🤖 machine learning

Detecting seizure onset and offset times using human intelligence: A critical-transitions-based approach

This paper introduces a robust seizure detection algorithm based on critical transitions that achieves near expert-level accuracy in identifying seizure onset and offset times across diverse morphologies and recording sessions, offering a versatile alternative to traditional machine learning approaches.

Original authors: Andrew Flynn, Cian McCafferty, Klaus Lehnertz, François David, Vincenzo Crunelli, Gordon Lightbody, Sebastian Wieczorek

Published 2026-07-30

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Andrew Flynn, Cian McCafferty, Klaus Lehnertz, François David, Vincenzo Crunelli, Gordon Lightbody, Sebastian Wieczorek

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine your brain as a bustling city where billions of tiny messengers (neurons) are constantly sending signals to keep you thinking, moving, and feeling. Usually, this city runs on a smooth, rhythmic traffic flow. But sometimes, for reasons we don't fully understand, the traffic lights all turn red at once, and a massive, chaotic gridlock erupts. In the medical world, this sudden, intense gridlock is called a seizure. For the 50 million people worldwide living with epilepsy, these episodes are unpredictable and can be dangerous. To help them, doctors and scientists use special sensors to listen to the brain's electrical chatter, hoping to spot the exact moment a seizure starts and stops.

The big challenge is that the brain is messy. Just like a city has construction zones, loud concerts, and random power surges, the brain has "noise" and weird electrical blips that look like seizures but aren't. For years, scientists have tried to build computer programs to automatically spot these seizures. Many of these programs use "machine learning," which is like teaching a computer by showing it thousands of examples until it figures out the pattern. But these programs can be like magic black boxes: they might work great, but no one knows why they made a specific decision, and sometimes they get confused by the noise or even make up patterns that aren't there. This paper asks a different question: instead of teaching a computer to guess, can we use simple, clear math to spot the exact moment the brain's traffic goes from "smooth flow" to "chaotic gridlock"?

This paper introduces a new way to catch seizures using a concept called "critical transitions." Think of a critical transition like a dam breaking. Before the break, the water is calm and low (the "non-seizure state"). Then, suddenly, the water level rises, the pressure builds, and the dam gives way, flooding the valley (the "seizure state"). The moment the dam breaks is the "critical transition." The authors, a team of mathematicians and neuroscientists, built a simple algorithm that acts like a very strict security guard watching a water gauge. They don't need to know the complex history of the water or use a black-box AI. They just watch the voltage (the electrical signal) and look for two specific things: a sudden jump up past a high line, and then a sustained stay above a lower line for a specific amount of time. If the signal jumps and stays up, the guard sounds the alarm: "Seizure starting!" If it drops and stays down, the guard says: "Seizure over."

The researchers tested this "security guard" on recordings from rats with a specific type of epilepsy. They compared the guard's alarms to the notes made by human experts who had carefully watched the same recordings. The results were surprisingly good. When they tuned the guard's sensitivity just right for each specific rat, the algorithm matched the experts' notes almost perfectly, catching the start and end of seizures with high accuracy. Even better, they found a single set of rules that worked well for all the rats, even though the rats had different seizure styles and sizes. This suggests that this math-based approach is robust and doesn't get easily fooled by the brain's background noise or the weird, fake signals that often trip up other detectors.

The paper also explicitly argues against relying solely on complex machine learning methods that lack "explainability." The authors point out that while AI is powerful, it can sometimes "hallucinate" or invent patterns that look real but aren't, and it's hard to trust a system you can't understand. Their approach, by contrast, is transparent: you can see exactly which voltage thresholds triggered the alarm. They also show that their method doesn't need to be retrained for every new patient or every new type of seizure; it just needs the right mathematical settings. While they don't claim to have solved epilepsy or replaced doctors, they demonstrate that a simple, math-driven tool can be just as effective as complex AI at spotting seizures, offering a reliable, clear, and versatile alternative for the future of brain monitoring.

Technical Summary: Detecting Seizure Onset and Offset Times Using Human Intelligence: A Critical-Transitions-Based Approach

Problem Statement
Current seizure detection algorithms face significant challenges in clinical and research settings. Many rely on extensive pre-processing, heuristic rules, or "black-box" machine learning techniques that lack explainability and may generate plausible but false predictions (hallucinations). These methods often struggle to balance sensitivity and specificity when confronted with variable seizure morphologies, interictal epileptiform discharges (IEDs), and recording artefacts. The variability in seizure characteristics across individuals, recording sites, and sessions further complicates the development of robust, generalizable detection tools.

Methodology
The authors propose an alternative seizure detection algorithm grounded in the mathematical theory of Critical Transitions (CTs). Rather than relying on machine learning, the approach models seizure onset and offset as sudden, large changes in the state of a complex system (the brain), transitioning between a non-seizure (NS) state and a seizure (S) state.

Core Concept: The algorithm defines the NS state as characterized by small-amplitude, weakly correlated fluctuations around 0 mV, and the S state by large-amplitude, strongly correlated fluctuations. Seizure onset and offset are identified as CTs between these two states.
Algorithm Mechanics: Inspired by a "non-ideal relay," the algorithm utilizes a moving window analysis with two voltage thresholds: an upper threshold ( $\alpha$ $α$ ) and a lower threshold ( $\beta$ $β$ ).
- Seizure Onset (NS $\to$ S): Detected when the absolute voltage exceeds $\alpha$ , followed by frequent crossings of a lower threshold $\beta$ for a minimum duration $\tau_S$ .
- Seizure Offset (S $\to$ NS): Detected when the voltage falls below $\beta$ and remains below the upper threshold $\alpha$ for a minimum duration $\tau_{NS}$ .
- Artefact Mitigation: To distinguish true CTs from high-frequency artefacts (e.g., muscle movement), the algorithm incorporates a condition checking the rate of change between consecutive data points. If the voltage jumps rapidly between thresholds (indicating noise), the event is flagged as an artefact rather than a seizure onset.
Evaluation Framework: The algorithm's performance is evaluated against expert annotations of seizure onset ( $\tau_1$ $τ_{1}$ ) and offset ( $\tau_2$ $τ_{2}$ ) times in Local Field Potential (LFP) recordings from Genetic Absence Epilepsy Rats from Strasbourg (GAERS).
- Classification: Time intervals detected by the algorithm ( $I^{(A)}_S$ ) are compared to expert intervals ( $I^{(E)}_S$ ) using True Positive (TP), False Positive (FP), True Negative (TN), and False Negative (FN) classifications. Overlap parameters ( $\rho_S, \rho_{NS}$ ) determine the strictness of the match required for a TP/TN classification.
- Metrics: Performance is quantified using Receiver-Operating-Characteristic (ROC) curves, specifically the Area Under the Curve (AUC) and the Distance from Ideal (DFI), which measures the Euclidean distance of a point on the ROC curve from the ideal performance point (0, 1).
- Parameter Tuning: The study systematically varies algorithm parameters ( $\alpha, \beta, \tau_S, \tau_{NS}, \tau_w$ ) to identify optimal settings for specific recording sessions and to derive a "general" set of parameters applicable across all sessions.

Key Results
The study was conducted on voltage recordings from multiple GAERS with varying seizure morphologies and significant artefact presence.

Parameter Sensitivity: Performance is highly sensitive to the choice of parameters. For a specific recording session (T2M), the optimal parameter set ( $P_3$ with $\delta_{\alpha\beta}=0.01$ ) yielded an AUC of approximately 0.975 and a DFI of 0.09, indicating near-ideal performance.
Generalizability: A single "general" set of parameters (derived from the average of optimal $\alpha$ values across sessions) was tested. While performance varied slightly across different rats (e.g., lower performance in Rat S due to significantly smaller voltage amplitudes), the algorithm maintained high accuracy (average AUC $\approx$ 0.89 with $\rho=0$ ) across diverse recording sessions without re-tuning.
Robustness: The algorithm successfully distinguished seizure activity from interictal epileptiform discharges and artefacts, maintaining robustness despite the variability in seizure morphology and the presence of noise.
Metric Comparison: The study highlights that relying solely on AUC can be misleading; the DFI metric provides complementary insight, particularly in identifying the specific parameter values that yield the best single-point performance.

Significance and Claims
The paper claims to demonstrate that seizure detection can be effectively achieved through a mathematics-based approach rooted in the theory of critical transitions, without reliance on heuristic rules or unexplainable machine learning models.

Explainability: The algorithm is fully transparent, based on defined mathematical states and transitions, offering an alternative to "black-box" AI methods.
Complementarity: The authors suggest this approach can complement existing machine learning techniques, potentially improving overall system performance by providing a robust, explainable baseline.
Versatility: The method shows potential for application beyond epilepsy, detecting critical transitions in other complex systems (e.g., thermoacoustic systems, aircraft wings), though the paper focuses specifically on its efficacy in detecting epileptic seizures in rodent models.
Modesty: The authors acknowledge that while the algorithm performs well, it currently monitors only a single feature (voltage amplitude). They note that future performance improvements could be realized by incorporating additional features (e.g., variance, autocorrelation) or adaptive thresholds, but emphasize that the current single-feature approach already achieves near-expert-level performance in many cases.

More like this