One-shot Learning for iEEG Seizure Detection Using End-to-end Binary Operations: Local Binary Patterns with Hyperdimensional Computing

Abstract: This paper presents an efficient binarized algorithm for both learning and classification of human epileptic seizures from intracranial electroencephalography (iEEG). The algorithm combines local binary patterns with brain-inspired hyperdimensional computing to enable end-to-end learning and inference with binary operations. The algorithm first transforms iEEG time series from each electrode into local binary pattern codes. Then atomic high-dimensional binary vectors are used to construct composite representations of seizures across all electrodes. For the majority of our patients (10 out of 16), the algorithm quickly learns from one or two seizures (i.e., one-/few-shot learning) and perfectly generalizes on 27 further seizures. For other patients, the algorithm requires three to six seizures for learning. Overall, our algorithm surpasses the state-of-the-art methods for detecting 65 novel seizures with higher specificity and sensitivity, and lower memory footprint.

• The paper demonstrates a novel one-shot seizure detection approach combining LBP feature extraction with hyperdimensional computing for robust iEEG analysis.
• The method leverages end-to-end binary operations to achieve rapid learning from minimal seizure examples, significantly reducing computational overhead.
• Experimental results on 16 patients show superior specificity and sensitivity, paving the way for efficient implantable seizure monitoring devices.

One-shot Learning for iEEG Seizure Detection Using End-to-end Binary Operations

Introduction

The paper proposes an innovative algorithmic approach for the detection of epileptic seizures using intracranial electroencephalography (iEEG) data, implemented through a novel method combining local binary patterns (LBP) with hyperdimensional (HD) computing. This work is situated within the context of addressing the need for efficient seizure detection algorithms, especially for patients with drug-resistant epilepsy undergoing pre-surgical long-term iEEG recordings. The challenges addressed include effective seizure detection from limited ictal examples, low computational complexity for potential implantable device deployment, and adaptability to patient-specific seizure dynamics.

Methodology

Local Binary Patterns (LBP) and Symbolization

The algorithm employs LBP as a feature extraction technique, transforming time series data into binary pattern codes. The LBP code captures relational dynamics between consecutive time points in iEEG signals, which, although simple, robustly preserve the dominant features of the signal dynamics, particularly for differentiating between interictal and ictal states.

Hyperdimensional (HD) Computing

HD computing is harnessed here due to its strengths in large-scale distributed representation and fast learning capabilities. Atomic vectors, representing basic inputs like electrode names and LBP codes, populate an item memory (IM). The method combines these vectors using bundling and binding operations. This encodes spatial and temporal iEEG dynamics into high-dimensional binary vectors, facilitating efficient learning and generalization from minimal examples—crucial for one-shot or few-shot learning setups.

Algorithmic Contributions

The paper delineates a processing pipeline that maintains end-to-end binary operations, minimizing computational overhead while maximizing detection efficiency. Key stages include:

1. Feature Extraction: LBP codes convert iEEG sequences into bit strings.
2. High-dimensional Encoding: Binary high-dimensional vectors encode comprehensive temporal and electrode-wise dynamics.
3. Binary Classification: A straightforward associative memory mechanism classifies new data points, utilizing simple distance calculations over binary vectors.
4. Postprocessing: Contextual patient-specific thresholds reduce false positives, optimizing real-world applicability.

Experimental Results

Utilizing data from 16 patients and various iEEG electrodes, the algorithm demonstrates its capacity for rapid learning and high generalization. The approach effectively requires only one to six seizures for training across the patient cohort, achieving superior specificity and sensitivity compared to prevailing methods such as LBP with SVM and LGP with a Multilayer Perceptron (MLP). The average detection delay is recorded at approximately 18.2 seconds, aligning favorably with clinical requirements for timely intervention.

Discussion and Implications

This algorithm represents a significant stride towards practical seizure detection technologies that are both computationally minimal and robust. The scalability of this method, alongside its universal parameter set, underscores its potential for broad deployment across diverse patient profiles with variable electrode configurations. From a future perspective, as implantable and portable monitoring devices evolve, the integration of such low-latency, lightweight algorithms will be pivotal for continuous, real-time monitoring and intervention in epilepsy management systems.

Conclusion

This research introduces a robust and efficient methodology for seizure detection through an elegant synergy of LBP and HD computing. By leveraging binary operations throughout the learning and inference phases, the paper furnishes a scalable, patient-agnostic solution that outperforms existing techniques in sensitivity and memory efficiency, embodying a promising avenue for future developments in biosignal processing for neurological applications.