainer, the model is the explanation.
1. Feature extraction
Windowed features are computed per electrode from the filtered trace — amplitude, slope, and shape descriptors that distinguish a genuine action potential from a noise excursion of similar magnitude.
2. Feature selection with ANNIGMA
Fuzzy systems degrade badly as the input dimension grows: the rule base explodes combinatorially and interpretability evaporates. ANNIGMA (Artificial Neural Net Input Gain Measurement Approximation) ranks inputs by the gain they contribute through a trained network's weights, letting me cut the feature set to those that actually carry signal before any rules are generated.
3. Rule generation and GA weight optimisation
Rules are generated over the surviving features, then each rule is assigned a weight optimised by a genetic algorithm. This matters more than it sounds: weight optimisation is what lets a compact rule base compete with an opaque model on raw accuracy, and it is what keeps the rule count low enough to read.
Results
Evaluated across six biologically diverse chips, the classifier reached a best F1-score of 97.74% with a rule base of 337 interpretable IF-THEN rules. Every classification decision traces to a specific set of fired rules, and every rule names the electrodes and firing characteristics it depends on.
| Property | Result |
|---|---|
| Best F1-score | 97.74% |
| Chips evaluated | 6, biologically diverse |
| Rule base size | 337 IF-THEN rules |
| Rule weighting | Genetic algorithm optimised |
| Feature selection | ANNIGMA gain ranking |
| Explanation type | Intrinsic — no post-hoc approximation |
fuzzy classifier vs. threshold baseline across the six chips
Why the rules matter
The accuracy number is the part that gets a paper accepted. The rules are the part that makes the work useful.
Because each rule names the electrodes and firing patterns associated with spike events, the rule base is itself a description of the culture — which electrodes carry the informative activity, and what that activity looks like. A neuroscientist can read it, disagree with it, and test the disagreement. That is not a property you can retrofit onto a convolutional network with a saliency map.
It also changes the failure mode. When a black-box detector performs poorly on a new chip, the diagnosis is guesswork. When a rule-based detector performs poorly, you can look at which rules fired and see the reason.
Where it goes next
- Interval Type-2 fuzzy sets — Type-1 membership functions assume the uncertainty in a linguistic term is itself precise. IT2 sets model that second-order uncertainty directly, which suits biological variability better. A separate IT2 classifier exists as groundwork.
- Cross-chip transfer — quantifying how much of a rule base learned on one culture survives on another, and which rules are chip-specific.
- Closed-loop use — moving from offline detection to a stimulation loop where detection latency becomes a hard constraint.
Publication
“A Fuzzy-Based Approach for Interpretable Spike Detection in Living Neural Biocomputers”
IEEE World Congress on Computational Intelligence, FUZZ-IEEE 2026 — Maastricht, Netherlands.
Supervised by Prof. Hani Hagras and Dr. Michael Barros, University of Essex.