Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Efficient Patient Fine-Tuned Seizure Detection with a Tensor Kernel Machine

Seline J. S. de Rooij, Frederiek Wesel, Borbála Hunyadi

TL;DR

The paper addresses seizure detection on wearable EEG by tackling the data scarcity and inter-patient variability that hinder patient-specific models. It introduces Tensor Kernel Ridge Regression (T-KRR) with a weight tensor constrained to rank-$R$ CPD, learned in the primal form and approximated with Laplace-based kernels, enabling efficient updates via ALS. Transfer learning is realized by initializing the patient-fine-tuned (PF) model with the patient-independent (PI) weights and updating a subset of CPD factors, achieving performance on par with patient-specific SVMs while using substantially fewer parameters. This approach offers a memory-efficient and rapidly adaptable framework suitable for wearable devices, with the potential for online transfer learning in future work. The study demonstrates on behind-the-ear EEG data from temporal-lobe seizures that updating just one CPD factor can significantly boost performance and that PF matches PS SVM performance with about half the parameter count of PS SVM and an order of magnitude fewer than PI SVM, highlighting practical benefits for resource-constrained wearables.

Abstract

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.

Efficient Patient Fine-Tuned Seizure Detection with a Tensor Kernel Machine

TL;DR

The paper addresses seizure detection on wearable EEG by tackling the data scarcity and inter-patient variability that hinder patient-specific models. It introduces Tensor Kernel Ridge Regression (T-KRR) with a weight tensor constrained to rank- CPD, learned in the primal form and approximated with Laplace-based kernels, enabling efficient updates via ALS. Transfer learning is realized by initializing the patient-fine-tuned (PF) model with the patient-independent (PI) weights and updating a subset of CPD factors, achieving performance on par with patient-specific SVMs while using substantially fewer parameters. This approach offers a memory-efficient and rapidly adaptable framework suitable for wearable devices, with the potential for online transfer learning in future work. The study demonstrates on behind-the-ear EEG data from temporal-lobe seizures that updating just one CPD factor can significantly boost performance and that PF matches PS SVM performance with about half the parameter count of PS SVM and an order of magnitude fewer than PI SVM, highlighting practical benefits for resource-constrained wearables.

Abstract

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.
Paper Structure (15 sections, 8 equations, 2 figures, 1 table)

This paper contains 15 sections, 8 equations, 2 figures, 1 table.

Table of Contents

  1. Introduction
  2. Classification Method
  3. Notation and Preliminaries
  4. Kernel Machines
  5. Tensor Kernel Ridge Regression
  6. Transfer Learning with T-KRR
  7. Data Processing
  8. Data
  9. Preprocessing and Feature Extraction
  10. Results and Discussion
  11. Validation and Hyperparameter Tuning
  12. Classification Performance
  13. Model complexity
  14. Conclusion
  15. Acknowledgements

Figures (2)

  • Figure 1: AUC against the number of iterations of the ALS algorithm for T-KRR. Blue shows the AUROC and orange the AUPRC. Square dots represent the $\text{T-KRR}_{PF}$ model and round dots the $\text{T-KRR}_{PS}$ model with random initialization.
  • Figure 2: Barplot of the performance for the different classifiers. The error bars show the 95% confidence interval.