Recent developments in wearable devices have made accurate and efficient seizure detection more important than ever. A challenge in seizure detection is that patient-specific models typically outperform patient-independent models. However, in a wearable device one typically starts with a patient-independent model, until such patient-specific data is available. To avoid having to construct a new classifier with this data, as required in conventional kernel machines, we propose a transfer learning approach with a tensor kernel machine. This method learns the primal weights in a compressed form using the canonical polyadic decomposition, making it possible to efficiently update the weights of the patient-independent model with patient-specific data. The results show that this patient fine-tuned model reaches as high a performance as a patient-specific SVM model with a model size that is twice as small as the patient-specific model and ten times as small as the patient-independent model.

Recent advancements in wearable EEG devices have highlighted the importance of accurate seizure detection algorithms, yet the ever-increasing size of the generated datasets poses a significant challenge to existing seizure detection methods based on kernel machines. Typically, this problem is mitigated by significantly undersampling the majority class, but in practice, these methods tend to suffer from too many false alarms. Recent works have proposed tensor networks to enable large-scale classification with kernel machines. In this paper, we explore the use of a probabilistic tensor method, the tensor-network Kalman filter for LS-SVMs (TNKF-LSSVM), for seizure detection, as we hypothesize that using more data will improve the detection performance. We show that the TNKF-LSSVM performs comparably to a regular LSSVM in detecting seizures when both are trained on the same dataset. Additionally, the TNKF-LSSVM can provide meaningful uncertainty quantification, and it is able to handle large-scale datasets beyond the capabilities of the LS-SVM (i.e., $N \gt 10 ^{5})$. However, for the presented model configuration detection performance does not seem to improve with more input data.

Epilepsy is one of the most common neurological conditions, affecting nearly 1% of the global population. It is defined by the seemingly random occurrence of spontaneous seizures. Anti-epileptic drugs provide adequate treatment for about 70% of patients. The remaining 30%, on the other hand, continue to have seizures, which has a significant impact on their quality of life as they are constantly unsure when these seizures will occur. Reliable seizure detection methods would thus have a significant impact on the lives of these patients. Despite ongoing research efforts involving academia and industry in large international collaborations, epileptic seizure detection and especially prediction is still an unsolved problem. The key to the solution could lie within ultralong-term, reallife datasets that are currently being generated using wearable sensors. However, due to the size of these datasets, conventional learning techniques such as least-square support vector machines (LS-SVMs) can become intractable. Therefore, this work proposes the use of a recently developed tensor network Kalman filtering approach for LS-SVMs (TNKFLSSVM) to detect epileptic seizures [1]. In the TNKF-LSSVM algorithm, the dual problem of the LS-SVM is solved using a recursive Bayesian filtering approach. This way the least-square problem can be solved row-by-row using a Kalman filter, thereby avoiding explicit matrix inversions, while also being able to provide confidence bounds on the estimates. By making use of the tensor-train format [2] to represent the matrices and vectors in the Kalman equations, it is even possible to avoid the construction of the (N + 1) × (N + 1) covariance matrix1. To be able to apply the TNKF-LSSVM algorithm for seizure detection there are still some issues that need to be tackled. One such problem is that the TNKF-LSSVM only performs well when the dataset is properly balanced, which is generally not the case for seizure datasets. Furthermore, for the TNKF-LSSVM to work efficiently for large scale problems the modes of the tensortrains representing the matrices and vectors should be as small as possible, thus it must hold that N + 1 = Q i ni, such that ni is ‘small’ for all i. To overcome both of these challenges we propose using the SMOTE method to oversample the seizure class, such that a balanced training set can be generated that has good factorization properties. Some preliminary results using a small subset of data from a public EEG dataset [3] show that taking the above considerations into account, the TNKF-LSSVM method can have performance that is competitive with a regular LS-SVM. Where the TNKFLSSVM method has the benefit of scaling log-linearly with the size of the dataset (in terms of memory usage) and can provide an uncertainty estimate of the detection. Future work will need 1N is the number of data points in the training set and 1 is added for the bias. to show whether this scaling up works as expected for the entire dataset.