MT-NAM: An Efficient and Adaptive Model for Epileptic Seizure Detection

TL;DR

This paper addresses real-time, hardware-efficient epileptic seizure detection from scalp EEG by combining Neural Additive Models (NAM) with a distillation-based MT-NAM variant. It shows that MT-NAM, using micro regression trees to replace NAM feature networks, delivers up to ~50× faster inference with minimal loss in sensitivity, and that a test-time update rule (T3A) can restore NAM-level sensitivity during deployment. The approach is evaluated on the CHB-MIT dataset, achieving high window-based sensitivity and specificity while guaranteeing 100% event-based sensitivity in several configurations. The work demonstrates the practicality of combining interpretable feature-wise modeling, knowledge distillation, and online adaptation for real-time seizure detection and hardware implementation.

Abstract

Enhancing the accuracy and efficiency of machine learning algorithms employed in neural interface systems is crucial for advancing next-generation intelligent therapeutic devices. However, current systems often utilize basic machine learning models that do not fully exploit the natural structure of brain signals. Additionally, existing learning models used for neural signal processing often demonstrate low speed and efficiency during inference. To address these challenges, this study introduces Micro Tree-based NAM (MT-NAM), a distilled model based on the recently proposed Neural Additive Models (NAM). The MT-NAM achieves a remarkable 100$\times$ improvement in inference speed compared to standard NAM, without compromising accuracy. We evaluate our approach on the CHB-MIT scalp EEG dataset, which includes recordings from 24 patients with varying numbers of sessions and seizures. NAM achieves an 85.3\% window-based sensitivity and 95\% specificity. Interestingly, our proposed MT-NAM shows only a 2\% reduction in sensitivity compared to the original NAM. To regain this sensitivity, we utilize a test-time template adjuster (T3A) as an update mechanism, enabling our model to achieve higher sensitivity during test time by accommodating transient shifts in neural signals. With this online update approach, MT-NAM achieves the same sensitivity as the standard NAM while achieving approximately 50$\times$ acceleration in inference speed.

Figures (8)

• Figure 1: The proposed model framework. (a)Feature extraction. During this phase 7 band power features aligned with Linelenght and Variance is extracted from each channel of raw EEG signal for every 1 second window size resulting in 9 features per channel and totally $M=9N$ number of features with $N$ being the channel number. (b)Training NAM. During training phase NAM is trained using the extracted features and offline prediction is applied for seizure detection. (c)MT-NAM Inference During inference the distilled version of NAM namely, Micro Tree-based NAM (MT-NAM) and online T3A update is used for predicting the seizures. Color Codes: Indicates the feature extraction paradigm (a), is the training paradigm (b) and is the inference paradigm during testing (c).
• Figure 2: (a)-(b) Window-based sensitivity and specificity across all subjects for the NAM and baseline models. (c) ROC curves (averaged across all subjects) for the NAM and baseline models. The zoomed-in area highlights the performance of NAM and LGBM at low false positive rates.
• Figure 3: Comparison of the outputs of (a) MT1-NAM, (b) MT2-NAM, (c) MT4-NAM, and (d) NAM. The left plots depict the outputs of three randomly selected feature functions (transparent lines) and their summation (opaque lines). On the right, an example of the micro decision tree used for approximation is displayed, with the color bar representing the proportion of samples passing through each node of the micro tree.
• Figure 4: Illustration of a feature function learned by NAM ($f(x)$) and its approximation ($\tilde{f}(x)$) using (a)MT-1, (b)MT-2, and (c)MT-4 approximations of NAM. The white regions in the plots correspond to regions with low data density (typically a few points) and the gray regions correspond to regions with high data density.
• Figure 5: Per-patient analysis of window-based specificity and sensitivity across different online update methods. The results indicate that, among all updating approaches, T3A demonstrates the greatest improvement in seizure sensitivity. The red color indicates the drop and green color indicates the increase in the metric over the offline performance after the online update is applied.