The idea of a device that could warn someone a seizure is coming before it actually starts sounds like science fiction, but it’s become one of the most active areas of epilepsy research happening right now. The honest answer to whether AI can do this is: it’s much closer to solved for detecting seizures as they happen than for genuinely predicting them in advance, and understanding that distinction matters for making sense of where the technology actually stands.
Detection and prediction are two different problems, and it’s worth separating them clearly. Detection means recognizing that a seizure is currently happening, or just happened, and alerting someone so help can arrive quickly. Prediction means forecasting that a seizure is about to happen before it starts, giving someone time to get somewhere safe or take preventive action. Detection is the technology that’s actually made it into real medical devices people can buy today. Prediction is still overwhelmingly a research problem, with genuinely promising results in lab settings but a long way to go before it’s something a person could rely on day to day.
On the detection side, there’s already an FDA-cleared, commercially available product doing this successfully. A smartwatch-based system originally developed at the MIT Media Lab uses machine learning to monitor electrodermal activity, essentially subtle changes in skin conductivity tied to the nervous system, along with motion and heart rate data, to identify generalized tonic-clonic seizures as they’re happening. In a supporting clinical trial involving 135 patients monitored simultaneously with video-EEG in epilepsy monitoring units, the device’s algorithm detected all 40 of the tonic-clonic seizures that occurred during the study. Later versions of the system have reported around 98 percent accuracy with a low false alarm rate in broader use. When the device detects a likely seizure, it automatically alerts a designated caregiver by call and text, including the wearer’s location, which matters enormously for reducing the risk associated with someone having a seizure alone. This kind of technology has a direct connection to reducing SUDEP risk, since prompt attention during or immediately after a convulsive seizure can genuinely be lifesaving, and a a device that can summon help within moments closes a gap that used to depend entirely on someone else happening to be nearby and paying attention.
Prediction is the harder, more ambitious problem, and it’s where most of the current AI research energy is actually going. The basic idea is to catch subtle changes in brain activity, sometimes minutes, sometimes longer, before a seizure actually begins, using either EEG signals directly or other physiological data as a stand-in. Deep learning models, including convolutional and recurrent neural networks and more recently transformer-based architectures, have been trained on large EEG datasets to try to recognize these pre-seizure patterns. Some of this research is producing genuinely strong numbers: one recent single-channel EEG model reported prediction accuracy above 90 percent and sensitivity near 98 percent on a benchmark dataset, successfully forecasting the large majority of seizures within its prediction window. Numbers like that sound close to solved, but there’s an important asterisk attached to almost all of them.
The asterisk is generalization. Seizure prediction models tend to perform very well when trained and tested on the same patient’s own data, since every person’s pre-seizure brain activity has its own distinct signature. The much harder challenge is building a model that works well across many different people without needing to be individually retrained for each one, and this is where current research consistently runs into trouble. A model that hits 95 percent accuracy for one patient’s data might perform far worse on someone else’s, and that inconsistency is one of the main reasons prediction technology hasn’t made the leap from research papers into consumer devices the way detection has. Researchers are actively working on this through approaches like federated learning, which allows models to learn from data across many patients without that data ever leaving each individual’s device, and synthetic data generation, used to fill in gaps for the seizure types and patient groups that are underrepresented in existing datasets.
There’s also a deeper, more uncomfortable problem underneath all of this: even human experts don’t fully agree on what counts as seizure activity when reading the same EEG. In one major effort to build a seizure detection algorithm for critical care settings, the team training the model had nine board-certified epileptologists independently annotate the same EEG snippets, and their agreement with each other averaged only around 55 percent. If trained specialists looking at the same data don’t consistently agree on what they’re seeing, it puts a natural ceiling on how confidently any AI model, no matter how sophisticated, can be judged against a “correct” answer that isn’t always fully settled to begin with.
Newer research directions are also branching out beyond scalp EEG entirely. Some groups are experimenting with in-ear electrodes as a far less obtrusive way to continuously monitor brain activity outside of a hospital setting, aiming for something closer to a hearing aid than a cap full of wires. Others are combining EEG with additional signals, like heart rate variability, skin temperature, and movement, betting that the fusion of multiple data streams will end up more reliable than any single signal alone, since seizures don’t just show up in brain activity, they show up as changes throughout the whole body.
None of this means a reliable, everyday seizure prediction device is right around the corner, but it also isn’t decades away in the way it might have sounded a decade ago. Detection has already crossed from research into real, life-saving products people wear today. Prediction is still working through the harder problem of generalizing across different people and different seizure types, but the pace of published research in just the last couple of years, including models specifically designed to run efficiently enough for a small wearable rather than a hospital computer, suggests this is a genuinely active, fast-moving field rather than a stalled one. For a condition defined by its unpredictability, even partial progress toward forecasting it represents a real shift in how epilepsy might eventually be managed.
References
Artificial intelligence-based epileptic seizure prediction strategies: A review. MDPI.
Kerr, W. T., et al. (2024). The present and future of seizure detection, prediction, and forecasting with machine learning, including the future impact on clinical trials. Frontiers in Neurology.
Real-time EEG-based epileptic seizure prediction using artificial intelligence: A systematic review. MedRxiv.
Ho, T. J., Ostrem, B. E. L., & Hillis, J. M. (2026). Artificial intelligence in wearable seizure detection devices: Current technologies and future directions. Frontiers in Neurology.
Jang, D., Jung, K.-Y., Jeon, Y.-G., et al. (2026). Single-channel EEG-based seizure prediction using deep learning. Scientific Reports.
Wearable artificial intelligence for epilepsy: Scoping review. Journal of Medical Internet Research.
Empatica. EpiMonitor and Embrace2 seizure monitoring. Empatica.com.
NeurologyLive. (2018). Embrace smartwatch gets FDA go-ahead for seizure monitoring. Neurologylive.com.