Δ BP-Net: Monitoring "Changes" in Blood Pressure Using PPG with Self-contrastive Masking.

Estimating blood pressure (BP) values from physiological signals (e.g., photoplethysmogram (PPG)) using deep learning models has recently received increased attention, yet challenges remain in terms of models' generalizability. Here, we propose taking a new approach by framing the problem as tracking the "changes" in BP over an interval, rather than directly estimating its value. Indeed, continuous monitoring of acute changes in BP holds promising implications for clinical applications (e.g., hypertensive emergencies). As a solution, we first present a self-contrastive masking (SCM) model, designed to perform pair-wise temporal comparisons within the input signal. We then leverage the proposed SCM model to introduce ΔBPNet, a model trained to detect elevations/drops greater than a given threshold in the systolic blood pressure (SBP) over an interval, from PPG. Using data from PulseDB, 1) we evaluate the performance of ΔBP-Net on previously unseen subjects, 2) we test ΔBP-Net's ability to generalize across domains by training and testing on different datasets, and 3) we compare the performance of ΔBP-Net with existing PPG-based BP-estimation models in detecting over-threshold SBP changes. Formulating the problem as a binary classification task (i.e., over-threshold SBP elevation/ drop or not), ΔBP-Net achieves 75.97%/73.19% accuracy on data from subjects unseen during training. Additionally, the proposed ΔBP-Net outperforms ΔSBP estimations derived from existing PPG-based BP-estimation methods. Overall, by shifting the focus from estimating the value of SBP to detecting overthreshold "changes" in SBP, this work introduces a new potential for using PPG in clinical BP monitoring, and takes a step forward in addressing the challenges related to the generalizability of PPG-based BP-estimation models.

Imaging the large-scale and cellular response to focal traumatic brain injury in mouse neocortex.

Traumatic brain injury (TBI) affects neural function at the local injury site and also at distant, connected brain areas. However, the real-time neural dynamics in response to injury and subsequent effects on sensory processing and behavior are not fully resolved, especially across a range of spatial scales. We used in vivo calcium imaging in awake, head-restrained male and female mice to measure large-scale and cellular resolution neuronal activation, respectively, in response to a mild TBI induced by focal controlled cortical impact (CCI) injury of the motor cortex (M1). Widefield imaging revealed an immediate CCI-induced activation at the injury site, followed by a massive slow wave of calcium signal activation that traveled across the majority of the dorsal cortex within approximately 30 s. Correspondingly, two-photon calcium imaging in primary somatosensory cortex (S1) found strong activation of neuropil and neuronal populations during the CCI-induced traveling wave. A depression of calcium signals followed the wave, during which we observed atypical activity of a sparse population of S1 neurons. Longitudinal imaging in the hours and days after CCI revealed increases in the area of whisker-evoked sensory maps at early time points, in parallel to decreases in cortical functional connectivity and behavioral measures. Neural and behavioral changes mostly recovered over hours to days in our mild-TBI model, with a more lasting decrease in the number of active S1 neurons. Our results provide novel spatial and temporal views of neural adaptations that occur at cortical sites remote to a focal brain injury.

A Siamese Convolutional Neural Network for Identifying Mild Traumatic Brain Injury and Predicting Recovery.

Timely diagnosis of mild traumatic brain injury (mTBI) remains challenging due to the rapid recovery of acute symptoms and the absence of evidence of injury in static neuroimaging scans. Furthermore, while longitudinal tracking of mTBI is essential in understanding how the diseases progresses/regresses over time for enhancing personalized patient care, a standardized approach for this purpose is not yet available. Recent functional neuroimaging studies have provided evidence of brain function alterations following mTBI, suggesting mTBI-detection models can be built based on these changes. Most of these models, however, rely on manual feature engineering, but the optimal set of features for detecting mTBI may be unknown. Data-driven approaches, on the other hand, may uncover hidden relationships in an automated manner, making them suitable for the problem of mTBI detection. This paper presents a data-driven framework based on Siamese Convolutional Neural Network (SCNN) to detect mTBI and to monitor the recovery state from mTBI over time. The proposed framework is tested on the cortical images of Thy1-GCaMP6s mice, obtained via widefield calcium imaging, acquired in a longitudinal study. Results show that the proposed model achieves a classification accuracy of 96.5%. To track the state of the injured brain over time, a reference distance map is constructed, which together with the SCNN model, are employed to assess the recovery state in subsequent sessions after injury, revealing that the recovery progress varies among subjects. The promising results of this work suggest that a similar approach could be potentially applicable for monitoring recovery from mTBI, in humans.

An LSTM-based adversarial variational autoencoder framework for self-supervised neural decoding of behavioral choices.

Objective.This paper presents data-driven solutions to address two challenges in the problem of linking neural data and behavior: (1) unsupervised analysis of behavioral data and automatic label generation from behavioral observations, and (2) extraction of subject-invariant features for the development of generalized neural decoding models.Approach. For behavioral analysis and label generation, an unsupervised method, which employs an autoencoder to transform behavioral data into a cluster-friendly feature space is presented. The model iteratively refines the assigned clusters with soft clustering assignment loss, and gradually improves the learned feature representations. To address subject variability in decoding neural activity, adversarial learning in combination with a long short-term memory-based adversarial variational autoencoder (LSTM-AVAE) model is employed. By using an adversary network to constrain the latent representations, the model captures shared information among subjects' neural activity, making it proper for cross-subject transfer learning.Main results. The proposed approach is evaluated using cortical recordings of Thy1-GCaMP6s transgenic mice obtained via widefield calcium imaging during a motivational licking behavioral experiment. The results show that the proposed model achieves an accuracy of 89.7% in cross-subject neural decoding, outperforming other well-known autoencoder-based feature learning models. These findings suggest that incorporating an adversary network eliminates subject dependency in representations, leading to improved cross-subject transfer learning performance, while also demonstrating the effectiveness of LSTM-based models in capturing the temporal dependencies within neural data.Significance. Results demonstrate the feasibility of the proposed framework in unsupervised clustering and label generation of behavioral data, as well as achieving high accuracy in cross-subject neural decoding, indicating its potentials for relating neural activity to behavior.

The Feeling of Time Passing Is Associated with Recurrent Sustained Activity and Theta Rhythms Across the Cortex.

Introduction: We are constantly estimating how much time has passed, and yet know little about the brain mechanisms through which this process occurs. In this pilot study, we evaluated so-called subjective time estimation with the temporal bisection task, while recording brain activity from electroencephalography (EEG). Methods: Nine adult participants were trained to distinguish between two durations of visual stimuli as either "short" (400 msec) or "long" (1600 msec). They were then presented with stimulus durations in between the long and short stimuli. EEG data from 128 electrodes were examined with a novel analytical method that identifies segments of sustained cortical activity during the task. Results: Participants tended to categorize intermediate durations as "long" more frequently than "short" and were thus experiencing time as moving faster while overestimating the amount of time passing. Their mean bisection point (during which frequency of selecting short vs. long is equal) was closer to the geometric mean of task stimuli (800 msec) rather than the arithmetic mean (1000 msec). In contrast, sustained brain activity occurred closer to the arithmetic mean. The recurrence rate of this activity was highly related to the bisection point, especially when analyzed within naturally occurring theta oscillations (4-8 Hz) (r = -0.90). Discussion: Sustained activity across the cortex within the theta range may reflect temporal durations, whereas its repeated appearance relates to the subjective feeling of time passing.

Demographic Information Fusion Using Attentive Pooling In CNN-GRU Model For Systolic Blood Pressure Estimation.

Fusing demographic information into deep learning models has become of interest in recent end-to-end cuff-less blood pressure (BP) estimation studies in order to achieve improved performance. Conventionally, the demographic feature vector is concatenated with the pooled embedding vector. Here, using an attention-based convolutional neural network-gated recurrent unit (CNN-GRU), we present a new approach and fuse the demographic information into the attentive pooling module. Our results demonstrate that, under calibration-based testing protocol, the proposed approach provides improved systolic blood pressure (SBP) estimation accuracy (with R2=0.86 and mean absolute error (MAE)=4.90 mmHg) compared to both the baseline model with no demographic information fused, and the conventional approach of fusing demographic information. Our work showcases the feasibility of using attention-based methods to combine demographic features with deep learning models, and suggests new ways for fusing demographic information in deep learning models to achieve improved BP estimation accuracy.

Hierarchical Classification Strategy for Mitigating the Impact of The Presence of Pain in fNIRS-based BCIs.

Brain computer interfaces (BCIs) can find applications in assistive systems for patients who experience conditions that impede their motor abilities. A BCI uses signals acquired from the brain to control external devices. As physical pain influences cortical signals, the presence of pain can negatively impact the performance of the BCI. In this work, we propose a strategy to mitigate this negative impact. Cortical signals are acquired from test subjects while they performed two mental arithmetic tasks, in the presence and the absence of painful stimuli. The task of the BCI is to reliably classify the two mental arithmetic tasks from the cortical recordings, irrespective of the presence or the absence of pain. We propose to do this classification, hierarchically, in two levels. In the first level, the data is classified into those captured in the presence and the absence of pain. Depending on the results of the classification from the first level, in the second level, the BCI performs the classification of tasks using a classifier trained either in the presence or the absence of pain. A 1-dimensional convolutional neural network (1D-CNN) is used for classification at both levels. It is observed that using this hierarchical strategy, the BCI is able to classify the tasks with an accuracy greater than 90%, irrespective of the presence or the absence of pain. Given that the presence of physical pain has shown previously to reduce the classification accuracy of a BCI to almost chance levels, this mitigation strategy will be a significant step towards enhancing the performance of BCIs when they are used in assistive systems for patients.

A Synthetic Ultra-Wideband Transceiver for Millimeter-Wave Imaging Applications.

In this work, we present a transceiver front-end in SiGe BiCMOS technology that can provide an ultra-wide bandwidth of 100 GHz at millimeter-wave frequencies. The front-end utilizes an innovative arrangement to efficiently distribute broadband-generated pulses and coherently combine received pulses with minimal loss. This leads to the realization of a fully integrated ultra-high-resolution imaging chip for biomedical applications. We realized an ultra-wide imaging band-width of 100 GHz via the integration of two adjacent disjointed frequency sub-bands of 10-50 GHz and 50-110 GHz. The transceiver front-end is capable of both transmit (TX) and receive (RX) operations. This is a crucial component for a system that can be expanded by repeating a single unit cell in both the horizontal and vertical directions. The imaging elements were designed and fabricated in Global Foundry 130-nm SiGe 8XP process technology.

A Deep Learning Framework Based on Dynamic Channel Selection for Early Classification of Left and Right Hand Motor Imagery Tasks.

Ideal brain-computer interfaces (BCIs) need to be efficient and accurate, demanding for classifiers that can work across subjects while providing high classification accu-racy results from recordings with short duration. To address this problem, we present a new deep learning framework for discriminating motor imagery (MI) tasks from electroen-cephalography (EEG) signals. The framework consists of a 1D convolutional neural network-long short-term memory (CNN-LSTM), combined with a dynamic channel selection approach based on Davies-Bouldin index (DBI). Using data from BCI competition IV-IIa data, the proposed framework reports an average classification accuracy of 70.17% and 76.18% when using only 800 ms and 1500 ms of the EEG data after the task onset, respectively. The proposed framework dynamically balances individual differences, achieves comparable or better performance compared to existing work, while using short duration of EEG.

Identifying mild traumatic brain injury using measures of frequency-specified networks.

Objective. Early diagnosis of mild traumatic brain injury (mTBI) is challenging, yet crucial for providing patients with timely treatments and minimizing the risks of developing injury-related disorders. To tackle this problem, this paper presents a framework based on measures of frequency-specified brain functional networks identifying mTBI.Approach. Cortical activity of 15 control and 15 injury Thy1-GCaMP6s mice are recorded, using widefield calcium imaging, prior to and 20 minutes after inducing injury. Power spectral distribution (PSD) of the recorded cortical activities is examined, and the frequency bands with significant difference in PSD between the injury and control groups are identified. Frequency-specified functional networks are then constructed. Employing graph theoretical analysis, various network measures from the constructed frequency-specified functional networks are extracted and used as features. Several classifiers are utilized to evaluate the performance of the computed network measures, either individually or collectively as features, to classify mTBI from control.Main results. Spectral analysis reveals the presence of two dominant frequency bands (low:<1Hz) and high: [1-8] Hz) in the cortical activities recorded via calcium imaging. Comparison of the brain networks of control and injury groups shows significant reduction (p < 0.05) in global functional connectivity following injury, specially for the high frequency band network. Interestingly, graph measures of the high frequency band network provided higher classification accuracy results, compared to those computed from the low frequency band network, suggesting that mTBI network-based features are frequency dependent. Using all network measures collectively as a multi-measure feature vector and a convolutional neural networks classifier, a model for identifying mTBI is developed, offering an average classification accuracy of 97.28%.Significance. Results signifies the importance of considering frequency-specific analysis in functional networks for mTBI identification, and demonstrate the possibility of using network measures for early mTBI diagnosis.

The Effects of Filtering PPG Signal on Pulse Arrival Time-Systolic Blood Pressure Correlation.

Pulse arrival time (PAT), evaluated from electro-cardiogram (ECG) and photoplethysmogram (PPG) signals, has been widely used for cuff-less blood pressure (BP) estimation due to its high correlation with BP. However, the question of whether filtering the PPG signal impacts the extracted PAT values and consequently, the correlation between PAT and BP, has not been investigated before. In this paper, using data from 18 subjects, changes in the PAT values, and in the subject-specific PAT-systolic BP (SBP) correlation caused by filtering the PPG signal with variable cutoff frequencies in the range of 2 to 15 Hz are studied. For PAT extraction, three PPG characteristic points (foot, maximum slope and systolic peak) are considered. Results show that differences in the cutoff frequency can shift the PAT values and introduce a worst-case error of over 8.2 mmHg for SBP estimation, indicating that PPG signal filter settings can impact PAT-based BP estimations. Our study suggests that extracting the PAT from the maximum slope point of PPG signal filtered at 10 Hz provides the most stable correlation with SBP.

PulseDB: A large, cleaned dataset based on MIMIC-III and VitalDB for benchmarking cuff-less blood pressure estimation methods.

There has been a growing interest in developing cuff-less blood pressure (BP) estimation methods to enable continuous BP monitoring from electrocardiogram (ECG) and/or photoplethysmogram (PPG) signals. The majority of these methods have been evaluated using publicly-available datasets, however, there exist significant discrepancies across studies with respect to the size, the number of subjects, and the applied pre-processing steps for the data that is eventually used for training and testing the models. Such differences make conducting performance comparison across models largely unfair, and mask the generalization capability of various BP estimation methods. To fill this important gap, this paper presents "PulseDB," the largest cleaned dataset to date, for benchmarking BP estimation models that also fulfills the requirements of standardized testing protocols. PulseDB contains 1) 5,245,454 high-quality 10 -s segments of ECG, PPG, and arterial BP (ABP) waveforms from 5,361 subjects retrieved from the MIMIC-III waveform database matched subset and the VitalDB database; 2) subjects' identification and demographic information, that can be utilized as additional input features to improve the performance of BP estimation models, or to evaluate the generalizability of the models to data from unseen subjects; and 3) positions of the characteristic points of the ECG/PPG signals, making PulseDB directly usable for training deep learning models with minimal data pre-processing. Additionally, using this dataset, we conduct the first study to provide insights about the performance gap between calibration-based and calibration-free testing approaches for evaluating generalizability of the BP estimation models. We expect PulseDB, as a user-friendly, large, comprehensive and multi-functional dataset, to be used as a reliable source for the evaluation of cuff-less BP estimation methods.

Cuff-Less Blood Pressure Estimation From Photoplethysmography via Visibility Graph and Transfer Learning.

This paper presents a new solution that enables the use of transfer learning for cuff-less blood pressure (BP) monitoring via short duration of photoplethysmogram (PPG). The proposed method estimates BP with low computational budget by 1) creating images from segments of PPG via visibility graph (VG), hence, preserving the temporal information of the PPG waveform, 2) using pre-trained deep convolutional neural network (CNN) to extract feature vectors from VG images, and 3) solving for the weights and bias between the feature vectors and the reference BPs with ridge regression. Using the University of California Irvine (UCI) database consisting of 348 records, the proposed method achieves a best error performance of 0.00±8.46 mmHg for systolic blood pressure (SBP), and -0.04±5.36 mmHg for diastolic blood pressure (DBP), respectively, in terms of the mean error (ME) and the standard deviation (SD) of error, ranking grade B for SBP and grade A for DBP under the British Hypertension Society (BHS) protocol. Our novel data-driven method offers a computationally-efficient end-to-end solution for rapid and user-friendly cuff-less PPG-based BP estimation.

Cuff-Less Blood Pressure Estimation via Small Convolutional Neural Networks.

Deep learning-based cuff-less blood pressure (BP) estimation methods have recently gained increased attention as they can provide accurate BP estimation with only one physiological signal as input. In this paper, we present a simple and effective method for cuff-less BP estimation by training a small-scale convolutional neural network (CNN), modified from LeNet-5, with images created from short segments of the photoplethysmogram (PPG) signal via visibility graph (VG). Results show that the trained modified LeNet-5 model achieves an error performance of 0.184±7.457 mmHg for the systolic BP (SBP), and 0.343±4.065 mmHg for the diastolic BP (DBP) in terms of the mean error (ME) and the standard deviation (SD) of error between the estimated and reference BP. Both the SBP and the DBP accuracy rank grade A under the British Hypertension Society (BHS) protocol, demonstrating that our proposed method is an accurate way for cuff-less BP estimation.

PulseLab: An Integrated and Expandable Toolbox for Pulse Wave Velocity-based Blood Pressure Estimation.

In this paper, we introduce PulseLab, a comprehensive MATLAB toolbox that enables estimating the blood pressure (BP) from electrocardiogram (ECG) and photoplethysmogram (PPG) signals using pulse wave velocity (PWV)-based models. This universal framework consists of 6 sequential modules, covering end-to-end procedures that are needed for estimating BP from raw PPG/ECG data. These modules are "dataset formation", "signal pre-processing", "segmentation", "characteristic-points detection", "pulse transit time (PTT)/ pulse arrival time (PAT) calculation", and "model validation". The toolbox is expandable and its application programming interface (API) is built such that newly-derived PWV-BP models can be easily included. The toolbox also includes a user-friendly graphical user interface (GUI) offering visualization for step-by-step processing of physiological signals, position of characteristic points, PAT/PTT values, and the BP regression results. To the best of our knowledge, PulseLab is the first comprehensive toolbox that enables users to optimize their model by considering several factors along the process for obtaining the most accurate model for cuff-less BP estimation.

A Variational Encoder Framework for Decoding Behavior Choices from Neural Data.

In this paper, using an adversarial variational encoder model, we propose a two-step data-driven approach to extract cross-subject feature representations from neural activity in order to decode subjects' behavior choices. First, various characteristics of the recorded behavior are computed and passed as features to a clustering model in order to categorize different behavior choices in each trial and create labels for the data. Then, we utilize a variational encoder to learn the latent space mappings from neural activity. An attached adversary network is used in a discriminative setting to detach the subject's individuality from the representations. Recorded cortical activity from Thy1-GCaMP6s transgenic mice during a motivational licking experiment was used in this study. Experimental results demonstrate the capabilities of the proposed method in extracting discriminative representations from neural data to decode behavior by achieving an average classification accuracy of 88.8% across subjects.

A Data-Driven Framework for Intention Prediction via Eye Movement With Applications to Assistive Systems.

Fast and accurate human intention prediction can significantly advance the performance of assistive devices for patients with limited motor or communication abilities. Among available modalities, eye movement can be valuable for inferring the user's intention, as it can be tracked non-invasively. However, existing limited studies in this domain do not provide the level of accuracy required for the reliable operation of assistive systems. By taking a data-driven approach, this paper presents a new framework that utilizes the spatial and temporal patterns of eye movement along with deep learning to predict the user's intention. In the proposed framework, the spatial patterns of gaze are identified by clustering the gaze points based on their density over displayed images in order to find the regions of interest (ROIs). The temporal patterns of gaze are identified via hidden Markov models (HMMs) to find the transition sequence between ROIs. Transfer learning is utilized to identify the objects of interest in the displayed images. Finally, models are developed to predict the user's intention after completing the task as well as at early stages of the task. The proposed framework is evaluated in an experiment involving predicting intended daily-life activities. Results indicate that an average classification accuracy of 97.42% is achieved, which is considerably higher than existing gaze-based intention prediction studies.

Early classification of motor tasks using dynamic functional connectivity graphs from EEG.

Objective. Classification of electroencephalography (EEG) signals with high accuracy using short recording intervals has been a challenging problem in developing brain computer interfaces (BCIs). This paper presents a novel feature extraction method for EEG recordings to tackle this problem.Approach. The proposed approach is based on the concept that the brain functions in a dynamic manner, and utilizes dynamic functional connectivity graphs. The EEG data is first segmented into intervals during which functional networks sustain their connectivity. Functional connectivity networks for each identified segment are then localized, and graphs are constructed, which will be used as features. To take advantage of the dynamic nature of the generated graphs, a long short term memory classifier is employed for classification.Main results. Features extracted from various durations of post-stimulus EEG data associated with motor execution and imagery tasks are used to test the performance of the classifier. Results show an average accuracy of 85.32% about only 500 ms after stimulus presentation.Significance. Our results demonstrate, for the first time, that using the proposed feature extraction method, it is possible to classify motor tasks from EEG recordings using a short interval of the data in the order of hundreds of milliseconds (e.g. 500 ms). This duration is considerably shorter than what has been reported before. These results will have significant implications for improving the effectiveness and the speed of BCIs, particularly for those used in assistive technologies.

Classification of EEG Data Based on the Spatio-Temporo-Rhythmic Characteristics of the Task-Discriminating Functional Sub-networks.

We propose a new approach that utilizes the dynamic state of cortical functional connectivity for the classification of task-based electroencephalographic (EEG) data. We introduce a novel feature extraction framework that locates functional networks in the cortex as they convene at different time intervals across different frequency bands. The framework starts by applying the wavelet transform to isolate, then augment, EEG frequency bands. Next, the time intervals of stationary functional states, within the augmented data, are identified using the source-informed segmentation algorithm. Functional networks are localized in the brain, during each segment, using a singular value decomposition-based approach. For feature selection, we propose a discriminative-associative algorithm, and use it to find the sub-networks showing the highest recurrence rate differences across the target tasks. The sequences of augmented functional networks are projected onto the identified sub-networks, for the final sequences of features. A dynamic recurrent neural network classifier is then used for classification. The proposed approach is applied to experimental EEG data to classify motor execution and motor imagery tasks. Our results show that an accuracy of 90% can be achieved within the first 500 msec of the cued task-planning phase.

Detecting mTBI by Learning Spatio-temporal Characteristics of Widefield Calcium Imaging Data Using Deep Learning.

Early diagnosis of mild traumatic brain injury (mTBI) is of great interest to the neuroscience and medical communities. Widefield optical imaging of neuronal populations over the cerebral cortex in animals provides a unique opportunity to study injury-induced alternations in brain function. Using this technique, along with deep learning, the goal of this paper is to develop a framework for the detection of mTBI. Cortical activities in transgenic calcium reporter mice expressing GCaMP6s are obtained using widefield imaging from 8 mice before and after inducing an injury. Two deep learning models for differentiating mTBI from normal conditions are proposed. One model is based on the convolutional neural network-long short term memory (CNN-LSTM), and the second model is based on a 3D-convolutional neural network (3D-CNN). These two models offer the ability to capture features both in the spatial and temporal domains. Results for the average classification accuracy for the CNN-LSTM and the 3D-CNN are 97.24% and 91.34%, respectively. These results are notably higher than the case of using the classical machine learning methods, demonstrating the importance of utilizing both the spatial and temporal information for early detection of mTBI.