Heba Khamis

Seizure state detection of temporal lobe seizures by autoregressive spectral analysis of scalp EEG

OBJECTIVE To investigate a novel application of autoregression (AR) spectral techniques for seizure detection from scalp EEG. METHODS EEGs were recorded from twelve patients with left temporal lobe epilepsy. The Burg maximum entropy AR method was applied to the signals from four electrodes near the epileptic focus for each patient, and the AR spectra were parameterized based on scalp EEG features described by a neurologist, thus mimicking clinical seizure identification. The parameters measured spectral peak power, sharpness, and location in a delta/low theta frequency range. An optimized nonlinear seizure detection index, which accounted for spatial and temporal persistence of behavior, was then calculated. RESULTS Performance was optimized using recordings from two patients (315h, 18 seizures). For the remaining 10 patients (1624h, 83 seizures) results are presented as a Receiver Operating Characteristic graph, yielding an overall event-based true positive rate of 91.57% and epoch-based false positive rate of 3.97%. CONCLUSIONS Performance of the AR seizure identification method is comparable to other approaches. Techniques such as artifact removal are expected to improve performance. SIGNIFICANCE There is a real potential for this seizure detection method to be of practical clinical use in long-term monitoring.

Frequency–moment signatures: A method for automated seizure detection from scalp EEG

OBJECTIVES To investigate patient-specific automated epileptic seizure detection from scalp EEG using a new technique: frequency-moment signatures. METHODS Signatures were calculated from 32s blocks of data of electrode differences from the right (RH) and left hemisphere (LH). Discrete Fourier transforms of 15 data subsets were calculated per block per hemisphere. The spectral powers at a given frequency from the RH and LH were combined into a single quantity. The signature elements were found by subtracting normalised central moments of the subset distribution from the mean, to measure the consistency of the spectral power at a given frequency over all subsets. The seizure measure was the logarithm of the probability that a signature belonged to a control set of non-seizure signatures. RESULTS Following the optimisation of signature parameters using three one-day recordings from each of 12 patients, performance was tested on a separate set of data from the same patients. The method had a sensitivity of 91.0% (total 34 seizures) with 0.020 false positives per hour (total 618 h). CONCLUSIONS Frequency-moment signatures promise automated seizure detection sensitivities comparable to visual identification and other published methods, with improved false detection rates. SIGNIFICANCE This technique has the potential to be used more widely in EEG analysis.

Detection of temporal lobe seizures and identification of lateralisation from audified EEG

OBJECTIVE To investigate the accuracy of human listeners in identifying epileptic seizures and seizure lateralisation from audified EEG signals. METHODS EEG data from 17 temporal lobe epilepsy patients (9 male, 8 female; aged 23-55) was converted to audio format by 60× time compression. Using a subset of 19% of the data, five auditory participants (2 female, 3 male; aged 23-58) were trained to identify seizures and their lateralisation by listening to audified EEG signals from difference electrodes P3-T5 and P4-T6. Following training, seizure detection performance of the auditory participants was tested using the remaining data. RESULTS Allowing a 5s auditory time margin for successful detection, the mean sensitivity of the five auditory participants was 89.6% (SD 8.3%) with a false detection rate of only 0.0068/h (SD 0.0077/h). The mean accuracy of seizure lateralisation identification was 77.6% (SD 7.1%). CONCLUSIONS With a limited amount of training, humans can detect seizures and seizure lateralisation from audified EEG signals of electrodes P3-T5 and P4-T6 with accuracy comparable to visual assessment of full EEG traces (21 electrodes) by an expert encephalographer. SIGNIFICANCE A more efficient and accurate clinical tool for assessing EEG data based on audification may be developed, which will improve diagnosis and treatment of epilepsy.

QRS Detection Algorithm for Telehealth Electrocardiogram Recordings

<italic>Objective:</italic> QRS detection algorithms are needed to analyze electrocardiogram (ECG) recordings generated in telehealth environments. However, the numerous published QRS detectors focus on clean clinical data. Here, a “UNSW” QRS detection algorithm is described that is suitable for clinical ECG and also poorer quality telehealth ECG. <italic>Methods:</italic> The UNSW algorithm generates a feature signal containing information about ECG amplitude and derivative, which is filtered according to its frequency content and an adaptive threshold is applied. The algorithm was tested on clinical and telehealth ECG and the QRS detection performance is compared to the Pan–Tompkins (PT) and Gutiérrez–Rivas (GR) algorithm. <italic>Results:</italic> For the MIT-BIH Arrhythmia database (virtually artifact free, clinical ECG), the overall sensitivity (<italic>Se</italic>) and positive predictivity (+<italic>P</italic>) of the UNSW algorithm was >99%, which was comparable to PT and GR. When applied to the MIT-BIH noise stress test database (clinical ECG with added calibrated noise) after artifact masking, all three algorithms had overall <italic>Se</italic> >99%, and the UNSW algorithm had higher +<italic>P</italic> (98%, <italic>p</italic> < 0.05) than PT and GR. For 250 telehealth ECG records (unsupervised recordings; dry metal electrodes), the UNSW algorithm had 98% <italic>Se</italic> and 95% +<italic>P</italic> which was superior to PT (+<italic>P</italic>: <italic>p</italic> < 0.001) and GR (<italic>Se</italic> and +<italic>P</italic>: <italic>p</italic> < 0.001). <italic>Conclusion:</italic> This is the first study to describe a QRS detection algorithm for telehealth data and evaluate it on clinical and telehealth ECG with superior results to published algorithms. <italic>Significance:</italic> The UNSW algorithm could be used to manage increasing telehealth ECG analysis workloads.

Classification of Texture and Frictional Condition at Initial Contact by Tactile Afferent Responses

Adjustments to friction are crucial for precision object handling in both humans and robotic manipulators. The aim of this work was to investigate the ability of machine learning to disentangle concurrent stimulus parameters, such as normal force ramp rate, texture and friction, from the responses of tactile afferents at the point of initial contact with the human finger pad. Three textured stimulation surfaces were tested under two frictional conditions each, with a 4 N normal force applied at three ramp rates. During stimulation, the responses of fourteen afferents (5 SA-I, 2 SA-II, 5 FA-I, 2 FA-II) were recorded. A Parzen window classifier was used to classify ramp rate, texture and frictional condition using spike count, first spike latency or peak frequency from each afferent. This is the first study to show that ramp rate, texture and frictional condition could be classified concurrently with over 90 % accuracy using a small number of tactile sensory units.

Decoding tactile afferent activity to obtain an estimate of instantaneous force and torque applied to the fingerpad.

Dexterous manipulation is not possible without sensory information about object properties and manipulative forces. Fundamental neuroscience has been unable to demonstrate how information about multiple stimulus parameters may be continuously extracted, concurrently, from a population of tactile afferents. This is the first study to demonstrate this, using spike trains recorded from tactile afferents innervating the monkey fingerpad. A multiple-regression model, requiring no a priori knowledge of stimulus-onset times or stimulus combination, was developed to obtain continuous estimates of instantaneous force and torque. The stimuli consisted of a normal-force ramp (to a plateau of 1.8, 2.2, or 2.5 N), on top of which -3.5, -2.0, 0, +2.0, or +3.5 mNm torque was applied about the normal to the skin surface. The model inputs were sliding windows of binned spike counts recorded from each afferent. Models were trained and tested by 15-fold cross-validation to estimate instantaneous normal force and torque over the entire stimulation period. With the use of the spike trains from 58 slow-adapting type I and 25 fast-adapting type I afferents, the instantaneous normal force and torque could be estimated with small error. This study demonstrated that instantaneous force and torque parameters could be reliably extracted from a small number of tactile afferent responses in a real-time fashion with stimulus combinations that the model had not been exposed to during training. Analysis of the model weights may reveal how interactions between stimulus parameters could be disentangled for complex population responses and could be used to test neurophysiologically relevant hypotheses about encoding mechanisms.

An Eight-Legged Tactile Sensor to Estimate Coefficient of Static Friction: Improvements in Design and Evaluation

According to the laws of friction, in order to initiate a sliding motion between two objects, a tangential force larger than the maximum static friction force is required. This process is governed by a material constant called the coefficient of static friction. Therefore, it is of great utility for robots to know the coefficient of static friction between its gripper and the object being manipulated, especially when a stable and precise grip on an object is necessary. Furthermore, it is most useful if the robot can estimate the coefficient of static friction upon touching an object at the very beginning of a manipulation task, instead of having to further explore the object before it tries to move the object. Motivated by this issue, we have designed and in this paper, further improved a novel eight-legged tactile sensor to estimate the coefficient of static friction between a planar surface and the sensing components of the prototype sensor which will also serve as the gripper. While the basic principle of the sensor is still unchanged, here we highlight some improvements to the sensors design and evaluation, including more robustly controlled frictional angles vital for the accurate sensing and the use of a programmable xyz-stage during evaluation. The coefficients of static friction between the sensor and nine different materials were estimated and compared to a measurement obtained via traditional methods as a reference. For all testing materials, the estimated ranges cover the corresponding reference values. Good conformance with the reference coefficients is also visually indicated from a least-square fitted line of the estimated coefficients, which has a gradient close to one and an r2 value greater than 0.9.

The Bayesian Decoding of Force Stimuli from Slowly Adapting Type I Fibers in Humans.

It is well known that signals encoded by mechanoreceptors facilitate precise object manipulation in humans. It is therefore of interest to study signals encoded by the mechanoreceptors because this will contribute further towards the understanding of fundamental sensory mechanisms that are responsible for coordinating force components during object manipulation. From a practical point of view, this may suggest strategies for designing sensory-controlled biomedical devices and robotic manipulators. We use a two-stage nonlinear decoding paradigm to reconstruct the force stimulus given signals from slowly adapting type one (SA-I) tactile afferents. First, we describe a nonhomogeneous Poisson encoding model which is a function of the force stimulus and the force’s rate of change. In the decoding phase, we use a recursive nonlinear Bayesian filter to reconstruct the force profile, given the SA-I spike patterns and parameters described by the encoding model. Under the current encoding model, the mode ratio of force to its derivative is: 1.26 to 1.02. This indicates that the force derivative contributes significantly to the rate of change to the SA-I afferent spike modulation. Furthermore, using recursive Bayesian decoding algorithms is advantageous because it can incorporate past and current information in order to make predictions—consistent with neural systems—with little computational resources. This makes it suitable for interfacing with prostheses.

An eight-legged tactile sensor to estimate coefficient of static friction

It is well known that a tangential force larger than the maximum static friction force is required to initiate the sliding motion between two objects, which is governed by a material constant called the coefficient of static friction. Therefore, knowing the coefficient of static friction is of great importance for robot grippers which wish to maintain a stable and precise grip on an object during various manipulation tasks. Importantly, it is most useful if grippers can estimate the coefficient of static friction without having to explicitly explore the object first, such as lifting the object and reducing the grip force until it slips. A novel eight-legged sensor, based on simplified theoretical principles of friction is presented here to estimate the coefficient of static friction between a planar surface and the prototype sensor. Each of the sensors eight legs are straight and rigid, and oriented at a specified angle with respect to the vertical, allowing it to estimate one of five ranges (5 = 8/2 + 1) that the coefficient of static friction can occupy. The coefficient of friction can be estimated by determining whether the legs have slipped or not when pressed against a surface. The coefficients of static friction between the sensor and five different materials were estimated and compared to a measurement from traditional methods. A least-squares linear fit of the sensor estimated coefficient showed good correlation with the reference coefficient with a gradient close to one and an r2 value greater than 0.9.

Tactile afferents encode grip safety before slip for different frictions.

Adjustments to frictional forces are crucial to maintain a safe grip during precision object handling in both humans and robotic manipulators. The aim of this work was to investigate whether a population of human tactile afferents can provide information about the current tangential/normal force ratio expressed as the percentage of the critical load capacity - the tangential/normal force ratio at which the object would slip. A smooth stimulation surface was tested on the fingertip under three frictional conditions, with a 4 N normal force and a tangential force generated by motion in the ulnar or distal direction at a fixed speed. During stimulation, the responses of 29 afferents (12 SA-I, 2 SA-II, 12 FA-I, 3 FA-II) were recorded. A multiple regression model was trained and tested using cross-validation to estimate the percentage of the critical load capacity in real-time as the tangential force increased. The features for the model were the number of spikes from each afferent in windows of fixed length (50, 100 or 200 ms) around points spanning the range from 50% to 100% of the critical load capacity, in 5% increments. The mean regression estimate error was less than 1% of the critical load capacity with a standard deviation between 5% and 10%. A larger number of afferents is expected to improve the estimate error. This work is important for understanding human dexterous manipulation and inspiring improvements in robotic grippers and prostheses.