Approximately 1 in every 150 children is diagnosed with epilepsy during the first ten years of life [1]. These children experience seizures, which disrupt their lives and directly harm the developing brain. Electroencephalography (EEG) is the main tool used clinically to diagnose seizures and epilepsy. However, the interpretation of EEGs requires time-consuming expert analysis [2]. Automated detection systems are a powerful tool that can help address the issue by reducing expert annotation time. Research on the automatic detection of seizures in pediatric EEG has been limited. Most seizure detection methods have been developed and tested using larger numbers of adult EEG [3], [4]. However, research has shown that brain events in EEG change with ageing [5], [6]. Therefore, model trained on EEGs from adults may not be be suitable for children. To test this hypothesis, we trained a seizure detection model on adult EEG and tested on adult and pediatric EEG recordings.
In this talk, we will discuss the use of electroencephalograms (EEG) in Intensive Care Units (ICU). We will review the use of EEGs as a multi-dimensional biomarker. We will review applications of artificial intelligence (AI) and machine learning (ML) for each type of biomarker. We will review cases highlighting biomarker usage in clinical management. Continuous EEG (CEEG) is an invaluable tool in the ICU since it yields multi-multi-dimensional biomarkers. AI can overcome or ameliorate limitations of CEEG applications in the ICU. Real-time analysis and interpretation of CEEG data is essential to influence clinical decision-making and clinical outcomes. ML models and AI integration into the decision-making process provides standardization and automation. Opportunities exist for the integration of real-time annotation and AI-based decision-support to achieve better patient outcomes.
Cardiovascular disease (CVD) has been a pressing medical issue in the United States for over a century and has been a leading cause of death [1], [2] with a great impact on mortality, morbidity, and healthcare cost. The Centers for Disease Control and Prevention (CDC) [3] reported that CVD is responsible for one death every 34 seconds and approximately 697,000 deaths in 2020 alone. Additionally, between 2017 to 2018, CVD directly and indirectly cost the United States economy approximately 378 billion dollars [2].
Multivariate signals are signals consisting of multiple signals measured simultaneously over time and are most commonly acquired by sensor networks. The emerging field of graph signal processing (GSP) promises to analyse dynamic characteristics of multivariate signals, while at the same time taking the network, or spatial structure between the signals into account. To do so, GSP decomposes the multivariate signals into graph frequency signals, which are ordered by their magnitude. However, the meaning of the graph frequencies in terms of this ordering remains poorly understood. Here, we investigate the role the ordering plays in preserving valuable dynamic structures in the signals, with neuroimaging applications in mind. In order to overcome the limitations in sample size common to neurophysiological data sets, we introduce a minimalist simulation framework to generate arbitrary amounts of data. Using this artificial data, we find that lower graph frequency signals are less suitable for classifying neurophysiological data than higher graph frequency signals. We further introduce a baseline testing framework for GSP. Using this framework, we conclude that dynamic, or spectral structures are poorly preserved in GSP, high-lighting current limitations of GSP for neuroimaging.
Accommodation is the eye's ability to focus up close by changing the shape of the lens. Accommodation affects the development of myopia and glaucoma and its age-related decline results in presbyopia. Presbyopia affects 100% of the population in the fifth decade of life. An understanding of accommodation is required to develop the best treatments for these maladies, but how the lens changes shape is still in dispute after more than 165 years. The fundamental issue is whether the change in lens shape results from all zonules (circumferential suspensory ligaments that connect the lens of the eye to the ciliary body) relaxing, which causes central and peripheral lens surface steepening, or whether instead just the anterior and posterior zonules relax while the equatorial zonules are under increased tension, which causes the lens surface to peripherally flatten and centrally steepen. The alternatives are illustrated in Figure 1.
Persons with Parkinson's Disease (PD) frequently have speech and voice disorders. Regular speech therapy with a speech-language pathologist is essential to mitigate progressive symptom deterioration. Speech-related therapies, such choral singing groups are alternative approaches designed to be more naturalistic and enhance participant enjoyment. It is important to measure and quantify the effects of these therapies on the vocal features of PD patients to determine efficacy. We performed a prospective crossover study of 25 PD patients attending discussion or choral-singing groups for 12 weeks each (Parkinsonics NCT02753621). Every six weeks, each participant produced several recordings of the sustained vowels /a:/ and /e:/ at ‘normal’ and ‘maximum’ loudness. The goal was to identify if there are signal-processing-based features that can help track changes in the voice of PD patients over time. Voice features were extracted from these recordings using the Automatic Voice Condition Analysis (AVCA) library and were compared using non-parametric statistical tests. Results suggest that neither therapy caused any significant improvements in the analyzed phonatory aspects of the patients' voices. Future work should require use of connected speech to analyze articulation and comparison with a control group of participants with PD not attending any therapy to evaluate if therapy can mitigate the progressive effects of PD on the voice of patients.
Clinicians currently use handwriting as one of the methods to establish the presence and monitor the progression of neurodegenerative diseases (NDs). While common handwriting evaluation methods are valuable means to detect fine motor and cognitive impairments associated with NDs, these are observer-dependent and subjective. In the present study, we analyzed a broad array of interpretable features, some proposed for the first time in this study, obtained from online handwriting data of participants with NDs and control subjects (CTRL). ND participants have Alzheimer's disease (AD), Parkinson's disease (PD), or Parkinson's disease mimics (PDM). Hand-writing data from three different neuropsychological tasks was used: Copy Text task, Copy Cube task, and Copy Image task. Then, we arranged three complementary sets of features and conducted a statistical analysis to test their significance between groups. Overall results suggested that subjects with AD reported a significantly higher $(p < 0.05)$ amount of data points and total duration with respect to the CTRL group in almost all the tasks under assessment. On the other hand, subjects with PD showed a significantly lower $(p < 0.05)$ horizontal width (both on tablet and in the air). Even though the AD and PDM groups showed a significantly lower velocity and acceleration $(p < 0.05)$ , their number of inversions in velocity and acceleration was significantly greater $(p < 0.05)$ , which indicates disfluency in writing. The features that we have used were found to provide good results in differentiating the studied groups and could be considered as part of diagnostic tools for the assessment and monitoring of NDs in clinical trials.
It is well understood that the CREB protein is highly involved in neuronal mechanisms underlying memory and learning in mammalian brain, and deficiencies in CREB activity can result in transition to certain pathological conditions. In this paper, we use some published experimental data, along with a neuronal system composed of the Izhikevich neuron model, to characterize how CREB abnormalities can alter neuronal signals and the system behavior. The abnormal data are extracted from intracellular recordings collected from the neurons of transgenic mice expressing VP16-CREB - a constitutively active form of CREB - whereas the normal data are obtained from the wild-type mice neurons. Upon estimating the neuron model parameters from the experimental data, we observe that the model exhibits good fit to both normal and abnormal data, for various synaptic input currents. To study the effect of CREB abnormalities on the considered neuronal system, we use the information theoretic redundancy parameter. It basically measures - for the system output neuron - the amount of spike count information overlap that exists between the states of the stimulus currents injected to the input neurons. Our analysis reveals a noticeable increase in the information redundancy, when CREB behaves abnormally. This finding motivates further exploration of the biological implications of the information redundancy in neuronal systems, and its use as a parameter to model abnormalities in CREB and perhaps other important transcription factors involved in learning and memory.
Dynamical, causal and cross-frequency coupling analysis using the EEG has received significant interest for the analysis and diagnosis of neurological disorders [1]–[3]. Due to the high computational requirements needed for some of these methods, EEG channel selection is crucial [4]. Functional connectivity (FC) between EEG channels is often used for channel selection and connectivity analysis [4, S, 6]. Ideally, in the case of selecting channels for dynamical and causal analysis, FC methods should be able to account for linear and nonlinear spatial and temporal interactions between EEG channels. In neuroscience, FC is quantified using different measures of (dis) similarity to assess the statistical dependence between two signals [5]. However, the interpretation of FC measures can differ significantly from one measure to another[5, 7]. In the early diagnosis of AD, [7] showed correlations among various (dis)similarity measures, and therefore these measures can be grouped. Thus, one from each is sufficient to extract information from the data [7]. Therefore, the development of a generic measure of (dis)similarity is important in FC analysis.
The sinoatrial node (SAN), located in the right atrium wall, is the heart's biological pacemaker and determines heart rate due to the repetitive spontaneous action potentials for cardiac rhythmic contractions in the heart pacemaker cells. The funny current (If) and SAN, together with regulation by the sympathetic and parasympathetic nervous systems, modulate the frequency of the SAN action potential. The interaction of these systems is responsible for the rhythmic pacemaker activity, controlling heart rate, and abnormalities resulting in arrhythmias.
Cystometry measures the behavior of the bladder and is frequently used to evaluate lower urinary tract abnormalities. Cystometry is conducted using two catheters, one in the bladder and the other in the vagina or rectum, which increases discomfort and complexity of the test. In this work we evaluated a method to estimate detrusor pressure (PDET), the pressure generated by a bladder contraction, from only a single catheter measuring vesical pressure (PVES). Using twenty urodynamic studies, we used statistical inference and wavelet multiresolution analysis to maximize the correlation coefficient $(R)$ between estimated PDET and calculated PDET after detecting and eliminating artifacts. Moreover, the estimator design considered a prospective real-time implementation. Root main square (RMS) error and correlation coefficient were used to evaluate algorithm accuracy in estimating $\mathbf{P}_{\text{DET}}$ , while a statistical F -score evaluated the accuracy of artifact detection. The output of the proposed estimator compared with calculated PDET, and overall estimation performance showed that $RMS=10.7\pm 2.1\ \text{cmH}_{2}\mathrm{O}$ and $R=0.88\pm$ 0.6 $(\mathrm{N}=20)$ . Moreover, detection accuracy for cough and Valsalva events were 99.5% and 84.30/0, respectively. We conclude that estimating PDET from PVES only is feasible making single channel cystometry a possibility.
Epilepsy is one of the most common neurological disorders in the world [1], affecting about 50 million people worldwide [2]. Epileptic seizures occur when millions of neurons are synchronously excited, resulting in a wave of electrical activity in the cerebral cortex [3]. Electroencephalography (EEG) is a noninvasive tool that measures cortical activity with millisecond temporal resolution. EEGs record the electrical potentials generated by the cerebral cortex nerve cells [4]. Therefore, this tool is commonly used for the analysis and detection of seizures [5]. Epilepsy causes many difficulties in relation to the quality of life of the patient. It is therefore vital that automatic detection algorithms exist to aid neurologists to accurately classify the different types of seizures. Roy et al. [10] used different machine learning techniques to achieve an average F1-score of 0.561 using 2 s windows whilst Vanabelle et al. [11] used 1 s windows and achieved an accuracy of 51.33%, which shows that reducing the time window would also decrease the accuracy of classification. This paper aims to show that an NLP can be used for hierarchical classification, following upon an earlier work on combining simple partial and complex partial seizures [9]. The second aim is to show a pipeline that can be used to separate the seizures back into their original labels using neural networks. This method is quick, effective, and requires less training.
Continuous intraoperative monitoring with electroencephalography (EEG) is routinely used in carotid endarterectomy (CEA) to detect cerebral ischemia [1]. Visually observed changes in EEG related to ischemia include an ipsilateral decrease in amplitude of faster frequencies or an ipsilateral increase in activity of slower frequencies. In the literature, significant EEG change has been defined as a decrease in the amplitude in the alpha frequency band by 50% or greater or an increase in activity in the theta or delta frequency band by 50% or greater [2], [3]. Compared to raw EEG, quantitative EEG (QEEG) parameters can enhance visual EEG review. QEEG parameters are derived by applying Fourier transformation to raw EEG signals to generate power spectra [4]. Examples of QEEG parameters include delta, theta, alpha, beta, gamma power values, alpha power to delta power ratio (ADR), beta power to delta power ratio (BDR), alpha-plus-beta power to delta-plus-theta power ratio (ABDTR), spectral edge frequency 90% (SEF90) and amplitude-integrated EEG (aEEG). QEEG parameters have been previously investigated in detecting ischemia in a relatively small number of patients undergoing CEA [5]. In this study, we report on the analyses of QEEG parameters in a large population of patients who underwent CEA with EEG monitoring.
Muscle fatigue is defined as a decline in the ability to maintain a desired force against a load. Muscle fatigue may also be described as a decline in the muscle's maximum force during contraction. In contrast to muscle damage or weakness, characterized by a compromise in the ability of well-rested muscles to generate force, muscle fatigue is generally reversible with rest [2]. In a muscle experiencing fatigue, the nerves cannot sustain the high frequency signal necessary to reach the Maximum Contraction (MC) for a long time, resulting in a decline in muscle force during a sustained contraction. Due to its utility in providing information about nerve signaling and muscle's electrical activity, surface electromyography (sEMG) is currently the dominant method to detect muscle fatigue [2]. Mechanomyography (MMG) can reveal unique information that cannot be derived from the sEMG signal alone about the physiological behavior of muscles during contraction. However, more information may be needed about the ability of MMG to measure changes in muscle's activation patterns and mechanical properties that occur with muscle fatigue. Additionally, investigating the force-dependent characteristics of the MMG signal can provide information about physiological properties such as muscle activation strategies and fiber type distribution, which can be used to explore factors contributing to fatigue responses [1]. The purpose of this study is to examine and analyze the electrical and mechanical muscle responses to submaximal isometric contractions, as well as force-varying trapezoidal contractions in the rectus femoris muscle.
In the last decade, improvements in automated speech processing, powered by signal processing and machine learning, has led to new approaches for medical assessment. Additionally, previous research in clinical speech has identified interpretable measures that are sensitive to changes in the cognitive, linguistic, affective, and motoric domains. In order to include speech-based automatic approaches in clinical applications, factors such as robustness, specificity, and interpretability of speech features are crucial. We focused on the analysis of a multi-modal array of interpretable features obtained from the spoken responses of participants with Neurodegenerative Diseases (ND) and control participants (CN) to neuropsychological tests. ND participants have Alzheimer's disease (AD), Parkinson's disease (PD), or Parkinson's disease mimics (PDM). We first collected spoken responses to three tests, a modified version of the Stroop test (MST), a verb naming task (VNT), and a noun naming task (NNT). Then, we arranged two complementary sets of cognitive and acoustic features and analyzed their statistical significance between the groups studied. Our results suggested that AD participants had significantly greater reaction times and significantly lower response accuracy with respect to the other groups across tests. In addition, PDM participants, compared to CN and PD participants, took a significantly longer time to complete the MST and NNT, while all the groups of participants with NDs showed significantly lower confidence during the MST. Since the analyzed features provided good differentiation results, they can be used in diagnostic tools for the assessment of NDs.
Circulating tumor cells (CTCs) shed from the primary tumor, intravasate into blood, and translocate to distant tissues via circulation [1]. CTC enumeration allows cancer detection, treatment monitoring, and survival prediction [2], [3]. In the clinical setting immunofluorescence-based CTC enumeration is primarily used by expert cytopathologists. Manual enumeration requires cytopathologists with rich experience to read hundreds to thousands of images in hours. Despite the seemingly high number, this poor efficiency hinders the relevant clinical implementation. Therefore, high-automation enumeration is missing but highly desired [4]. Here, we proposed an automatic CTC segmentation and enumeration method in digital pathology by using deep learning techniques. To prepare for enumeration, peripheral blood mononuclear cells (PBMC) were extracted from cancer patient blood followed by infection with a reengineered adenovirus, i.e., rAd CTC , which is a CD46-targeting, DF3 promoter-selective, and GFP-overexpression adenovirus. The rAd CTC ensures detection specificity and efficiency of expression for CTCs. Subsequently, PBMCs were stained with anti-CD45 fluorescence-labeled antibody and DNA staining dye DAPI. Finally, the acquired fluorescence images were used for automatic segmentation and enumeration [5]. It is noteworthy that the fluorescence images used in this study contain three channels. The green, red, and blue signals respectively represent overexpressed GFP in infected cells, CD45 staining on leukocyte membranes, and nuclear staining.
Heart diseases are a leading cause of mortality globally with cardiovascular disease (CVD) accounting for around 17.9 million deaths as of 2019 [1]. Limited access to healthcare services in low- and middle-income countries may be a main reason of the high mortality. The financial burden associated with CVD is also high and may reach $70 billion in USA by 2030 [2]. Early detection of heart diseases can reduce adverse health events and lower related costs. Regular monitoring of these conditions can alert patients and healthcare providers about life-threatening abnormalities, which can reduce hospitalization rates. Use of a reliable, simple-to-use and cost-effective non-invasive techniques to detect heart conditions can expedite the diagnosis and treatment processes, improving patient management and quality of life.
In recent years, because of safety and timeliness of the ultrasound imaging, this technology has been widely used in the field of medical diagnosis [1]. In the process of ultrasound imaging, the beamforming process is the most important part, which directly determines the imaging quality [2]. At present, the most widely used algorithm is the traditional delay-and-sum (DAS), but DAS has some inherent disadvantages in low resolution and obviously artifacts [3]. For the purpose of solving these deficiencies, many advanced imaging methods have been proposed. Among them, the minimum variance (MV) designed by Capon is a kind of very potential algorithm due to its high resolution [4]. However, the effect of MV algorithm is mainly depended on the accuracy of the preset desired directional vector and the calculation of covariance matrix. Therefore, the MV has the problem of insufficient robustness [5]. In subsequent studies, many innovative methods had been used to overcome the shortcomings of MV algorithm [6], such as eigenspace-based MV (ESBMV).
Hip fractures are common in the geriatric population and represent a growing social and economic burden [1]. They are associated with decreased mobility, and the recovery period can be prolonged. Early mobilization is a critical component of the recovery process. Few studies have quantitatively measured activity levels in patients after hip fracture surgery, resulting in a lack of objective data about mobility status after hospital discharge. Furthermore, each year about 1.5 million elderly people are injured falling, and about 47,300 people aged ≥65 years suffer injuries from falls using walking aids that require an emergency room visit [2]. Patients using walkers are seven times more likely to fall than those that use canes. The risk of repeat falls has been shown to be especially high in patients who have already sustained a hip fracture.
An EEG seizure detection algorithm employed in a clinical setting is likely to encounter many EEG segments that are difficult to classify due to the complexity of EEG signals and small data sets frequently used to train seizure detectors. The detectors should therefore be able to notify the clinician when they are uncertain in their predictions and they should also be accurate for confident predictions. This would enable the clinician to focus mainly on the parts of the recording where confidence in predictions is low. Here we analyse the calibration of neonatal and adult seizure detection algorithms based on a convolutional neural network in terms of how well the output seizure/non-seizure probabilities estimate the corresponding empirical frequencies. We found that the detectors turned out to be overconfident, in particular when incorrectly predicting seizure segments as non-seizure segments. The calibration of both detectors, measured in terms of expected calibration error and overconfidence error, was improved noticeably with the use of Monte Carlo dropout. We find that a straightforward application of dropout during training and classification leads to a noticeable improvement in the calibration of EEG seizure detectors based on a convolutional neural network.
