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 [1] for detecting 65 novel seizures with higher specificity and sensitivity, and lower memory footprint.

Laelaps: An Energy-Efficient Seizure Detection Algorithm from Long-term Human iEEG Recordings without False Alarms

“…2A LBP code of length is associated with every sampling point by concatenating its bit with the successive −1 bits. We observe that the distribution of LBP codes is significantly different between the ictal and the interictal states [12]. More specifically, during the interictal state the histogram of LBP codes is flattened out (i.e., the counts are almost evenly distributed over all the possible codes).…”

“…After filtering and downsampling the raw iEEG signals, anbit LBP code is computed for every sampling point. The LBP codes with different lengths ( ∈ [4,8]) produce almost similar performance [12], hence we set = 6 because using larger code sizes impairs the applicability to non-stationary iEEG signals. In addition, larger codes increase the delay of classification since the code size determines the minimum size of the statistical analysis window, i.e., the size of such a window should be large enough that all symbols can at least theoretically occur once [14].…”

“…However, one major challenge with iEEG signals is to realize seizure detection algorithms with low complexity suitable for execution on implantable devices with limited battery size for long-term or even lifelong operation. Another challengecommon to all methods-is to reliably detect seizures from a small number of examples due to the uneven distribution with much longer interictal than ictal time periods, and the patientspecific seizure dynamics [7], [12]. To address these challenges, one promising option is to develop computationally efficient and fast learning algorithms by combining methods from symbolic dynamics [5], [13], [14], [15], [1] and brain-inspired hyperdimensional computing [16], [17].…”

Online Learning and Classification of EMG-Based Gestures on a Parallel Ultra-Low Power Platform Using Hyperdimensional Computing

“…Based on the use of these MAP operations, an encoder can be designed for various tasks, e.g., EMG [20], [22], [49], EEG [23], [50], ECoG [51], ExG [45], or in general pattern processing [52]. The encoder emits a hypervector representing the event of interest that is then fed into an associative memory (AM) for training and inference.…”

Analysis of Contraction Effort Level in EMG-Based Gesture Recognition Using Hyperdimensional Computing

“…In this paper, we propose building a general classification model based on hyperdimensional (HD) computing [7] to deal with varying muscle contraction effort levels. HD computing has shown promising results in classification tasks using biosignals such as EMG in recognizing hand gestures [8] and electrocorticography (ECoG) for seizure detection with oneshot learning [9]. With slight modifications to our previously introduced encoding scheme [8], we analyze the muscle contraction level variations in two different ways, depending on the application: If discrimination between different gestures is the only goal, the classifier should output the same gesture class regardless of the subject's effort level.…”