5th Dutch Bio-Medical Engineering Conference 2015
22-23 January 2015, Egmond aan Zee, The Netherlands
10:40   Motor Control, Neuro Control & Patient Models I
15 mins
Martijn Vlaar, Teodoro Solis-Escalante, Alfred Schouten
Abstract: Human motion control is facilitated by the central nervous system (CNS) in conjunction with the muscles and sensory feedback. Malfunctioning of one of the components can result in movement disorders with symptoms such as tremor and bradykinesia. The pathophysiology of most movement disorders remains unclear. Thus improved understanding of movement disorders can expedite diagnosis and enhance treatment. A common technique to study the sensory system is the event related potential (ERP), in which a brief sensory stimulus is applied to a participant many times while the evoked response in the brain is recorded using electroencephalogram (EEG). The ERP can be obtained for several types of stimuli (e.g. visual, auditory, mechanical, electrical) and is often used as a clinical tool. Campfens et al. [1] applied ramp-and-hold stretches to the wrist and found that stretches in flexion and extension direction gave a similar ERP on the scalp electrodes. The goal of this pilot study is to elaborate on these findings by using advanced signal processing and source localization techniques to find the subtle differences in the responses to flexion and extension stretches. We investigate the cortical response to a well-controlled mechanical stimulus applied to the wrists of two subjects (N=2) using a robotic manipulator (Wristalyzer by MOOG Inc., Nieuw-Vennep, The Netherlands). EEG was recorded at 2 kHz using a 128-channel amplifier (Refa system by TMSi, Oldenzaal, The Netherlands). Each subject was instructed to relax while his wrist is alternately flexed and extended (0.06 rad, 40 ms ramp and at least 500 ms hold). The wrist was moved 450 times in both directions to facilitate averaging and subsequent noise reduction. After removing noisy electrodes and artifact rejection an independent component analysis (ICA) was performed. The signal-to-noise (SNR) was calculated per component and the response at the component having the highest SNR was further investigated. The component with the highest SNR was fitted with a dipole (standard head model and electrode locations) and was for both subjects located in the contralateral sensorimotor cortex. Our results show a deflection in the ERP at approximately 20 ms after onset of movement, which can be explained by the time delay between the sensors and the brain. Movement in flexion and extension direction result in a very similar ERP, which was shown before in healthy persons and stroke survivors [1]. The ERP for both directions has a P40 and a N120 component. Subtle differences in the ERP were found between stretch directions, which were consistent over subjects. Noteworthy, the polarity of the ERP is the same for both movement directions, indicating an even nonlinear transformation in the CNS: f(-x)=f(x).
15 mins
Teodoro Solis-Escalante, Yuan Yang, Alfred Schouten, Frans van der Helm
Abstract: Classical paradigms for studying somatosensory function stimulate the peripheral nervous system to elicit time- and phase-locked responses in the brain. The electroencephalogram (EEG) measures electrical potentials at the scalp. These potentials represent electrical activity of neuronal groups in the brain. Averaging the EEG across multiple repetitions of the stimulus reduces the contribution of spontaneous brain activity and reveals stimulus-related responses. Rhythmic stimulation evokes steady-state responses at the stimulation rate and its harmonics [1, 2], indicating the nonlinear behavior of the nervous system. Steady-state responses have a higher signal-to-noise ratio than the responses evoked by transient stimuli, thus fewer repetitions of the stimulus are needed and the recording time is reduced. Lesion to the central or peripheral nervous system may alter the characteristics of the evoked responses, e.g., its topography; which highlights the clinical relevance of somatosensory evoked responses. Noteworthy, the absence of evoked responses following electrical stimulation of the median nerve successfully predicts poor recovery of upper limb function in acute stroke patients [3]. The goal of this work is to map the topography of nonlinear (frequency) components of somatosensory steady-state responses. Using a periodic stimulus with multiple frequency components, we evoked responses in harmonics and intermodulation products of the stimulation frequencies; effectively increasing the amount of information that can be retrieved. Analyzing the nonlinear components has the potential to enhance the signal-to-noise ratio (in the frequency domain) and to reveal interactions between brain structures. We analyzed high-density EEG recordings (126 channels) from eleven healthy young volunteers (right-handed), while a robotic manipulator (Wristalyzer, Moog Inc., The Netherlands) applied a mechanical stimulation to their right wrist. The participants’ task was to relax with eyes open. Using three frequency components (7, 13, and 29 Hz) we retrieved steady-state responses from a total of 30 frequencies, i.e., the three stimulation frequencies, and their second and third order harmonics and intermodulation products. We calculated inter-trial phase-coherence (ITPC) [4] for each electrode (563 ± 207 epochs), and obtained the average scalp topography for each frequency across participants. Our results show frequency specific scalp topographies consistent across participants. Main responses occur at the (i) contralateral hemisphere and in (ii) midfrontal areas. These can be found one at the time or in combination, with variations on the ITPC strength of each area. The variations on scalp topography could be explained by the response of underlying sources to specific frequencies, but also to varying levels of noise across frequencies. Due to the large number of epochs and the computation of phase-coherence, the former explanation seems more plausible. Future analysis will apply this methodology to EEG recordings of stroke patients to study neurophysiological changes linked to recovery of motor function.
15 mins
Marco Mercuri, Peter Karsmakers, Bart Vanrumste, Paul Leroux, Dominique Schreurs
Abstract: Fall incidents are considered one of the most dangerous risks among elderly people worldwide, which often result in serious physical and psychological consequences. Radar techniques have been recently investigated in healthcare aiming at contactless fall detection [1]-[3]. In fact, as opposite to wire-based devices or wireless body sensor networks (WBSNs), which involve sensors attached to the human body that transmit continuously data to physicians or caregivers, they present the benefit of avoiding both discomfort and the need for actions by the person, that tend to forget wearing them. In this way, a subject can freely carry out his/her activities while being monitored remotely detecting any dangerous situation. However, a single radar sensor is insufficient for real situations. In fact, in fall detection applications, it is necessary to assess the different changes of speed experienced by a subject during normal movements and fall events [1]-[3]. However, due to the Doppler Effect, a single radar is not able to detect the speed of a target moving perpendicular to the line of sight (LoS) of the antenna. In addition, depending on the position of a person in a room, his/her reflection can be obstructed by furniture. These limitations can be overcome by means of a wireless radar sensor network (WRSN). In fact, by combining the measured information from several sensors, a better estimate of the motion is obtained. Moreover, a radar can detect absolute distance so by using multiple sensors it is possible to detect positions. In this work, as opposed to [1]-[3], a WRSN aiming at in-door contactless fall detection and positioning is presented and discussed. The network integrates three radar nodes, but the approach is scalable to higher number of nodes. Experimental results conducted in a real room setting with human subjects have demonstrated the feasibility of the proposed approach in overcoming radar’s limitations, allowing the detection of fall incidents over all orientations and in-door positioning operations by implementing a trilateration technique. REFERENCES [1] P. Karsmakers, T. Croonenborghs, M. Mercuri, D. Schreurs, and P. Leroux, “Automatic in-door fall detection based on microwave radar measurements,” in Proc. Eur. Radar Conf., Amsterdam, The Netherlands, Oct. 31-Nov. 2, 2012, pp. 202-205. [2] M. Mercuri, P. J. Soh, G. Pandey, P. Karsmakers, G. A. E. Vandenbosch, P. Leroux, and D. Schreurs, “Analysis of an indoor biomedical radar-based system for health monitoring,” IEEE Trans. Microwave Theory Tech., vol. 61, no. 5, pp. 2061-2068, May 2013. [3] C. Garripoli, M. Mercuri, P. Karsmakers, P. J. Soh, G. A. E. Vandenbosch, C. Pace, P. Leroux, and D. Schreurs, “Embedded DSP-based Telehealth Radar System for Remote In-door Fall Detection,” IEEE J. Biomed. Health Inform., 2014, accepted to be published.
15 mins
Marijn van Dongen, Wouter Serdijn
Abstract: This work presents a neural stimulator system that employs a fundamentally different way of stimulating neural tissue as compared to classical constant current stimulation. A stimulation pulse is composed of a sequence of current pulses injected at a frequency of 1 MHz for which the duty cycle is used to control the stimulation intensity. The efficacy of this type of stimulation is first verified with simulations that use dynamic tissue and axonal models. Second, the models are verified using an in vitro measurement setup in which the response of Purkinje cells in the cerebellum of a mouse is measured as a function of a high-frequency stimulation signal using patch-clamp techniques [1]. The measurements confirm that high-frequency stimulation can recruit neurons in a similar way as classical constant current stimulation. Subsequently a prototype IC design is realized that implements a high-frequency stimulator system [2]. The system features 8 independent channels that connect to any of the 16 electrodes at the output. A sophisticated control system allows for individual control of each channel's stimulation and timing parameters. This flexibility makes the system suitable for complex electrode configurations and current steering applications. Simultaneous multichannel stimulation is implemented using a high frequency alternating technique, which reduces the amount of electrode switches by a factor 8. The system has the advantage of requiring a single inductor as its only external component. Furthermore it offers a high power efficiency, which is nearly independent on both the voltage over the load as well as on the number of simultaneously operated channels. Measurements confirm this: in multichannel mode the power efficiency can be increased for specific cases to 40% compared to 20% that is achieved by state-of-the-art classical constant current stimulators with adaptive power supply.
15 mins
Jörn Kretschmer, Achim Speck, Sunetra Basavaraju, Axel Riedlinger, Christoph Schranz, Knut Möller
Abstract: Mechanical ventilation is a life-saving intervention helping a patient to overcome phases of impaired lung function. The clinician will choose ventilator settings to ensure sufficient oxygenation and carbon dioxide removal. However, in critically ill patients, those settings might lead to a tradeoff between sufficient ventilation and further damage to the lung tissue. Mathematical models allow prediction of patient reaction to changes in therapy that may be exploited to optimize therapy settings. Complex models of high physiological detail allow simulation of various disease patterns but require a high number of clinical data in order to robustly identify model parameters. Thus, we propose to employ multiple models differing in detail and simulation focus. To enable automated usage of those models in decision support systems, numerical selection criteria need to be introduced that help to select the model that best fits the given data. Clinical data of ten mechanically ventilated patients were selected from a previous ARDS-Study [1]. Each patient was ventilated in volume-controlled mode applying different maneuvers to extract relevant information for parameter identification. The applied maneuvers included low flow maneuver (LF) to test lung mechanics under quasi-static conditions with constant and slow inspiration flow, dynamic slice maneuver (DS) to quantify dynamic properties of the lung tissue, and SCASS maneuver which occludes the airways after reaching a randomized inspiration volume to ensure a quasi-static pressure-volume relationship. Four different models of respiratory mechanics were each fitted to all maneuvers and fit results were analyzed with three numerical criteria. Mathematical models were first order model (FOM, 2 parameters), viscoelastic model (VEM, 4 parameters), recruitment model (PRM, 5 parameters) and viscoelastic recruitment model (PRVEM, 7 parameters). Selection criteria included coefficient of determination (CD), confidence interval (CI) and the corrected Akaike Information Criterion (AICc) [2]. CD describes how well a model is able to reproduce the recorded data, while AICc additionally includes the number of free model parameters as a penalty. CI values are parameter specific and describe how much a parameter value may be altered without affecting the fit outcome. Thus, if CI is high compared to the parameter value, the parameter is not sufficiently defined by the recorded patient data. Results of LF showed best CD and AICc in PRM and PRVEM models. PRM however presented high CI in multiple patients, disqualifying it from use in this maneuver. In DS, VEM and PRVEM both showed best CD and AICc, while also presenting high CI in multiple patients. Here FOM lead to the best overall results. In SCASS, VEM and PRVEM both showed best CD and AICc, with PRVEM being disqualified by high CI. In summary, using a combination of CD, CI and AICc allows identifying the model that best fits the given data while ensuring that identified parameter values are sufficiently defined by the data.
15 mins
Axel Riedlinger, Jörn Kretschmer, Knut Möller
Abstract: Mechanical ventilation is a life-saving therapy maintaining sufficient oxygenation and removal of carbon dioxide in patients with acute lung failure. However, inappropriate ventilator settings bare the risk of additional lung injury. The use of mathematical models can support clinicians in decision making by enabling them to test ventilator settings in a simulated environment that reproduces the physiological properties of an individual patient. Ventilator settings are then based on multiple simulations finding the one setting that leads to the clinical results specified by the clinician. Models of human pulmonary gas exchange allow the simulation of blood gas levels with respect to inspired oxygen fraction FiO2. Robust parameter identification is a basic condition for reliable model-based decision support in clinical practice, i.e. the identification algorithm must be able to find physiologically sound parameter values under noisy conditions. Parameter identification of a three-parameter pulmonary gas exchange model with shunt and two alveolar compartments was analysed. Three synthetic data sets were generated each representing a different degree of lung damage. Structural identifiability was investigated by employing different combinations of the model parameters and comparing the resulting simulations with the synthetic data sets. Practical identifiability was analyzed by multiple parameter identification attempts using different initial values for the model parameters. Results of the analysis with synthetic data were also evaluated retrospectively with three real patient data sets. The error surface of the objective function showed two global minima for all tested data sets. However, structural identifiability is still given because of the model being symmetric meaning that both minima are describing the same result. Analysis of practical identifiability showed different solutions for the applied initial parameter values. Initialisations of parameters are shown to have a strong effect on identification results. Robustness of identification could be improved by using a simple shunt model to calculate an initial value for the shunt fraction of the three-parameter model. Structural identifiability for the three-parameter model of human gas exchange was shown allowing the use in clinical practice. Practical identifiability was improved by using knowledge from a simpler model, i.e. calculation of an appropriate initial value for shunt using a simple model helps in identifying correct parameter values of the three-parameter model.