Epilepsy is the third most common neurological disorder in the United States, affecting about 3.4 million people – 1% of the population. One of the biggest challenges articulated by the epilepsy community is the unpredictability of seizures. While epileptic seizures do follow patterns that can broadly be described, pinpointing when exactly a seizure will occur in the moments before is a more challenging issue. Despite this, there is growing evidence that human biomarkers may show warning signs before a seizure occurs.

One example of this is seizure dogs, whose increased olfactory capabilities are able to be used to distinguish between seizure and non-seizure sweat, most likely due to the presence of volatile organic compounds (VOCs). A 2024 paper suggests that dogs were able to predict seizures from sweat samples in 94% of epileptic events. These findings indicate that there may be a measurable biological change that has strong predictive capability for seizures. In this sense, artificial intelligence methods can be employed in order to learn the features of these biomarkers for prediction.

Recent research also examines the most relevant models needed to predict seizures from EEG signals. After conducting a literature review, authors identified a study employing a deep learning model, known as long short-term memory (LSTM), that can detect seizure patterns in brain wave data within 15-120 minutes before a seizure occurs with a sensitivity of 99%. The study found LSTM models achieved an average accuracy of 93%, with average warning time of 21 minutes. Using another deep learning model known as a transformer, an additional study referenced by the authors was able to achieve a prediction sensitivity of 96%, with a warning time of 3 to 30 minutes. This high performance originates from the unique ability of these models to work with time-series data. For LSTMs, this is due to the model’s ability to process sequences iteratively while capturing long-term dependencies. For transformers, this is due to the direct connections between different tokens within the sequence. While these results are promising, there may be difficulty in using EEG data in the real world, as these procedures are considered invasive.

Using non-invasive methods, a 2025 research paper showed that autonomic nervous system (ANS) biomarkers, such as heart rate and temperature, can also be useful for predicting changes shortly before the seizure. Prior methods for predicting seizures using biomarkers obtained through wearable devices include supervised learning models, which are trained on using wearable data as input features. LSTMs are also used here, being trained on raw and Fourier-transformed data for sequence learning. In this paper, the authors hypothesize that combining features that include complementary information could improve the prediction accuracy. Specifically, the authors propose a system using LSTMs for processing sequences and deep canonical correlation analysis (DCCAEs) for finding correlations between ANS subsystems. This system was used in conjunction with a growing neural network, which allowed for the system to integrate different features from these regions as the architecture of the machine learning model was changed. Through this, they found that combining multiple biological signals significantly improved prediction accuracy, raising it from about 74% to 82%. This suggests that seizures may affect several body systems at once, and models that integrate multiple data streams perform better.

Looking ahead, researchers hope to develop wearable systems that employ transfer learning and growing neural networks to continuously learn and adapt to each individual’s seizure patterns. By combining brain activity, heart rate, and temperature, future devices may provide earlier and more reliable warnings. Returning to the past idea of VOCs, this hybrid of generalizable models and wearable methods can play a key role in creating a new, non-invasive input stream for seizure prediction. If successful, these technologies could transform seizure prediction from a clinical tool into an everyday safety system.