The Results section covers the performance comparison of XGBoost with CNN and the deployment system performance evaluation. Finally, we conclude our work and discuss possible future directions in the Conclusions section.The study of brain activity using electroencephalogram typically involves extracting information from signals associated with certain activities. In recent years, machine learning techniques have been applied to the classification of mTBI because it enables the extraction of complex and typically nonlinear patterns from the EEG data. Most of the work surveyed used rule-based techniques, such as k-Nearest Neighbors . Previous investigations have studied a variety of classification techniques, including classical machine learning such as SVM and deep learning such as Convolutional Neural Networks . These techniques have been shown to perform TBI classification with more than 80% accuracy. However, in most investigations we reviewed that implement machine learning for TBI detection, the primary focus was the study of classification techniques and performance of classification models rather than portable deployment. A few systems used a small, portable computer for deployment in some form. The Neuroberry platform used a Raspberry Pi 2 device to capture EEG signals but the focus was on enabling EEG signal availability on the Internet of Things domain. The Acute Ischemic Stroke Identification System utilized an Analog to Digital Converter front end with Raspberry Pi 3 to capture physical EEG signals. However, plastic plants pots this system transferred the captured data to an HPC running MATLAB for signal analysis and processing and did not focus on signal classification.

Zgallai et al. described a Raspberry Pi-based system that used deep learning to perform EEG signal classification. It was designed to identify a subject’s intended movement direction from a multichannel EEG signal to control wheel-chair movement in a closed-loop robotic system rather than as a general system for identification, analysis, and monitoring of a physiological condition such as mTBI. Bruno et al. highlighted challenges with existing medical diagnosis techniques and described a classification system from the perspective of real-time TBI diagnosis, but their work was focused on the algorithm to perform TBI diagnosis and not on the implementation of a deployment system. In our previous work, we developed and described a CNN based model to perform automated sleep stage scoring and mTBI classification. In addition, we did a limited deployment of the CNN model on a Raspberry Pi 4 system. In that work, the focus was on describing the CNN model configuration, evaluating its performance, and showcasing that deployment to RPi was feasible rather than designing a complete, portable classification system. We have reused the previously developed CNN model in the current work to provide a baseline performance comparison with a new XGBoost model developed for this work. Further, the two models enable us to demonstrate the versatility of the current system to operate with multiple types of predictive models. To the best of our knowledge, no standalone, portable system has yet been created using Raspberry Pi that can capture real-time EEG signals, detect the presence of mTBI, and classify mTBI sleep/wake epoch states.A previously published dataset as described in [3] was used to train and evaluate deployed models. This dataset was collected as part of a study involving 11 adult male mice subjects divided into two groups—mTBI and Sham. FPI procedure was used to induce mTBI in 5 subjects and the remaining 6 mice were used as Sham subjects.

To capture the EEG signal, three ball-tipped electrodes were placed in the skull of each subject, two frontal and one in the parieto-occipital region. In this work, we proposed and demonstrated an RPi based EEG acquisition, processing, and classification system for early mTBI detection. This system was implemented using a single channel EEG data obtained from mice. This system was demonstrated to operate in a portable, real-time, and standalone configuration and perform classification of real-time EEG epochs into four target classes . As shown in Table 1, the accuracy, precision, and recall results were identical across RPi and HPC. This confirmed that the predictive model behavior did not change when the training and deployment systems involved different system architectures, i.e., x64 based MacOS/Windows HPC for training vs. ARM-based RPi for deployment and prediction. Hence, it is possible to train a predictive model on a more powerful computer and deploy it to an embedded device such as RPi that has limited memory and processing resources. This is especially applicable to multilayered neural networks like CNN that typically have long training times on an HPC, and the training times would be prohibitively long on an embedded device like RPi. We calculated the epoch processing time on RPi by varying the number of epochs, as shown in Figure 4 and described in Table 2. While it was expected that the processing time would increase as the number of processed epochs is increased, the key inference was that the processing time was considerably smaller than the time required to collect the EEG epochs. At 256 Hz sampling rate and 64 s epoch size, the processing time ranged from 0.01% to 0.03% of the epoch collection time. Hence, we concluded that the system had ample time to process previously captured EEG epochs while new epochs were captured at practical EEG signal sampling rates. We employed two different approaches for supervised learning models used in this system, the CNN model developed in our previous work, and an XGBoost predictive model created in the current work. We compared classification metrics and performance of the XGBoost and CNN models on the deployment system as well as an HPC.

We observed that the XGBoost model exhibited better performance in terms of accuracy and inference time compared to the CNN based predictive model. In the case of XGBoost, the variation of inference time remained roughly within 2 µs between HPC and RPi. A low inference time was critical for the real-time operation of the classification system. One possible reason for the better accuracy performance in the case of XGBoost compared to CNN was that the classification model for XGBoost was created using hand-crafted features which enabled learning differentiating patterns for the four target classes better than the CNN model that automatically extracted the differentiating features. These results, however, were data-dependent, so they should be validated on different datasets to verify the generality of the model. We found that overall, XGBoost was better suited for deployment on RPi because of its faster inference time and better performance than CNN. By using two different predictive models for classification, we demonstrated the flexibility of the system to deploy improved classification models in the future. In this system, we used a DAC to generate EEG signal waveform form European data format files. This provided a reliable way to generate an EEG signal waveform without requiring an actual subject to capture the EEG signal from. We verified that the EEG waveform generated using the DAC on RPi was consistent with the EEG data stored in the EDF file. The verification was done by calculating MSE across the stored and generated signal, which was found to be 0.26, a small value indicating that the generated signal represented the stored signal accurately. Synthesizing EEG signals to replicate the complex and typically nonlinear signal patterns is challenging and the ability to generate EEG signals from an actual recording data file using a DAC simplifies the setup that is required to test an EEG classification deployment system hardware and software chain. It enables the use of several available open-access EEG data files to train classification models and test the deployment system. For future use, the signal generation capability of this system can be simplified for ease of use and expanded to work with a variety of EEG data file types. This can help accelerate mTBI related future research pertaining to portable classification systems that are often constrained by the lack of readily available live EEG signals to test a hardware classification system. In addition to early mTBI detection, blueberry pot the capability of the system to perform live classification on input EEG signals can be extended to cover mTBI related health and sleep monitoring applications in the future. Typically, after the initial diagnosis, TBI patients undergo EEG sleep monitoring in a hospital setup. A portable EEG sleep monitoring system, such as the one described in this work, can enable a subject to self-monitor in home settings and greatly enhances the accuracy, efficiency, and efficacy. The classification system developed in the current work can also provide a replacement of the labor-intensive manual sleep-stage scoring of EEG signals by human experts with an online and automated system with the capability to perform fast sleep staging. Further, our technical approaches can be extended to several other EEG applications, including detection of the onset of epileptic seizures, strokes, and other neurological conditions.

In this work, we used a relatively simple hardware system to capture and digitize EEG signals, which could be improved. Because we generated EEG signals from a datafile containing clean EEG data, this hardware did not include amplification and filtering stages. A practical system designed for field use would require additional hardware and software capabilities to capture and process EEG signals in real-time. We also used a relatively simple metric for comparison of generated and stored EEG signals. While we only used MSE as a metric for this system, for cases where components in the signal path could potentially cause phase changes in the signal, MSE should be coupled another metric such as cross-correlation to verify signal integrity. In terms of hardware, such a system would require amplification, preprocessing, and filtering stages. In software, decimation, normalization, Independent Components Analysis , physiological artifact removal , and filtering stages can be implemented. Further, we used an 8-bit ADC for this proof-of-concept system, but for devices designed for practical use, ADCs typically vary from 16-bit to 24-bit resolution. For example, the OpenBCI Cyton Biosensing system for sampling EEG and other physiological signals uses a 24-bit ADC. We will note that higher resolution ADCs also involve a relatively higher cost and have lower sampling rates as the number of resolution bits increases. In addition, the system in this work was designed for single-channel EEG generation and capture, which limits its use for multichannel EEG applications. It does not directly provide connectivity to wireless EEG headsets. However, several “hardware attached on top” devices are available for RPi, for example, the brain HAT, that makes it possible to connect wireless headsets seamlessly and we anticipate the system in this work to function as intended with the actual streaming EEG data outside the particulars of EEG headset interfacing.Horticultural crops have high economic, and enrich our lives through their aesthetic and nutritional value. Many horticultural species originate from tropical regions and are sensitive to cold at every stage of their lifecycle. Cold stress leads to lower productivity and post-harvest losses in these species, with poor economic and environmental outcomes. Better understanding of the protective mechanisms mediated by hormonal and other signaling pathways may offer solutions to reduce cold-stress induced losses. The papers included in this collection illustrate this concept, examining natural cold-tolerance mechanisms and practical ways for growers to alleviate chilling stress and to reduce crop losses. The studies were remarkably diverse in terms of the species studied , plant organs examined , and approaches used . The papers encompassed the use of basic science, aimed at identifying key genes and their roles in cold signal transduction and protective pathways in fruit and photosynthetic tissues; reverse genetics for proof-of-concept on the hypothesized role of a cold-tolerance transcription factor cloned from an understudied species; and emerging technologies, by using exogenous hormones and signaling compounds to mitigate the harmful effects of chilling. These studies are described below.C-repeat binding factor proteins constitute a transcription factor subfamily known to play a key role in plants against different types of abiotic stress including cold, heat, salinity or dehydration, and thus have been extensively studied. Overexpression of CBFs has been used for the development of genetically modified plants with enhanced stress tolerance and for the investigation of the molecular mechanisms underlying plant stress responses. Using this approach, Yang et al. found that overexpression of three newly identified longan CBF genes enhanced cold tolerance in Arabidopsis by increasing the content of the osmoprotectant proline, reducing the accumulation of reactive oxygen species , and stimulating the expression of cold-responsive genes.