Adelson Chua

EECS Department, University of California, Berkeley

Technical Report No. UCB/EECS-2025-22

May 1, 2025

http://www2.eecs.berkeley.edu/Pubs/TechRpts/2025/EECS-2025-22.pdf

Implantable devices that record neural activity and detect seizures have been adopted for issuing warnings or triggering neurostimulation to suppress epileptic seizures. Traditional seizure detection systems rely on high-accuracy offline-trained machine learning classifiers that need manual retraining when seizure patterns change over time. For an implantable seizure detection system, a low-power, at-the-edge, online learning algorithm can be used to dynamically adapt to neural signal drifts, maintaining high accuracy without external intervention. This dissertation describes an energy-efficient classification algorithm based on logistic regression that incorporates stochastic gradient descent for online and unsupervised updates, ensuring sustained high classification accuracies over long periods of time. The online learning framework was implemented on two different on-chip variants: one being focused solely on seizure detection (which is referred to as SOUL); and another that combines both seizure detection and prediction (which is referred to as SPIRIT), which leverages the detector’s outputs to continually improve the prediction accuracy without additional external inputs. The systems’ performance was evaluated using long-term datasets, including cases with drifting seizure features, demonstrating high prediction and detection accuracies over extended periods through on-chip adaptation. Both SOUL and SPIRIT managed to achieve comparable, if not better, detection and prediction accuracies versus other on-chip state-of-the-art work in this field. The online learning approach described in this thesis enabled the systems to maintain high accuracies over long periods of time while being very energy efficient. For SOUL, the combination of the proposed algorithmic approach and circuit-level optimizations resulted in an energy efficiency of 1.5 nJ/classification, which is at least 24x better than the state-of-the-art. It also consumes 0.1 mm2 of area making it the smallest seizure detector classifier in the literature by a factor of 10x. For SPIRIT, using the same architectural optimizations that made an energy-efficient SOUL, the energy efficiency for prediction was 17.2 nJ/classification, which is at least 5.6x better than the only other on-chip seizure predictor in the literature. Compared to the same work, SPIRIT’s power consumption is about 134x smaller at 17.2 µW, while also being 28x smaller at 0.14 mm2. SPIRIT is the first on-chip seizure predictor that can retrain in an unsupervised manner while being more energy efficient than state-of-the-art.

Advisors: Rikky Muller

---

BibTeX citation:

@phdthesis{Chua:EECS-2025-22,
    Author= {Chua, Adelson},
    Title= {Unsupervised Online Learning for Seizure Detection and Prediction},
    School= {EECS Department, University of California, Berkeley},
    Year= {2025},
    Month= {May},
    Url= {http://www2.eecs.berkeley.edu/Pubs/TechRpts/2025/EECS-2025-22.html},
    Number= {UCB/EECS-2025-22},
    Abstract= {Implantable devices that record neural activity and detect seizures have been adopted for issuing warnings or triggering neurostimulation to suppress epileptic seizures. Traditional seizure detection systems rely on high-accuracy offline-trained machine learning classifiers that need manual retraining when seizure patterns change over time. For an implantable seizure detection system, a low-power, at-the-edge, online learning algorithm can be used to dynamically adapt to neural signal drifts, maintaining high accuracy without external intervention. This dissertation describes an energy-efficient classification algorithm based on logistic regression that incorporates stochastic gradient descent for online and unsupervised updates, ensuring sustained high classification accuracies over long periods of time. The online learning framework was implemented on two different on-chip variants: one being focused solely on seizure detection (which is referred to as SOUL); and another that combines both seizure detection and prediction (which is referred to as SPIRIT), which leverages the detector’s outputs to continually improve the prediction accuracy without additional external inputs. The systems’ performance was evaluated using long-term datasets, including cases with drifting seizure features, demonstrating high prediction and detection accuracies over extended periods through on-chip adaptation.
Both SOUL and SPIRIT managed to achieve comparable, if not better, detection and prediction accuracies versus other on-chip state-of-the-art work in this field. The online learning approach described in this thesis enabled the systems to maintain high accuracies over long periods of time while being very energy efficient. For SOUL, the combination of the proposed algorithmic approach and circuit-level optimizations resulted in an energy efficiency of 1.5 nJ/classification, which is at least 24x better than the state-of-the-art. It also consumes 0.1 mm2 of area making it the smallest seizure detector classifier in the literature by a factor of 10x. For SPIRIT, using the same architectural optimizations that made an energy-efficient SOUL, the energy efficiency for prediction was 17.2 nJ/classification, which is at least 5.6x better than the only other on-chip seizure predictor in the literature. Compared to the same work, SPIRIT’s power consumption is about 134x smaller at 17.2 µW, while also being 28x smaller at 0.14 mm2. SPIRIT is the first on-chip seizure predictor that can retrain in an unsupervised manner while being more energy efficient than state-of-the-art.},
}

EndNote citation:

%0 Thesis
%A Chua, Adelson 
%T Unsupervised Online Learning for Seizure Detection and Prediction
%I EECS Department, University of California, Berkeley
%D 2025
%8 May 1
%@ UCB/EECS-2025-22
%U http://www2.eecs.berkeley.edu/Pubs/TechRpts/2025/EECS-2025-22.html
%F Chua:EECS-2025-22