A machine learning approach for automated wide-range frequency tagging analysis in embedded neuromonitoring systems.

EEG is a standard non-invasive technique used in neural disease diagnostics and neurosciences. Frequency-tagging is an increasingly popular experimental paradigm that efficiently tests brain function by measuring EEG responses to periodic stimulation. Recently, frequency-tagging paradigms have proven successful with low stimulation frequencies (0.5-6Hz), but the EEG signal is intrinsically noisy in this frequency range, requiring heavy signal processing and significant human intervention for response estimation. This limits the possibility to process the EEG on resource-constrained systems and to design smart EEG based devices for automated diagnostic. We propose an algorithm for artifact removal and automated detection of frequency tagging responses in a wide range of stimulation frequencies, which we test on a visual stimulation protocol. The algorithm is rooted on machine learning based pattern recognition techniques and it is tailored for a new generation parallel ultra low power processing platform (PULP), reaching performance of more that 90% accuracy in the frequency detection even for very low stimulation frequencies (<1Hz) with a power budget of 56mW.

Analysis of the evolution of COVID-19 disease understanding through temporal knowledge graphs.

The COVID-19 pandemic highlighted two critical barriers hindering rapid response to novel pathogens. These include inefficient use of existing biological knowledge about treatments, compounds, gene interactions, proteins, etc. to fight new diseases, and the lack of assimilation and analysis of the fast-growing knowledge about new diseases to quickly develop new treatments, vaccines, and compounds. Overcoming these critical challenges has the potential to revolutionize global preparedness for future pandemics. Accordingly, this article introduces a novel knowledge graph application that functions as both a repository of life science knowledge and an analytics platform capable of extracting time-sensitive insights to uncover evolving disease dynamics and, importantly, researchers' evolving understanding. Specifically, we demonstrate how to extract time-bounded key concepts, also leveraging existing ontologies, from evolving scholarly articles to create a single temporal connected source of truth specifically related to COVID-19. By doing so, current knowledge can be promptly accessed by both humans and machines, from which further understanding of disease outbreaks can be derived. We present key findings from the temporal analysis, applied to a subset of the resulting knowledge graph known as the temporal keywords knowledge graph, and delve into the detailed capabilities provided by this innovative approach.

Scalable and energy efficient seizure detection based on direct use of compressively-sensed EEG data on an ultra low power multi-core architecture.

Extracting information from dense multi-channel neural sensors for accurate diagnosis of brain disorders necessitates computationally expensive and advanced signal processing approaches to analyze the massive volume of recorded data. Compressive Sensing (CS) is an efficient method for reducing the computational complexity and power consumption in the resource-constrained multi-site neural systems. However, reconstructing the signal from compressed measurements is computationally intensive, making it unsuitable for real-time applications such as seizure detection. In this paper, a seizure detection algorithm is proposed to overcome these limitations by circumventing the reconstruction phase and directly processing the compressively sampled EEG signals. The Lomb-Scargle Periodogram (LSP) is used to extract the spectral energy features of the compressed data. Performance of the seizure detector using non-linear support vector machine (SVM) classifier, tested on 24 patients of the CHB-MIT data-set for compression ratios (CR) of 1-64x, is 96-93%, 92-87%, 0.95-0.91, and <1 s for sensitivity, accuracy, the area under the curve, and latency, respectively. A power-efficient classification method based on the utilization of dual linear SVM classifiers is proposed. The proposed classification method based on the dual linear SVM classification achieved better classification performance compared to commonly used classifiers, such as K-nearest neighbor, random forest, artificial neural network, and linear SVM, while consuming low power in comparison to non-linear SVM kernels. The hardware-optimized implementation of this algorithm is proposed on a low-power multi-core SoC for near-sensor data analytics: Mr. Wolf. Optimized implementation of this algorithm on Mr. Wolf platform leads to detecting a seizure with an energy budget of 18.4 µJ and 3.9 µJ for a compression ratio of 24x using non-linear SVM classifier and the dual linear SVM based classification method, respectively.

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

This paper presents a wearable electromyographic gesture recognition system based on the hyperdimensional computing paradigm, running on a programmable parallel ultra-low-power (PULP) platform. The processing chain includes efficient on-chip training, which leads to a fully embedded implementation with no need to perform any offline training on a personal computer. The proposed solution has been tested on 10 subjects in a typical gesture recognition scenario achieving 85% average accuracy on 11 gestures recognition, which is aligned with the state-of-the-art, with the unique capability of performing online learning. Furthermore, by virtue of the hardware friendly algorithm and of the efficient PULP system-on-chip (Mr. Wolf) used for prototyping and evaluation, the energy budget required to run the learning part with 11 gestures is 10.04 mJ, and 83.2  µJ per classification. The system works with a average power consumption of 10.4 mW in classification, ensuring around 29 h of autonomy with a 100 mAh battery. Finally, the scalability of the system is explored by increasing the number of channels (up to 256 electrodes), demonstrating the suitability of our approach as universal, energy-efficient biopotential wearable recognition framework.