In our recent work, I focused on three-phase seizure segmentation: identifying ictal onset, intra-ictal transition, and seizure termination directly from SEEG recordings. Rather than treating seizures as a single undifferentiated event, this framing emphasizes their temporal structure and makes the analysis more useful for presurgical interpretation.
Clinical review of intracranial EEG is still heavily dependent on expert visual inspection. That process is valuable, but it is also time-intensive and can vary across readers. A phase-based computational framework can support that review by providing consistent temporal boundaries that are interpretable and easy to audit.
Precise phase segmentation also helps with downstream analysis. Once onset, transition, and termination are localized in time, it becomes easier to study propagation dynamics across channels, compare seizure organization across patients, and relate electrophysiology to surgical hypotheses.
A seizure is a dynamic signal, so changepoint detection is a natural fit. In this project, I used envelope-based multivariate features together with the Pruned Exact Linear Time (PELT) algorithm to detect boundaries between seizure phases. The goal was not only good performance, but also a method that remains physiologically interpretable.
The feature set combines signal energy and spectral structure, including RMS envelope, relative bandpower, line length, and spectral entropy. That combination gives the model access to amplitude shifts, rhythmic structure, and changes in signal complexity without turning the pipeline into a black box.
What I find most promising is that the framework supports both temporal precision and clinical interpretability. If an algorithm proposes a boundary, it should be possible to inspect the relevant channels, understand the feature behavior, and decide whether the output is clinically meaningful.
That balance matters in epilepsy research. Methods that are accurate but opaque are harder to trust in high-stakes settings. Methods that are interpretable but too weak to be useful do not help enough. I am interested in approaches that move both of those constraints in the right direction.
This line of work connects naturally to broader questions about seizure propagation, seizure onset zone localization, and representation learning from intracranial EEG. The published paper is one step in that direction, and I expect future work to build on it with richer multichannel modeling and tighter integration with clinical decision support.
If you are interested in the paper itself, it is now online via Springer and is also listed on my publications page.
]]>That makes changepoint detection a natural methodological choice. Rather than asking a model to label every time point independently, a changepoint framework asks a more structured question: when does the signal statistically change in a meaningful way?
SEEG recordings often contain abrupt or progressive shifts in amplitude, rhythmic organization, spectral content, and synchrony across channels. Those transitions are exactly what clinicians inspect visually when they review seizures. A changepoint model tries to formalize that process.
This is appealing for two reasons. First, it produces outputs that are easy to interpret: proposed boundary times. Second, it keeps the analysis close to the temporal logic of the clinical problem instead of forcing the data into a less natural classification setup.
Of course, not every change in the signal is clinically meaningful. Intracranial EEG is noisy, patient-specific, and heterogeneous across seizures. A useful changepoint system has to distinguish true physiological transitions from incidental fluctuations.
That is why feature design matters. If the input representation captures relevant seizure dynamics, then changepoint detection becomes much more informative. If the features are poorly matched to the phenomenon, even a sophisticated optimization algorithm will return boundaries that are technically valid but clinically unhelpful.
I am especially interested in changepoint methods because they sit in a productive middle ground. They are more formal and scalable than pure visual review, but they do not have to become opaque black-box predictors.
If a model identifies a seizure onset boundary, we should be able to inspect the channels and the feature trajectories that drove that decision. In a clinical context, that level of transparency is not optional. It is part of what makes a computational result worth trusting.
I expect changepoint methods to remain useful not only for seizure phase segmentation, but also for studying propagation structure, functional reorganization, and other temporally evolving phenomena in intracranial EEG.
For me, the key question is not whether changepoint detection should replace clinical expertise. It should not. The question is how to build tools that make temporal structure easier to quantify, inspect, and use.
]]>Here are a few pitfalls I think are especially important.
The most common failure mode is still data leakage. If windows from the same seizure or patient appear in both training and test sets, performance can become artificially optimistic very quickly.
In epilepsy applications, patient-level separation is usually the right default. A model that performs well only because it has already seen a very similar seizure from the same person is not telling us much about generalization.
Many neuroscience labels are approximate, noisy, or based on human interpretation. Seizure onset time, propagation boundaries, emotional state labels, and cognitive annotations often carry uncertainty that standard pipelines ignore.
That does not mean the labels are unusable. It means the model evaluation should reflect their limitations. If the ground truth is inherently uncertain by a few seconds, then a metric that pretends otherwise can be misleading.
EEG and SEEG are not just collections of independent samples. The order of events matters. Transitions matter. Context matters.
Pipelines that flatten everything into disconnected windows can miss the very structure that makes the problem interesting. Even when window-based analysis is necessary, it helps to think carefully about what temporal information is being discarded.
High accuracy is attractive, but in biomedical settings it is rarely enough on its own. Clinicians and collaborators usually want to know why a system made a prediction, which channels were informative, and whether the output aligns with known physiology.
This is one reason I often prefer methods that preserve some connection between features, signal dynamics, and decision boundaries. A model that is slightly less accurate but substantially easier to interpret can be more valuable in practice.
Referencing scheme, filtering, artifact handling, channel selection, and segmentation strategy can change results substantially. These steps are sometimes treated as routine housekeeping, but they are often part of the scientific question itself.
A pipeline is only as defensible as its preprocessing assumptions. If those assumptions are weak or poorly documented, the downstream machine-learning results are harder to trust.
The hardest part of EEG and SEEG machine learning is usually not the model architecture. It is making sure the data split, label definition, temporal framing, and evaluation criteria actually match the problem you claim to solve.
That is where most of the real rigor lives.
]]>