Non-linear SVM Seizure Classification SoC for Low Latency Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing seizure detection systems face challenges in achieving high detection accuracy with low latency, often resulting in either high latency or low detection rates, making them unsuitable for real-time seizure suppression in epileptic patients.
Innovation Solution
A system-on-chip (SoC) with a hardware-efficient log-linear engine implementing non-linear support vector machine (NLSVM) for seizure detection, incorporating an eight-channel scalable electroencephalography (EEG) architecture, time division band-pass filters, and low-noise analog front-end circuits to achieve high detection accuracy and low false alarm rates within 2 seconds.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of time
If a Linear Support Vector Machine (LSVM) classifier is used, then latency is reduced (less than 2 s), but detection rate decreases (84.4%) and false alarm rate increases (max. 14.7%)
Solution Approach 1:
The patent transitions from a linear classifier (LSVM) to a non-linear classifier (NLSVM with RBF kernel), fundamentally changing the mathematical model parameters. This transformation enables the system to capture complex non-linear patterns in EEG data, achieving 95.1% detection rate while maintaining low latency through hardware optimization
Solution Approach 2:
The patent replaces software-based classification with a hardware-implemented NLSVM engine. By mapping the non-linear support vector machine algorithm to dedicated hardware circuits, the system achieves real-time processing with less than 2 seconds latency while maintaining 95.1% detection accuracy, resolving the contradiction between speed and accuracy
2Measurement precision
If detection accuracy is improved, then false alarm rate increases, making the system unsuitable for practical use
Solution Approach 1:
The patent employs an NLSVM with RBF kernel function, changing from linear to non-linear parameter space transformation. This enables the system to achieve 95.1% detection accuracy while maintaining only 0.94% false alarm rate by properly tuning the kernel parameters gamma and sigma to optimize the decision boundary
Solution Approach 2:
The system implements a feedback mechanism where the NLSVM classifier continuously learns from EEG data patterns, adjusting its decision boundary to distinguish true seizure events from false alarms. The hardware implementation provides real-time feedback processing, maintaining high accuracy while suppressing false alarms through adaptive thresholding
Data Source
AI summary
This disclosure is directed to a machine-based patient-specific seizure classification system. In general, an example system may comprise a non-linear SVM seizure classification system-on-chip (SoC) with multichannel EEG data acquisition and storage for epileptic patients is presented. The SoC may integrate a hardware-efficient log-linear Gaussian Basis Function engine, floating point piecewise linear natural log, and low-noise, high dynamic range readout circuits. In at least one example implementation, the SoC may consume 1.83 μJ/classification while classifying 8 channel results with an average detection rate, average false alarm rate and latency of 95.1%, 0.94% and <2 s, respectively.