Autoregressive models for biomedical signal processing.

Autoregressive models are ubiquitous tools for the analysis of time series in many domains such as computational neuroscience and biomedical engineering. In these domains, data is, for example, collected from measurements of brain activity. Crucially, this data is subject to measurement errors as well as uncertainties in the underlying system model. As a result, standard signal processing using autoregressive model estimators may be biased. We present a framework for autoregressive modelling that incorporates these uncertainties explicitly via an overparameterised loss function. To optimise this loss, we derive an algorithm that alternates between state and parameter estimation. Our work shows that the procedure is able to successfully denoise time series and successfully reconstruct system parameters.Clinical relevance- This new paradigm can be used in a multitude of applications in neuroscience such as brain-computer interface data analysis and better understanding of brain dynamics in diseases such as epilepsy.

Artificial intelligence-enhanced epileptic seizure detection by wearables.

OBJECTIVE: Wrist- or ankle-worn devices are less intrusive than the widely used electroencephalographic (EEG) systems for monitoring epileptic seizures. Using custom-developed deep-learning seizure detection models, we demonstrate the detection of a broad range of seizure types by wearable signals. METHODS: Patients admitted to the epilepsy monitoring unit were enrolled and asked to wear wearable sensors on either wrists or ankles. We collected patients' electrodermal activity, accelerometry (ACC), and photoplethysmography, from which blood volume pulse (BVP) is derived. Board-certified epileptologists determined seizure onset, offset, and types using video and EEG recordings per the International League Against Epilepsy 2017 classification. We applied three neural network models-a convolutional neural network (CNN) and a CNN-long short-term memory (LSTM)-based generalized detection model and an autoencoder-based personalized detection model-to the raw time-series sensor data to detect seizures and utilized performance measures, including sensitivity, false positive rate (the number of false alarms divided by the total number of nonseizure segments), number of false alarms per day, and detection delay. We applied a 10-fold patientwise cross-validation scheme to the multisignal biosensor data and evaluated model performance on 28 seizure types. RESULTS: We analyzed 166 patients (47.6% female, median age = 10.0 years) and 900 seizures (13 254 h of sensor data) for 28 seizure types. With a CNN-LSTM-based seizure detection model, ACC, BVP, and their fusion performed better than chance; ACC and BVP data fusion reached the best detection performance of 83.9% sensitivity and 35.3% false positive rate. Nineteen of 28 seizure types could be detected by at least one data modality with area under receiver operating characteristic curve > .8 performance. SIGNIFICANCE: Results from this in-hospital study contribute to a paradigm shift in epilepsy care that entails noninvasive seizure detection, provides time-sensitive and accurate data on additional clinical seizure types, and proposes a novel combination of an out-of-the-box monitoring algorithm with an individualized person-oriented seizure detection approach.

Obfuscated Memory Malware Detection in Resource-Constrained IoT Devices for Smart City Applications.

Obfuscated Memory Malware (OMM) presents significant threats to interconnected systems, including smart city applications, for its ability to evade detection through concealment tactics. Existing OMM detection methods primarily focus on binary detection. Their multiclass versions consider a few families only and, thereby, fail to detect much existing and emerging malware. Moreover, their large memory size makes them unsuitable to be executed in resource-constrained embedded/IoT devices. To address this problem, in this paper, we propose a multiclass but lightweight malware detection method capable of identifying recent malware and is suitable to execute in embedded devices. For this, the method considers a hybrid model by combining the feature-learning capabilities of convolutional neural networks with the temporal modeling advantage of bidirectional long short-term memory. The proposed architecture exhibits compact size and fast processing speed, making it suitable for deployment in IoT devices that constitute the major components of smart city systems. Extensive experiments with the recent CIC-Malmem-2022 OMM dataset demonstrate that our method outperforms other machine learning-based models proposed in the literature in both detecting OMM and identifying specific attack types. Our proposed method thus offers a robust yet compact model executable in IoT devices for defending against obfuscated malware.

A Multi-Modal AI-Driven Cohort Selection Tool to Predict Suboptimal Non-Responders to Aflibercept Loading-Phase for Neovascular Age-Related Macular Degeneration: PRECISE Study Report 1.

Patients diagnosed with exudative neovascular age-related macular degeneration are commonly treated with anti-vascular endothelial growth factor (anti-VEGF) agents. However, response to treatment is heterogeneous, without a clinical explanation. Predicting suboptimal response at baseline will enable more efficient clinical trial designs for novel, future interventions and facilitate individualised therapies. In this multicentre study, we trained a multi-modal artificial intelligence (AI) system to identify suboptimal responders to the loading-phase of the anti-VEGF agent aflibercept from baseline characteristics. We collected clinical features and optical coherence tomography scans from 1720 eyes of 1612 patients between 2019 and 2021. We evaluated our AI system as a patient selection method by emulating hypothetical clinical trials of different sizes based on our test set. Our method detected up to 57.6% more suboptimal responders than random selection, and up to 24.2% more than any alternative selection criteria tested. Applying this method to the entry process of candidates into randomised controlled trials may contribute to the success of such trials and further inform personalised care.

Weak Monotonicity With Trend Analysis for Unsupervised Feature Evaluation.

Performance in an engineering system tends to degrade over time due to a variety of wearing or ageing processes. In supervisory controlled processes there are typically many signals being monitored that may help to characterize performance degradation. It is preferred to select the least amount of information to obtain high quality of predictive analysis from a large amount of collected data, in which labeling the data is not always feasible. To this end a novel unsupervised feature selection method, robust with respect to significant measurement disturbances, is proposed using the notion of "weak monotonicity" (WM). The robustness of this notion makes it very attractive to identify the common trend in the presence of measurement noises and population variation from the collected data. Based on WM, a novel suitability indicator is proposed to evaluate the performance of each feature. This new indicator is then used to select the key features that contribute to the WM of a family of processes when noises and variations among processes exist. In order to evaluate the performance of the proposed framework of the WM and suitability, a comparative study with other nine state-of-the-arts unsupervised feature evaluation and selection methods is carried out on well-known benchmark datasets. The results show a promising performance of the proposed framework on unsupervised feature evaluation in the presence of measurement noises and population variations.

Evaluating Rehabilitation Progress Using Motion Features Identified by Machine Learning.

Evaluating progress throughout a patient's rehabilitation episode is critical for determining the effectiveness of the selected treatments and is an essential ingredient in personalised and evidence-based rehabilitation practice. The evaluation process is complex due to the inherently large human variations in motor recovery and the limitations of commonly used clinical measurement tools. Information recorded during a robot-assisted rehabilitation process can provide an effective means to continuously quantitatively assess movement performance and rehabilitation progress. However, selecting appropriate motion features for rehabilitation evaluation has always been challenging. This paper exploits unsupervised feature learning techniques to reduce the complexity of building the evaluation model of patients' progress. A new feature learning technique is developed to select the most significant features from a large amount of kinematic features measured from robotics, providing clinically useful information to health practitioners with reduction of modeling complexity. A novel indicator that uses monotonicity and trendability is proposed to evaluate kinematic features. The data used to develop the feature selection technique consist of kinematic data from robot-aided rehabilitation for a population of stroke patients. The selected kinematic features allow for human variations across a population of patients as well as over the sequence of rehabilitation sessions. The study is based on data records pertaining to 41 stroke patients using three different robot assisted exercises for upper limb rehabilitation. Consistent with the literature, the results indicate that features based on movement smoothness are the best measures among 17 kinematic features suitable to evaluate rehabilitation progress.

Effective Assessments of a Short-Duration Poor Posture on Upper Limb Muscle Fatigue Before Physical Exercise.

A forward head and rounded shoulder posture is a poor posture that is widely seen in everyday life. It is known that sitting in such a poor posture with long hours will bring health issues such as muscle pain. However, it is not known whether sitting in this poor posture for a short period of time will affect human activities. This paper investigates the effects of a short-duration poor posture before some typical physical activities such as push-ups. The experiments are set up as follows. Fourteen male subjects are asked to do push-ups until fatigue with two surface electromyography (sEMG) at the upper limb. Two days later, they are asked to sit in this poor posture for 15 min with eight sEMG sensors located at given back muscles. Then they do the push-ups after the short-duration poor posture. The observations from the median frequency of sEMG signals at the upper limb indicate that the short-duration poor posture does affect the fatigue procedure of push-ups. A significant decreasing trend of the performance of push-ups is obtained after sitting in this poor posture. Such effects indicate that some parts of the back muscles indeed get fatigued with only 15 min sitting in this poor posture. By further investigating the time-frequency components of sEMG of back muscles, it is observed that the low and middle frequencies of sEMG signals from the infraspinatus muscle of the dominant side are demonstrated to be more prone to fatigue with the poor posture. Although this study focuses only on push-ups, similar experiments can be arranged for other physical exercises as well. This study provides new insights into the effect of a short-duration poor posture before physical activities. These insights can be used to guide athletes to pay attention to postures before physical activities to improve performance and reduce the risk of injury.

Feature Learning in Assistive Rehabilitation Robotic Systems.

Feature learning plays a crucial role in data analysis when the measured data is in a high dimensional space. This paper applied the feature learning technique in data set collected from human movement experiments in an upper limb rehabilitation robotic device. The results showed that the proposed feature learning technique can identify key features to characterize the upper limb movements of humans, even though human variations exist. Four representative features were obtained out of 72 statistic features with very good prediction performance. In future, this feature learning technique can be used in building the link between the movement quality measured from robotic device to the well-known clinic measurements.

EMU: A transparent 3D robotic manipulandum for upper-limb rehabilitation.

This paper introduces the EMU, a three-dimensional robotic manipulandum for rehabilitation of the upper extremity for patients with neurological injury. The device has been designed to be highly transparent, have a large workspace, and allow the use of the hand for interaction with real-world objects to provide additional contextual cues during exercises. The transparency is achieved through the use of a capstan transmission for the drive joints; a hybrid serial parallel kinematics minimising moving inertia; and lightweight materials. An experimental protocol is reported here which demonstrates the transparency through a comparison to out-of-robot movements, and with an existing rehabilitation robotic device. Additionally, an adjustable gravity compensation method is constructed, which minimises the torque required at the shoulder to carry the subject's arm. These characteristics allow the EMU to serve as a multi-purpose platform for the further development of novel robot assisted rehabilitation strategies.

Probing to Observe Neural Dynamics Investigated with Networked Kuramoto Oscillators.

The expansion of frontiers in neural engineering is dependent on the ability to track, detect and predict dynamics in neural tissue. Recent innovations to elucidate information from electrical recordings of brain dynamics, such as epileptic seizure prediction, have involved switching to an active probing paradigm using electrically evoked recordings rather than traditional passive measurements. This paper positions the advantage of probing in terms of information extraction, by using a coupled oscillator Kuramoto model to represent brain dynamics. While active probing performs better at observing underlying system synchrony in Kuramoto networks, especially in non-Gaussian measurement environments, the benefits diminish with increasing relative size of electrode spatial resolution compared to synchrony area. This suggests probing will be useful for improved characterization of synchrony for suitably dense electrode recordings.

Fourier Tract Sampling (FouTS): A framework for improved inference of white matter tracts from diffusion MRI by explicitly modelling tract volume.

Diffusion MRI tractography algorithm development is increasingly moving towards global techniques to incorporate "downstream" information and conditional probabilities between neighbouring tracts. Such approaches also enable white matter to be represented more tangibly than the abstract lines generated by the most common approaches to fibre tracking. However, previously proposed algorithms still use fibre-like models of white matter corresponding to thin strands of white matter tracts rather than the tracts themselves, and therefore require many components for accurate representations, which leads to poorly constrained inverse problems. We propose a novel tract-based model of white matter, the 'Fourier tract', which is able to represent rich tract shapes with a relatively low number of parameters, and explicitly decouples the spatial extent of the modelled tract from its 'Apparent Connection Strength (ACS)'. The Fourier tract model is placed within a novel Bayesian framework, which relates the tract parameters directly to the observed signal, enabling a wide range of acquisition schemes to be used. The posterior distribution of the Bayesian framework is characterised via Markov-chain Monte-Carlo sampling to infer probable values of the ACS and spatial extent of the imaged white matter tracts, providing measures that can be directly applied to many research and clinical studies. The robustness of the proposed tractography algorithm is demonstrated on simulated basic tract configurations, such as curving, twisting, crossing and kissing tracts, and sections of more complex numerical phantoms. As an illustration of the approach in vivo, fibre tracking is performed on a central section of the brain in three subjects from 60 direction HARDI datasets.

Solving the Bloch equation with periodic excitation using harmonic balancing: application to Rabi modulated excitation.

In waveform design for magnetic resonance applications, periodic continuous-wave excitation offers potential advantages that remain largely unexplored because of a lack of understanding of the Bloch equation with periodic continuous-wave excitations. Using harmonic balancing techniques the steady state solutions of the Bloch equation with periodic excitation can be effectively solved. Moreover, the convergence speed of the proposed series approximation is such that a few terms in the series expansion suffice to obtain a very accurate description of the steady state solution. The accuracy of the proposed analytic approximate series solution is verified using both a simulation study as well as experimental data derived from a spherical phantom with doped water under continuous-wave excitation. Typically a five term series suffices to achieve a relative error of less than one percent, allowing for a very effective and efficient analytical design process. The opportunities for Rabi frequency modulated continuous-wave form excitation are then explored, based on a comparison with steady state free precession pulse sequences.

Effect of sensory experience on motor learning strategy.

It is well known that the central nervous system automatically reduces a mismatch in the visuomotor coordination. Can the underlying learning strategy be modified by environmental factors or a subject's learning experiences? To elucidate this matter, two groups of subjects learned to execute reaching arm movements in environments with task-irrelevant visual cues. However, one group had previous experience of learning these movements using task-relevant visual cues. The results demonstrate that the two groups used different learning strategies for the same visual environment and that the learning strategy was influenced by prior learning experience.

The dynamics of the epileptic brain reveal long-memory processes.

The pattern of epileptic seizures is often considered unpredictable and the interval between events without correlation. A number of studies have examined the possibility that seizure activity respects a power-law relationship, both in terms of event magnitude and inter-event intervals. Such relationships are found in a variety of natural and man-made systems, such as earthquakes or Internet traffic, and describe the relationship between the magnitude of an event and the number of events. We postulated that human inter-seizure intervals would follow a power-law relationship, and furthermore that evidence for the existence of a long-memory process could be established in this relationship. We performed a post hoc analysis, studying eight patients who had long-term (up to 2 years) ambulatory intracranial EEG data recorded as part of the assessment of a novel seizure prediction device. We demonstrated that a power-law relationship could be established in these patients (ß = - 1.5). In five out of the six subjects whose data were sufficiently stationary for analysis, we found evidence of long memory between epileptic events. This memory spans time scales from 30 min to 40 days. The estimated Hurst exponents range from 0.51 to 0.77 ± 0.01. This finding may provide evidence of phase-transitions underlying the dynamics of epilepsy.

Modelling extracellular electrical stimulation: part 3. Derivation and interpretation of neural tissue equations.

OBJECTIVE: A common approach in modelling extracellular electrical stimulation is to represent neural tissue by a volume conductor when calculating the activating function as the driving term in a cable equation for the membrane potential. This approach ignores the cellular composition of tissue, including the neurites and their combined effect on the extracellular potential. This has a number of undesirable consequences. First, the two natural and equally valid choices of boundary conditions for the cable equation (i.e. using either voltage or current) lead to two mutually inconsistent predictions of the membrane potential. Second, the spatio-temporal distribution of the extracellular potential can be strongly affected by the combined cellular composition of the tissue. In this paper, we develop a mean field volume conductor theory to overcome these shortcomings of available models. APPROACH: This method connects the microscopic properties of the constituent fibres to the macroscopic electrical properties of the tissue by introducing an admittivity kernel for the neural tissue that is non-local, non-instantaneous and anisotropic. This generalizes the usual tissue conductivity. A class of bidomain models that is mathematically equivalent to this class of self-consistent volume conductor models is also presented. The bidomain models are computationally convenient for simulating the activation map of neural tissue using numerical methods such as finite element analysis. MAIN RESULTS: The theory is first developed for tissue composed of identical, parallel fibres and then extended to general neural tissues composed of mixtures of neurites with different and arbitrary orientations, arrangements and properties. Equations describing the extracellular and membrane potential for the longitudinal and transverse modes of stimulation are derived. SIGNIFICANCE: The theory complements our earlier work, which developed extensions to cable theory for the micro-scale equations of neural stimulation that apply to individual fibres. The modelling framework provides a number of advantages over other approaches currently adopted in the literature and, therefore, can be used to accurately estimate the membrane potential generated by extracellular electrical stimulation.

Modelling extracellular electrical stimulation: part 4. Effect of the cellular composition of neural tissue on its spatio-temporal filtering properties.

OBJECTIVE: The objective of this paper is to present a concrete application of the cellular composite model for calculating the membrane potential, described in an accompanying paper. APPROACH: A composite model that is used to determine the membrane potential for both longitudinal and transverse modes of stimulation is demonstrated. MAIN RESULTS: Two extreme limits of the model, near-field and far-field for an electrode close to or distant from a neuron, respectively, are derived in this paper. Results for typical neural tissue are compared using the composite, near-field and far-field models as well as the standard isotropic volume conductor model. The self-consistency of the composite model, its spatial profile response and the extracellular potential time behaviour are presented. The magnitudes of the longitudinal and transverse components for different values of electrode-neurite separations are compared. SIGNIFICANCE: The unique features of the composite model and its simplified versions can be used to accurately estimate the spatio-temporal response of neural tissue to extracellular electrical stimulation.

Rabi resonance in spin systems: theory and experiment.

The response of a magnetic resonance spin system is predicted and experimentally verified for the particular case of a continuous wave amplitude modulated radiofrequency excitation. The experimental results demonstrate phenomena not previously observed in magnetic resonance systems, including a secondary resonance condition when the amplitude of the excitation equals the modulation frequency. This secondary resonance produces a relatively large steady state magnetisation with Fourier components at harmonics of the modulation frequency. Experiments are in excellent agreement with the theoretical prediction derived from the Bloch equations, which provides a sound theoretical framework for future developments in NMR spectroscopy and imaging.

Modeling individual human motor behavior through model reference iterative learning control.

A computational model is proposed in this paper to capture learning capacity of a human subject adapting his or her movements in novel dynamics. The model uses an iterative learning control algorithm to represent human learning through repetitive processes. The control law performs adaptation using a model designed using experimental data captured from the natural behavior of the individual of interest. The control signals are used by a model of the body to produced motion without the need of inverse kinematics. The resulting motion behavior is validated against experimental data. This new technique yields the capability of subject-specific modeling of the motor function, with the potential to explain individual behavior in physical rehabilitation.

Bayesian model evidence for order selection and correlation testing.

Model selection is a critical component of data analysis procedures, and is particularly difficult for small numbers of observations such as is typical of functional MRI datasets. In this paper we derive two Bayesian evidence-based model selection procedures that exploit the existence of an analytic form for the linear Gaussian model class. Firstly, an evidence information criterion is proposed as a model order selection procedure for auto-regressive models, outperforming the commonly employed Akaike and Bayesian information criteria in simulated data. Secondly, an evidence-based method for testing change in linear correlation between datasets is proposed, which is demonstrated to outperform both the traditional statistical test of the null hypothesis of no correlation change and the likelihood ratio test.

Patient-specific bivariate-synchrony-based seizure prediction for short prediction horizons.

This paper evaluates the patient-specific seizure prediction performance of pre-ictal changes in bivariate-synchrony between pairs of intracranial electroencephalographic (iEEG) signals within 15min of a seizure in patients with pharmacoresistant focal epilepsy. Prediction horizons under 15min reduce the durations of warning times and should provide adequate time for a seizure control device to intervene. Long-term continuous iEEG was obtained from 6 patients. The seizure prediction performance was evaluated for all possible channel pairs and for different prediction methods to find the best performing channel pairs and methods for both pre-ictal decreases and increases in synchrony. The different prediction methods involved changes in window duration, signal filtering, thresholding approach, and prediction horizon durations. Performance for each patient, for all seizures, was first compared with an analytical-Poisson-based random predictor. The performance of the top 5% of channel pairs for each patient closely matched the top 5% of analytical-Poisson-based random predictor performance indicating that patient-specific, bivariate-synchrony-based seizure prediction could be random in general (under the assumption that channel-pair prediction times are statistically independent). Analysis of the spatial patterns of performance showed no clear relationship to the seizure onset zone. For each patient the best channel pair showed better performance than Poisson-based random prediction for a selected subset of prediction thresholds. Given the caveats of comparing with this form of random prediction, alarm time surrogates were employed to assess statistical significance of a four-fold out-of-sample cross-validation analysis applied to the best channel-pairs. The cross-validation analysis obtained reasonable testing performance for most patients when performance was compared to random prediction based on alarm time surrogates. The most significant case was a patient whose testing set sensitivity and false positive rate were 0.67±0.09 and 3.04±0.29h(-1), respectively, for decreases in synchrony, an intervention time of 15min and a seizure onset period of 5min. For each testing set for this patient, performance was better than that obtained by random prediction at the significance level of 0.05 (average sensitivity of 0.47±0.05). Moreover, there were 9 seizures in each testing set which gives greater power to this cross-validation result, although the cross-validation was performed on the best channel pair selected by within-sample optimization for all seizures of the patient. Further validation with larger datasets from individual patients is needed. Improvements in prediction performance should be achievable through investigations of multivariate synchrony combined with non-linear classification methods.