SSBE 2019 Annual Meeting
Aug 27, 2019 - Campus Biotech, chemin de mines 9, Genève
Selected Oral
Contributions
AUTOMATED OPTIMIZATION OF SPINAL CORD STIMULATION FOR ENABLING MOTOR ACTIVITY AFTER SPINAL CORD INJURY.
Pedro Abranches1,2, Henri Lorach1,2, Salif Komi1,2, Robin Demesmaeker1,2, Laura McCracken1,2, Edeny Baaklini1,2, Fabien B. Wagner1,2, Jocelyne Bloch2,3, Grégoire Courtine1,2
1 Center for Neuroprosthetics and Brain Mind Institute, École Polytechnique Fédérale de Lausanne, Switzerland
2 Department of Clinical Neuroscience, Lausanne University Hospital, Lausanne, Switzerland
3 Department of Neurosurgery, Lausanne University Hospital, Lausanne, Switzerland
INTRODUCTION: We recently demonstrated that spatiotemporal stimulation of the spinal cord enables the restoration of leg motor control and walking capabilities in individuals with severe spinal cord injury [1]. However, the fine tuning of these stimulation protocols (electrode configurations, stimulation amplitude, frequency, pulse width, and timing of different stimulation trains) is an extremely time-consuming process that does not guarantee optimal convergence. An automated approach to this optimization problem is needed to translate this strategy into an actual therapy which can be used in standard physiotherapy sessions.
METHODS: To this purpose, we developed a framework based on a Bayesian Optimization approach, where a Gaussian Process Upper Confidence Bound (GP-UCB) algorithm was implemented to learn and choose pre-defined stimulations in a constructed search-space. The GP- UCB algorithm chooses a specific stimulation protocol (i.e. a set of active electrodes, amplitude and frequency), delivers a 500-ms pulse train to an electrode array placed over the lumbosacral spinal cord and records the electromyographic activity (EMG) of the main leg muscles. It then computes a reward function based on muscle selectivity indices for a targeted functionality such as hip flexion [1], and updates its posterior mean and variance, which translate into predictions of reward and uncertainty.
This approach was tested in patients with severe spinal cord injury enrolled in our ongoing clinical study STIMO [1]. An electrode array (Specify 5-6- 5, Medtronic) was surgically implanted epidurally over the posterior aspect of the lumbosacral spinal cord and connected to an implantable pulse generator (Activa RC, Medtronic). All experimental sessions were conducted with the patient awake, lying supine on a table and with non-invasive EMG sensors (Trigno EMG, Delsys). A dataset comprised of different stimulation parameters and corresponding EMG signals was first built for prototyping the algorithm. After tuning of the hyperparameters, the algorithm was then used during the experimental sessions to explore in real- time the space of stimulation parameters in order to find the parameters with the best reward value.
RESULTS: Using the previously mentioned dataset, our algorithm was able to find the best stimulation under 20 iterations out of 600 possible choices in a consistent manner. In Fig. 1, a representation of the algorithm choices during a trial is depicted. As expected, we see the algorithm exploring the space of parameters in the first iterations. Nevertheless, after 18 iterations it converges to a parameter subspace associated with high reward values. During the real-time experiments, in a search space of 470 possible stimulations, the algorithm explored consistently stimulations that produced the desired muscle functionality within 15 iterations.
Fig. 1: Stimulations parameters explored by the GP-UCB algorithm and associated reward values for each iteration during one run of the algorithm.
DISCUSSION & CONCLUSIONS: One of the main challenges with automatic optimization of stimulation parameters is the dimensionality of the search space. In our case, we addressed this issue by constraining the search space based on prior knowledge of possible electrode configurations. We show that in this setting, the algorithm does a structured exploration of the search space, converging quickly to high reward subspaces without stopping in local maxima. Moreover, it shows promising results with respect to real-time experiments. These results pave the way towards real-time automatic optimization of stimulation parameters for targeted spinal cord stimulation after injury.
REFERENCES: 1 F. B.Wagner, J.-B. Mignardot, C. G. Le Go_-Mignardot et al. (2018), Targeted neurotechnology restores walking in humans with spinal cord injury. Nature, vol. 563, no. 7729, p. 65.
ACKNOWLEDGEMENTS: Innosuisse (CTI grant number 25761.1)
NANOINDENTATION MAPPING OF THE HETEROGENOUS MATERIAL PROPERTIES OF THE HUMAN CORNEA
Miguel Ángel Ariza-Gracia1, Murielle Lerch1, Jiri Nohava2, Philippe Büchler 1
1 ARTORG Center for Biomedical Engineering Research, Switzerland
2 Anton Paar TriTec, Switzerland INTRODUCTION: Chemical crosslinking (CXL)
is a novel technique in refractive surgery. CXL stiffens the cornea locally, which results in a modification of ocular optical properties. As the CXL agent is delivered to the tissue using passive diffusion, the resulting spatial distribution of the stiffening remains poorly described.
In this study, we use nanoindentation tests to identify the spatial distribution of the corneal stiffness before and after CXL. Experimental measurements will be used to identify the parameters of a material model of the cornea.
Since cornea is formed by a collagen network filled with water1, the tissue was described using a biphasic material model in which the matrix is characterized by a viscoelastic material.
METHODS: Using a bioindenter from Anton Paar TriTec2, nanoindentation creep tests were carried out in 11 control and 11 CXL donor eyes.
A spherical ruby tip (R=1 mm) was used to measure in three different corneal locations: central (C: 0-1 mm), para-central (B: 1-2.5 mm), and peripheral (A: 2.5-4 mm). Force-controlled experiments were performed; a load of 50 µN was gradually applied with a ramp of 30 s, followed by a plateau of constant force for 180 s. Displacement was recorded during the whole experiment.
This experiment setup was reproduced numerically using the FE software FEBio. Cornea was modelled using a poroviscoelastic material while the indenter was described as a rigid body. The average of the experimental force was applied on the indenter.
A grid-based optimization was carried out to identify the 4 material parameters of the model (E:
Young’s modulus; k: permeability; γ: viscoelastic coefficient; τ: relaxation time; ν=0.075). These mechanical parameters were calibrated using a
grid-based optimization based on ~10,000 simulations for each region (A, B, C) of the control and CXL samples.
The loss function compared the average numerical and experimental displacement of the indenter.
The minimum and the local minima were analyzed to determine the optimal material parameters (see Fig.1-left) and characterize the quality of the optimization.
RESULTS: Control and CXL corneas showed a heterogeneous distribution of the material properties. The E was higher in the center (C) and decreasing towards the periphery (A). Also, CXL increased E up to 3 times in the center (C) and up to 2 times at the mid-periphery (B). On the other hand, the behavior was similar for control and CXL samples at the periphery (A) (see Fig.1- right). The remaining parameters were also affected: k decreased only at the periphery for CXL; γ decreased uniformly for CXL; and τ remained constant.
DISCUSSION: Nanoindentation is able to characterize the spatial modification in corneal biomechanics associated with CXL. Results of the parameter identification showed that a poroviscoelastic models successfully reproduce the experimental data. Based on the parameters identified for the different region, we were able to determine the region of the cornea affected by CXL. Moreover, results indicate that CXL not only significantly increases the Young’s modulus, but also decreases the permeability of the tissue.
REFERENCES: 1 Forrester J.V. et al, The Eye, 4th Ed. Elsevier, 2016 . 2 Nohava J et al, J Biomed Mat Research, 106(5):1413-1420, 2018.
ACKNOWLEDGEMENTS: Ariza-Gracia was funded by project 786692 (H2020-MSCA).
TARGETED SPINAL STIMULATION AFTER SPINAL CORD INJURY:
FROM FIRST-IN-HUMAN TO LARGE-SCALE CLINICAL APPLICATION
Robin Demesmaeker
1,2, Salif Komi
1,2, Edeny Baaklini
1,2, Andreas Rowald
1,2, Henri Lorach
1,2, Miroslav Caban
3,4, Marina d’Ercole
4, Anne Watrin
4, Maryse van ‘t Klooster
4, Aude Yulzari
4, Rik
Buschman
5, Joachim von Zitzewitz
4, Vincent Delattre
4, Hendrik Lambert
4, Fabien B. Wagner
1,2, Jocelyne Bloch
2,6and Grégoire Courtine
1,2,61
Center for Neuroprosthetics and Brain Mind Institute, École Polytechnique Fédérale de Lausanne, Switzerland
2Department of Clinical Neuroscience, Lausanne University Hospital, Switzerland
3GTX medical, The Netherlands
4Institute of Bioengineering, École Polytechnique Fédérale de Lausanne, Switzerland
5Department of Neurosurgery, Lausanne University Hospital,
Switzerland
6Medtronic, USA
INTRODUCTION: Recent evidence suggests that epidural electrical stimulation of the lumbar spinal cord combined with intensive locomotor training can restore locomotion in individuals with chronic spinal cord injury (SCI) 1,2,3. We developed activity-specific stimulation programs modulated in space and time that are tailored to the patient’s needs. These initial results encouraged us to expand our clinical approach with technical improvements, enlargement of the patient population, and dissemination to a large multinational, multicenter scale.
METHODS: In the first-in-human clinical study (STIMO), 6 individuals with chronic spinal cord injury (>2 years post injury, AIS scores C-D) were implanted with a commercially available spinal electrode array generally used for chronic pain treatment (Surescan Specify 5-6-5, Medtronic) linked to an implanted pulse generator (ActivaRC, Medtronic) built for deep brain stimulation. After surgery, a custom-built communication system and lab environment enable the personalized application of the stimulation. After initial optimization, patients followed an intensive 5 months rehabilitation program including different locomotor activities such as walking and cycling.
In a second part of this study (STIMObridge), patients (including AIS scores A-B) are implanted with an electrode paddle array specifically designed to target the posterior roots, and thus improve spatiotemporal stimulation. Moreover, an advanced personalized MRI-based computational model is used to guide the surgical placement of the implant. After surgery, patients follow the same training protocol as in the first study.
Furthermore, an international multi-center clinical trial is going to test the portability of the therapy to multiple clinical centers and will evaluate the feasibility and preliminary efficacy of the therapy when applied in the subacute phase after SCI (~
after 4 weeks). Stimulation optimization and
training sessions will be integrated in the dense standard clinic-specific rehabilitation programs.
RESULTS: Spatiotemporally patterned spinal cord stimulation immediately restored walking in 7 participants and intensive locomotor training led to the recovery of voluntary movements even in the absence of stimulation. The optimized spinal cord implant showed an increased selectivity in the recruitment of spinal roots and a better coverage of the targeted lumbar region. A newly developed software platform combines the whole range of original lab environments, thus paving the way towards the multicenter clinical trial. This software guides the optimization and evaluation of spinal cord stimulation paradigms by expert users and therapists while enabling patients to perform different motor activities on their own.
DISCUSSION & CONCLUSIONS: Spatio- temporal stimulation of the spinal cord enables voluntary motor abilities in patients with SCI.
Through a series of subsequent clinical studies, the stimulation system and therapy are progressively adapted and improved with the ultimate aim to disseminate this treatment across clinical centers worldwide and to address a broad variety of injuries and motor activities.
REFERENCES: 1 C. A. Angeli, et al. (2018), Recovery of Over-Ground Walking after Chronic Motor Complete Spinal Cord Injury. New England Journal of Medicine, Vol. 379, p. 1244. 2 M. L.
Gill, et al. (2018), Neuromodulation of lumbosacral spinal networks enables independent stepping after complete paraplegia. Nature Medicine, vol. 24, p. 1677. 3 F. B. Wagner, et al.
(2018), Targeted neurotechnology restores walking in humans with spinal cord injury. Nature, vol.
563, no. 7729, p. 65.
ACKNOWLEDGEMENTS: Clinical study approved under Swissethics 04/2014 PB_2016- 00886 and Swissmedic 2016-MD-0002.
A RECURRENT NEURAL NETWORK TO CLASSIFY ATRIAL FIBRILLATION FROM PPG INTERBEAT INTERVAL
Elsa Genzoni
1, Jérôme Van Zaen
1, Mathieu Lemay
1, Etienne Pruvot
3, Fabian Braun
1, and Jean-Marc Vesin
21
Swiss Center for Electronics and Microtechnology (CSEM), Neuchâtel, Switzerland
2
Applied Signal Processing Group, École Polytechnique Fédérale de Lausanne, Switzerland
3
Arrhythmia Unit, Cardiology service, Heart and Vascular Department, Lausanne University Hospital, Lausanne, Switzerland
INTRODUCTION: Atrial fibrillation (AF) is the most common arrhythmia affecting millions of individuals worldwide and is frequently associated with cardiac diseases and comorbidities1. The paroxysmal nature of AF in its early stage makes it especially difficult to diagnose with ambulatory electrocardiogram (ECG) monitors. AF is often diagnosed during hospitalization at time of complications2. The recent development of photoplethysmographic (PPG) technology allows to monitor pulse rate with simple and low-cost means3. The increasing availability of wearable PPG devices, such as smartwatch, appears as a promising solution for AF screening in large populations. In this study, we evaluate the performance of a PPG peak detector and a recurrent neural network to detect AF from PPG signals collected with a wrist-worn device.
METHODS: Simultaneous PPG and 12-lead ECG recording were analyzed in 21 patients referred for catheter ablation of cardiac arrhythmias. The ECG signals were automatically labelled as AF or no- AF episodes (sinus rhythm (SR), supra- and ventricular arrhythmias) by a system composed of a QRS detector (to extract RR intervals) and a classifier, which was validated against expert’s annotations. After aligning the PPG and ECG signals, the probability of PPG data to be AF was estimated within 30 seconds windows using a stepwise algorithm composed of: a pulse detector extracting interbeat intervals; a quality index estimator discarding non-consistent pulses; a neural network with a gated recurrent unit and a logistic layer detecting AF. The neural network was trained on ECG-based RR intervals from the PhysioNet Long-Term AF Database. Performance metrics were computed by comparing PPG-based predictions to the ECG-based reference.
RESULTS: 1130 non-overlapping windows (187 AF, 943 no-AF) of PPG signals were analyzed.
The labels attributed to both PPG and ECG signals were compared and the results are provided in Table 1. The algorithm achieved an accuracy,
sensitivity and specificity of 96% for all when discriminating AF from no-AF episodes.
Table 1. Confusion matrix
PPG predictions
AF No-AF
ECG references
AF 179 8
No-AF 36 907
DISCUSSION & CONCLUSIONS: Our approach identifies with high accuracy AF episodes from PPG-based sequences containing SR as well as various arrhythmia types. Our algorithm, recently embedded in wearable device, paves the way for the screening of AF with limited upstream interventions. Further studies are needed with a more balanced distribution between AF and no-AF to strengthen the reliability of the results.
Moreover, the robustness of the peak detection algorithm and neural network classifier should be tested against additional types of arrhythmias.
REFERENCES: 1 M. Zoni-Berisso, F. Lercari, T.
Carazza, and S. Domenicucci, “Epidemiology of atrial fibrillation: European perspective,” Clin.
Epidemiol., vol. 6, p. 213, Jun. 2014. 2 N. Hannon et al., “Stroke associated with atrial fibrillation – incidence and early outcomes in the North Dublin population stroke study.,” Cerebrovasc Dis, vol.
29, pp. 43–49, 2010.3 T. Tamura, Y. Maeda, M.
Sekine, and M. Yoshida, “Wearable Photoplethysmographic Sensors—Past and Present,” Electronics, vol. 3, no. 2, pp. 282–302, 2014.
ACKNOWLEDGEMENTS: The authors would like to thank collaborators from the CHUV for providing the clinical data. The authors would also like to thank the collaborators from the CSEM who were deeply involved in the development of PPG-based technologies for the last 15 years.
3D BIOFIDELIC MODELING OF FALLS FROM CLINICAL DXA IMAGES
L. Grassi
1, I. Fleps
2, H. Sahlstedt
1, S.P. Väänänen
3, S. Ferguson
2, B. Helgason
2, H. Isaksson
11
Department of Biomedical Engineering, Lund University, Lund, Sweden
2
Institute for Biomechanics, ETH Zürich, Zürich, Switzerland
3
Department of Applied Physics, University of Eastern Finland, Kuopio, Finland
INTRODUCTION: Osteoporotic hip fractures result from low trauma such as a fall from standing height, and represent a major socioeconomic issue.
Fracture prevention is crucial, but current diagnostics based on bone mineral density (BMD) fails to identify 30-50% of the subjects at risk [1].
3D finite element (FE) models can improve the prediction accuracy but need to be compatible with the current clinical settings where 2D radiological images (DXA) are used. The aim of this study was to combine a biofidelic fall simulation technique, based on 3D computed tomography (CT) data [2]
with an algorithm to reconstruct 3D femoral shape and BMD distribution from a 2D DXA image [3].
METHODS: The CT scan (0.781x0.781x0.3 mm voxel size) of the pelvis of a 54-year old female donor was retrieved. A simulated DXA was obtained from CT using CTXA Hip module in QCT Pro [4]. The left femur shape and BMD distribution were reconstructed from the 2D DXA using a statistical shape and appearance model (SSAM) and a genetic optimization algorithm which found the SSAM instance whose 2D projection best resembled the target DXA [3]. The 3D reconstructed femur was automatically converted into an FE mesh and inserted into the validated biofidelic FE modeling pipeline proposed by Fleps et al [2]. Results were presented for (i) reconstruction accuracy (point-to-surface distance and BMD distribution) and (ii) FE predictions (peak force, fracture outcome, strain patterns of elements having past ultimate strain).
RESULTS: Mean point-to-surface distance was 1.18 mm. The reconstruction was more accurate in the head (0.89 mm) and neck (0.75 mm) regions (figure 1).
Fig. 1: Reconstruction error for the shape (left) and BMD (right).
The mean of the absolute BMD difference was 0.12±0.13 g/cm3 (figure 1). The FE predictions showed a difference between CT-based and reconstructed models of 17 % for the peak force.
CT and DXA-based simulations both predicted cervical femur fractures (figure 2).
Fig. 2: Comparison of volumetric strains: CT- based (left) and DXA-based (right) FE predictions.
DISCUSSION & CONCLUSIONS: This study proposed a method to generate complex 3D biofidelic FE models of sideways fall from a clinically available 2D DXA scan. This enables to perform mechanistic predictions of fracture risk in a normal clinical scenario where 2D DXA is available. The reconstruction accuracy was within one voxel size of the original CT, and BMD error was 7% of the maximum BMD. Reconstructed FE models predicted a fracture as the outcome of the biofidelic fall, same as the CT-based FE model, despite 17% difference in peak force. Future works will extend the validation to the additional 10 samples presented in [2]. If successfully implemented, this technique could be used to assess fracture risk on bigger clinical cohorts.
REFERENCES: 1 Pasco et al (2006) Osteoporos Int. 17(9):1404-09. 2 Fleps et al (2019) JBMR Preprint. 3 Väänänen et al (2015) Med Imag Anal 24(1):125-134. 4 Cann et al (2014) Plos One 9(3):e91904.
ACKNOWLEDGEMENTS: funding from Swiss National Science Foundation (project IZSEZ0 186596), the Royal Physiographic Society of Lund and Swedish Research Council (2015-4795).
Ethics approval from UBC Clinical Research Ethics Board (Study ID H06-70337).
VALIDATION OF AN EXPLICIT FINITE ELEMENT MODEL FOR SIMULATING IMPACT RESPONSE OF CaP-Ti CRANIAL IMPLANTS
Susanne Lewin
1, Ingmar Fleps
2, Dominique Neuhaus
2, Jonas Åberg
1,3, Cecilia Persson
1, Stephen. J.
Ferguson
2, Caroline Öhman-Mägi
1, Benedikt Helgason
21
Division of Applied Materials Science, Uppsala University, Sweden
2
Institute for Biomechanics, ETH Zürich, Zürich, Switzerland
3
OssDsign AB, Uppsala, Sweden
INTRODUCTION: Cranioplasty is a procedure performed following neurosurgery or trauma that aims to repair cranial defects using autologous bone or synthetic implants. A high overall complication rate (~20%) has been reported, the most common being bone resorption and infection [1], [2]. Conclusive agreement on the most optimal treatment is lacking [2], [3]. The present study focuses on a relatively new patient specific titanium–reinforced calcium phosphate (CaP-Ti) implant system, which has shown promising clinical results [4]. The specific aim is to validate finite element models (FEMs) of this implant system, against impact test data, for potential future use in subject specific pre-surgical planning.
METHODS: A generic implant system with a slightly curved circular shape was modelled in a CAD software (Rhinoceros 3D, release 5, Robert McNeel & Associates). Six implants, that were produced based on the CAD data, were subjected to impact where a mass of 5 kg was dropped from a height of 0.12 m onto the specimens. Force time- response was measured with an impact load cell.
Displacement of the implant system was measured with a high speed camera. A FEM was created based on the CAD data using Ansa (vers. 17.1.0, Beta CAE Systems, Switzerland). CaP material properties were defined as non-linear based on literature data [5]. The non-linear tensile properties of the titanium part of the implant system were measured directly on laser-sintered specimens. The FEM was solved in a commercial explicit FE solver (LS-Dyna, Livermore, USA).
Fig. 1: Experimental setup (left). Comparison of measured (gray) and FEM derived (blue) force- displacement response (right).
RESULTS: The FEM derived force-displacement response was found to closely match the measured response albeit with slight overestimation of maximum impact force (Fig. 1). Comparison of fracture patterns on the physical specimens after impact testing was also qualitatively matched by strain patterns derived from FEM results (Fig. 2).
Fig. 2: Principal compressive (left) and tensile (right) strains on surface of the implant system according to FEM results. These patterns qualitatively matched crack initiation on the physical speicimens.
DISCUSSION & CONCLUSIONS: A close match between mechanical tests and FE simulations was observed over the whole loading range of the impact. We believe that this demonstrates that these types of FEMs can reproduce the mechanical response of these implants under impact. In the future, the FEMs could be helpful in terms of implant design and pre-surgical biomechanical assessment.
REFERENCES: 1E. Neovius et al. (2010), J Plast Reconstr Aesthet Surg 63:1615–1623. 2S. E. C. M.
van de Vijfeijken et al. (2018) World Neurosurg 117:443-452. 3A. Moles et al. (2018) World Neurosurg 111:e395–e402.L. Kihlström et al.
(2019) World Neurosurg 122:e399–e407. 5I.
Ajaxon et al. (2017) J Mech Behav Biomed Mat 74:428-437.
ACKNOWLEDGEMENTS: Funding by EU, Eurostars project E9741.
HABITUAL LOADING PRIOR TO BRIDGING IS AN INDICATOR OF FRACTURE CALLUS STIFFNESS PROGRESSION IN INDIVIDUAL MICE
Graeme R. Paul (1), Esther Wehrle (1), Jianhua Zhang (1), Gisela A. Kuhn (1), Ralph Müller (1) 1. Institute for Biomechanics, ETH Zurich, Zurich, Switzerland
INTRODUCTION: Fracture healing is driven by mechanical forces within both soft tissue and newly formed bone (1). Assessing these mechanical forces in vivo poses a challenge, and hence computational models have been used to develop assumptions about mechanical loading within the defect region (2). However, these assumptions do not take into account individual physiological loading patterns, or predicting fracture healing progression on a per case basis. We have developed a system to capture physiological loading on an individualised basis and compare the self-induced mechanical forces within the fracture region with longitudinal monitoring of healing progression in a mouse femur defect model.
We were able to demonstrate that habitual loading at day 7 was a predictive factor of callus stiffness at bridging (day 14-21 post-op) and into the remodeling period (day 28-42), with higher loading correlating with improved callus stiffness properties.
METHODS: A right femur osteotomy was performed on 12 female C56BL/6J mice and an external fixator (RISystems, Landquart, Switzerland) with removable sidebars was used for fixation. Strain gauges (Vishay-Micro- measurements, Wendell, USA) were attached to aluminum sidebars (Fig 1a) and strains within these tethered fixators were captured during walking of the mice. A minimum of three runs were performed (distance of ~0.5m) and the most consistent movement burst was taken for analysis (Fig 1b).
Peak-to-peak amplitudes were calculated, requiring a peak prominence of at least half the maximum signal height, to ensure capture of a full foot-strike movement. The relative strain per peak-to-peak set was then calculated and the median relative strain determined. The strain output was converted to a percentage of bodyweight per mouse (BW) via sensor calibration. MicroCT images were acquired weekly and callus stiffness was determined using
finite element analysis. This stiffness was taken as a proxy for fracture healing progression.
RESULTS: Loading measurements from day 7 were correlated with the stiffness for weekly time points (Fig 1c). As measured by Spearman’s correlation coefficient, mechanical forces do not correlate with stiffness at the time of acquisition, indicating loading is independent of osteotomy stiffness. The highest correlation is displayed at day 21 (r=0.85, p<0.001) and, in general, a positive significant (r>0.6, p<0.05) correlation was observed between the forces at day 7 and stiffness from day 14–42. This indicates that the amount of mechanical stimulation induced by the mouse prior to bone formation is a leading factor in the fracture healing outcome. Additionally, two mice developed non- union fractures and displayed more than one standard deviation less mechanical loading at day 7 (Group mean BW loading=352%±95, Non-unions (223%, 246%), supporting the hypothesis that a shortage of initial mechanical stimulation may be responsible for the development of non-unions.
DISCUSSION & CONCLUSIONS: Via the use of in vivo measurements, we observed that self- induced mechanical forces pre-bridging were strongly correlated with fracture properties at bridging and during remodeling. Going forwards, the development of individualised models is possible, with the intention of creating targeted individualised mechanical boundary conditions, allowing quantified mechanical loading intervention to prevent non-union fractures and improve the outcome of fracture healing.
REFERENCES: 1Morgan et.al, J Biomech, 43:2418-2424, 2010, 2Betts et al., TERMIS.0230.
2017.
ACKNOWLEDGEMENTS: We thank ERC Advanced MechAGE, ERC-2016-ADG-741883 and CSCS.
NON-INVASIVE BLOOD PRESSURE MONITORING DURING ANESTHESIA INDUCTION
M. Proença
1, Y. Ghamri
2, F. Braun
1, G. Bonnier
1, G. Hofmann
2, Ph. Renevey
1, J. Solà
1, A. Axis
2, P. Schoettker
2,3, and Mathieu Lemay
11
Centre Suisse d’Electronique et de Microtechnique (CSEM), Switzerland
2
Department of Anesthesia, University Hospital Lausanne (CHUV), Switzerland
3
Faculté de Biologie et Médecine, Université de Lausanne (UNIL), Switzerland
INTRODUCTION: In most anesthetized patients, arterial pressure (AP) is routinely monitored by means of automated oscillometric brachial cuffs.
With intermittent measurements at regular intervals (e.g. 3-5 minutes), rapid intraoperative hypotensive events – associated with postoperative complications [1] – may go unnoticed. We investigated the possibility of detecting rapid deleterious changes in AP by pulse oximetry waveform analysis.
METHODS: The study was approved by the local ethics committee and registered at ClinicalTrials.gov (NCT02651558). In patients scheduled for an elective surgery necessitating invasive AP monitoring, we evaluated the performance of our pulse oximetry waveform analysis algorithm (oBPM®) [2] in detecting acute changes in AP, defined by changes of 20 % in invasive AP occurring in less than one minute.
The tracking ability of our algorithm was evaluated using four-quadrant plots and Bland-Altman analysis by comparison with the invasive reference (arterial catheterization). We report hereafter a preliminary analysis on the 15 first patients of the study.
RESULTS: Figure 1 shows the concordance between invasively-assessed mean AP (MAP) changes (MAPINV) and non-invasively-assessed MAP changes determined from pulse oximetry waveform analysis by our oBPM® algorithm (MAPoBPM). The same analysis was performed for systolic (SAP) and diastolic AP (DAP). For all three pressures, excellent concordance rates (100 %) and correlation coefficients (0.95) were found between both methods. Bland-Altman analysis showed good agreements, with percentage errors (mean SD) of 1.4 10.6 %, -1.6 11.3 %, and 0.5 10.8 %, for SAP, DAP, and MAP, respectively. The proportion of percentage errors falling within 20 % was 93.3
%, 95.5 %, and 96.4 % for SAP, DAP, and MAP, respectively.
DISCUSSION & CONCLUSIONS: Pulse oximeters are part of standard monitoring equipment in operating rooms and intensive care
units. Our proposed approach has shown the ability to detect acute AP variations in a non-selected population with no additional equipment required.
These findings suggest it could be used for the early detection of hypotensive events, or the ‘smart’
triggering of oscillometric measurements, thereby helping in reducing the duration and clinical consequences of hypotensive events.
Fig. 1: Four-quadrant plot between the invasively- assessed (MAPINV) and non-invasively assessed changes (MAPoBPM) in mean arterial pressure.
The concordance rate (CR) – i.e. the percentage of paired values showing a concordant direction of change – is of 100 %, and the correlation coefficient (r) of 0.95. A color code is used to identify changes occurring in periods of high (light color) or low (dark color) pressure.
REFERENCES: 1 M. Walsh, P.J. Devereaux, A.X.
Garg, et al (2013) Anesthesiology 119(3):507-15.
2 J. Solà, M. Proença, F. Braun, et al (2016) Curr Dir Biomed Eng 2(1):267-71.
GLYCOSAMINOGLYCANS IN AURICULAR CARTILAGE INTERACT STRONGLY WITH OTHER EXTRACELLULAR MATRIX
MACROMOLECULES
Manula S.B Rathnayake1, Colet E.M. ter Voert 1 and Kathryn S. Stok 1
1 Department of Biomedical Engineering, The University of Melbourne, Parkville, Australia
INTRODUCTION: Glycosaminoglycans (GAG) in cartilage interact chemically and structurally with other extracellular matrix macromolecules to create a load-bearing functional environment.
Differences in these interactions from one cartilage type to another have been overlooked in the literature. To investigate these interactions, GAG in cartilage can be readily digested with enzymes. The ability of enzymes to digest GAG depends on tissue permeability and the affinity of GAG for surrounding macromolecules. In this study, articular cartilage (ART) and auricular cartilage (AUR) were digested with hyaluronidase to identify differences in GAG interactions, with a view to future studies on the role of GAG in tissue mechanics. Hyaluronidase is an enzyme which attacks the hyaluronan backbone where aggregated proteoglycans with sulphated GAG (sGAG) side chains are attached [1]
METHODS: Bovine ART and AUR plugs (Ø5 × 2 mm) were cored. Samples (n = 36) were divided into three groups per cartilage type: control, active, and blank. Samples were lyophilised overnight, and dry weights were measured. Active and blank groups were digested in 2000 U/mL hyaluronidase solution or blank phosphate buffer, respectively, at 37°C for 24 hours [2]. sGAG content of the digested plugs and control groups was determined with dimethyl methylene blue assay. Measured sGAG content was normalised with the dry weight measured before digestions. During analysis two outlying data points from ART blank group were removed. Significant differences (p<0.05) in normalised sGAG content of control and treated groups were compared using a one-way ANOVA test. Mean loss of sGAG content was calculated in active and blank groups, compared to controls.
RESULTS: ART showed higher sGAG content than AUR (ART:191 ± 36, AUR:120 ± 18 µg/mg of cartilage dry weight). Upon digestion active and blank groups showed significantly less sGAG content compared to the control groups (Fig.1).
Hyaluronidase resulted in over 100% sGAG loss in ART. However, it only removed 79% of sGAG from AUR indicating that hyaluronidase was not efficient in removing sGAG in AUR.
ART Control ART Blank ART Active AUR Control AUR Blank AUR Active
0 50 100 150 200 250
sGAG content (µg/mg of cartilage dry weight)
* *
*
#*
#Fig. 1: sGAG content of bovine ART and AUR cartilage samples after hyaluronidase digestion.,
*p < 0.05 compared to control, #p < 0.05 compared to blank.
DISCUSSION & CONCLUSIONS: Unlike the ART extracellular matrix, the AUR extracellular matrix has an elastin meshwork in addition to collagen fibres. Elastin has positive lysine amino groups, enabling it to form bonds with negatively charged sGAG [3]. In bovine and chick aortas, elastin fibres have specific ultrastructural attachment to the proteoglycans with heparan sulphate GAG chains [4]. This suggests that similar interactions could be expected in AUR as it is rich in elastin. However, the specific type of sGAG present in AUR extracellular matrix is not explicitly known. In future, identifying these specific sGAG will provide further insight into their mechanical role in different cartilage types.
REFERENCES: 1 P. Hoffman, K. Meyer and A.
Linker (1956) J Biol Chem 219:653-63. 2 L.
Nimeskern, L. Utomo, I. Lehtoviita, et al (2016) J Biomech 49: 344-52. 3 C. Fornieri, M. Baccarani- Contri, D. Quaglino Jr, et al (1987). J Cell Biol 105:1463-69. 4 B. Radhakrishnamurthy, H.A. Ruiz Jr, and G.S. Berenson (1977) J Biol Chem 252:
4831-41.
ACKNOWLEDGEMENTS: This study was supported by internal funding from the University of Melbourne.
WOUND IMAGE SEGMENTATION WITH DEEP NEURAL NETWORKS
H. R. Orefice1*, G. Scebba1*, L. Tüshaus1, S. Catanzaro2, M. Berli2, W. Karlen11 Mobile Health Systems Lab, Institute of Robotics and Intelligent Systems, Department of Health Sciences and Technology, ETH Zürich, Zurich, Switzerland
2 Balgrist University Hospital, Zurich, Switzerland
INTRODUCTION: Chronic wounds have a large negative impact on quality of life of diabetic patients and can lead to severe complications. In clinical practice, the wound assessment procedure is mainly based on visual examination. The visual wound inspection can be highly subjective and lacks accurate objective measurements to assess the wound healing conditions across multiple visits [1].
Image analysis techniques could provide a potential solution to overcome these limitations. In a digital image, the dimensions of the wound bed, which is one of the most important markers for the wound assessment, can be determined by performing semantic segmentation, a task where each pixel is classified as a foreground (wound bed) or background pixel.
In this work, we investigated the benefits of applying a deep learning (DL) neural network to segment wound images.
METHODS: We proposed the popular U-Net [2], a deep neural network successfully used to segment diverse biomedical images. The model architecture enables context understanding and precise localization due to its symmetric contracting and expanding path [2]. The U-Net was compared to a random segmentation baseline and to a handcrafted feature-based technique, which combines pyramid mean-shift filtering with color-based segmentation, inspired by the work of L. Wang et al [1].
We used a newly acquired wound image dataset, which contains 365 high-resolution images. The data was collected after institutional ethics approval (ETH 2017-N-07) from 49 patients recruited at the Balgrist University Hospital between March 2018 and May 2019. A research assistant manually produced the ground-truth masks for each wound image. The performance of the implemented models were evaluated by computing the global accuracy, as the number of correctly classified pixels over the total number of pixels, and the F1 score, which is the harmonic average between precision and recall.
RESULTS: The U-Net outperformed the handcrafted feature-based technique in the provided metrics (Table 1), showing a F1 score of 0.773, compared to 0.526 for the handcrafted technique and to 0.363 for the random segmentation baseline.
DISCUSSION & CONCLUSIONS: We presented a data-driven model able to accurately perform semantic segmentation on wound images. Despite the handcrafted technique having a more comprehensive pipeline, the results showed a much poorer performance compared to the U-Net. Image analysis techniques that rely on handcrafted image features often mismatch the high variability of wounds’ shape and color and, therefore, do not achieve optimal performance on wound segmentation tasks.
Contrarily, the proposed DL model accurately segmented test wound images (not used for training), indicating high generalization capability.
Qualitatively, the U-Net also achieved promising results on the test data (Fig.1), corroborating the fact that DL models can improve wound image analysis with objective measurements for an accurate assessment of the wound healing conditions.
REFERENCES: 1 L. Wang, et al. (2013) Smartphone-Based Wound Assessment System for Patients With Diabetes, IEEE Trans Biomed Eng, (2013), 2:477-88. 2 O. Ronneberger, et al. (2015) U- Net: Convolutional Networks for Biomedical Image Segmentation, MICCAI, (2015), pp 234-41.
ACKNOWLEDGEMENTS: This work is part of the HMZ SKINTEGRITY project. The authors would like to thank S. Weidmann for data collection, D. Chishima and M. Patwari for preparing the dataset. This work was supported by SNF grant 150640.
*Both authors contributed equally to this work.
Table 1. Performance evaluation of U-Net, handcrafted features model, and mean random segmentation over 10 runs.
Accuracy F1
Random 0.500 0.363
Handcrafted 0.643 0.526
U-Net 0.870 0.773
Fig. 1: (a) Wound image test samples, (b) U-Net predictions, (c) handcrafted predictions, (d) ground-truth masks.
(a) (b) (c) (d)
IMPLANTABLE SENSORICS SYSTEM FOR AUTOMATED HEALTH MONITORING
Fabian Rutschi, Martin Bertsch, René Mathys, Volker M. Koch HuCE - BME Lab, Bern University of Applied Sciences, Switzerland
INTRODUCTION: This work was to conceptualize an encapsulation for an automated sensorics system aiming to monitor health parameters of dairy cows, like body temperature and movement activity, for early recognition of sickness, estrus and birth. Dairy cows have an estrus every 21 days for only a few hours with economic disadvantages, if it is missed [1].
METHODS:
Usable materials were evaluated on a list of predefined requirements. The system has to be biocompatible for 5 years under the fat layer in the uterus of the dairy cow. The encapsulation has to resist the peristaltic forces as well as the uterine pressure (UP). For this, the encapsulation was modeled in Siemens NX 12 3D CAD and imported into COMSOL Multiphysics. There, a peak pressure of 1500 mmHg was simulated as an isostatic pressure [2].
The same model as for the pressure was used for the temperature conduction from the outside boundary wall towards the temperature sensor. In COMSOL and MATLAB, different combinations of materials were used to determine a viable variant for the temperature conduction.
RESULTS: In a first step, the best fitting materials for the encapsulation were evaluated through benefit analysis. High-density Polyethylene (HDPE) and Polyetheretherketone (PEEK) were selected. The calculated minimum wall thickness was larger for HDPE than for PEEK (Table 1). Hence, the maximal pressure of 12000 mmHg was obtained for pressure-injection molded PEEK.
Table 1. Wall thickness and pressure calculations.
Wall thickness Maximal Pressure
HDPE 0.48 mm 1500 mmHg
HDPE PIM 1.5 mm 9000 mmHg
PEEK 0.37 mm 1500 mmHg
PEEK PIM 1 mm 12000 mmHg
Using the above defined wall thicknesses for pressure injection molding, the temperature conduction, the resistance to force and the water vapor permeationwere calculated.
The temperature conduction through 4 mm of material was within the previously defined hysteresis of 10 minutes, if direct contact between the capsule and sensor can be obtained. While 0.2 MPa isostatic pressure is applied, the forces inside the capsule wall lie below 3 MPa for HDPE and below 4.5 MPa for PEEK (Fig. 1). Those values lie below the yield forces of those two materials. The permeation through the capsule wall to reach an equilibrium of water concentrations lies at 13.7 days for HDPE and at 29.7 days for PEEK.
Fig. 1: Left: CAD of the encapsulation. Right:
Force distribution over the encapsulation out of HDPE.
DISCUSSION & CONCLUSIONS: PEEK has advantages in mechanical stress resistance and diffusion, while HDPE is better in temperature conduction for a lower cost per piece. The time it takes for water to permeate the encapsulation wall is smaller than the implantation duration, hence additional measures will have to be considered, like coating or resin casting of the electronic components [3]. For those reasons, an encapsulation out of HDPE and at least coated electronics will be produced, extensively tested and used for a clinical in vivo study of the sensorics system at the end of the year.
REFERENCES:
[1] Dietrich O. (2012): Etablierung einer neuen Methode zur automatisierten Brunsterkennung beim Rind. Dissertation, LMU München. [2] S.
San Aik, “Bovine Uterine Pressure and the Response to Oxytocin as measured by a new Apparatus,” 1980. Master Thesis, Massey University. [3] D. Moss and M. Basic, “Pressure Vessel Design Manual (Fourth Edition).” Elsevier Inc. 2012.
SSBE 2019 Annual Meeting
Aug 27, 2019 - Campus Biotech, chemin de mines 9, Genève
Selected Posters
MODEL-BASED CHARACTERIZATION OF BRAIN TUMOR GROWTH
Daniel Abler 1,2, Russell C. Rockne 2, Philippe Büchler 11 ARTORG Center for Biomedical Engineering, University of Bern, Switzerland
2 Beckman Research Institute, City of Hope National Cancer Center, Duarte, CA, USA
INTRODUCTION: Glioblastoma (GBM) is the most frequent malignant brain tumor in adults. Its invasive growth causes healthy-tissue deformation, so-called tumor mass-effect. Elevated tumor mass- effect is associated to poor prognosis in GBM patients [1], and tumor-induced mechanical stresses lead to neuronal loss and neurological dysfunction [2]. Despite presentation with similar imaging volumes, GBM can present with varying degrees of mass-effect. Here, we present and evaluate a computational framework for investigating the relation between tumor growth characteristics, mass-effect and its manifestation on clinical imaging.
METHODS: Our mathematical model [3] captures three dominant aspects of macroscopic GBM growth: Invasive growth is modelled as a reaction- diffusion process with normalized cancer cell concentration 𝑐(𝑥, 𝑡), diffusion tensor 𝑫(𝑥), and logistic growth with proliferation rate ρ(x):
()
(*= 𝛻 ⋅ 𝑫 𝛻𝑐 + 𝜌𝑐 1 − 𝑐 (1) To simulate the tissue-displacing mass-effect of the growing tumor, the domain is modeled as elastic continuum in which the actual deformation 𝒖(𝑥, 𝑡) of a tissue element is given by the combination of growth-induced strains 𝜺5678*9 and strains associated with the elastic response of the tissue.
The model assumes a linear constitutive relation between mechanical stress and strain, isotropic material properties, and a linear coupling between tumor concentration and growth-induced strain with isotropic coupling strength λ:
𝜺5678*9 𝑐 = 𝝀 𝑐 = 𝜆 𝕀 𝑐 (2) Identification of growth parameters (𝐷, 𝜌, 𝜆) is framed as a PDE-constrained optimization problem in which we seek the set of parameters 𝒑𝒐𝒑𝒕 that minimize an objective functional of the form 𝐽 = 𝑐 𝒙, 𝑡C − 𝑐C∗(𝒙) + 𝒖 𝒙, 𝑡C − 𝒖C∗(𝒙) (3) where 𝑐C∗(𝒙) and 𝒖C∗(𝒙) are image-derived estimates of tumor cell concentration and tissue displacement fields at observation time point 𝑘.
Simulated tumor cell concentration 𝑐 𝒙, 𝑡C and tissue deformation 𝒖 𝒙, 𝑡C fields are constrained by the forward model. We use the adjoint method to efficiently find solutions to the optimization problem for multiple model parameters 𝒑. The mechanically-coupled reaction-diffusion model is implemented in the finite-element library FEniCS
[4] and uses dolfin-adjoint [5] for deriving and solving the adjoint equations.
RESULTS: Fig.1 illustrates forward simulation and parameter estimation in a 2D model of the human brain with separate subdomains for white matter (WM), grey matter (GM), and cerebrospinal fluid (CSF), as well as distinct isotropic growth 𝒑 =
𝐷FG, 𝐷HG, 𝜌FG, 𝜌HG, 𝜆 and mechanical tissue parameters. Using the objective functional Eq. (3), the parameters of the forward simulation were recovered by adjoint optimization. Evaluation on noisy 2D synthetic data shows relative errors in parameter estimates to be <15% for 80% of cases across a set of realistic parameter combinations.
Fig. 1: Left: Simulated tumor evolution from seed with reference growth parameters 𝒑. Right:
Predicted growth configuration based on 𝒑𝒐𝒑𝒕. DISCUSSION & CONCLUSIONS: Current developments focus on evaluating the performance of this approach in 3D, and on characterizing GBM growth phenotypes from clinical imaging.
REFERENCES: 1Steed et al (2017) Sci. Rep, 8(1):2827 2Seano et al (2019) Nat Biomed Eng 3:230-245. 3Abler et al (2018) In: Gefen, Weihs (eds) Computer Methods in Biomechanics and Biomedical Engineering, 57-64. 4Alnæs et al, The FEniCS Project Version 1.5. 5Farrell et al (2018) SIAM J. Sci. Comput, 35 (4):C369–C393.
ACKNOWLEDGEMENTS: The research leading to these results has received funding from the European Union’s Horizon 2020 research and innovation programme under the Marie Skłodowska-Curie grant agreement No 753878.
EFFICIENT PRIVACY-PRESERVING NEURAL NETWORK INFERENCE FOR HEART ARRHYTHMIA DETECTION
Philipp Chervet1, Alexandra-Mihaela Olteanu1, Juan Ramón Troncoso-Pastoriza2, David Froelicher2, Jérôme Van Zaen1, Ricard Delgado-Gonzalo1, and Jean-Pierre Hubaux2
1 Embedded Software Group, CSEM, Neuchâtel, Switzerland
2 LCA1, École Polytechnique Fédérale de Lausanne, Switzerland
INTRODUCTION: Becoming widely popular, the use of machine learning as a service poses an increasing risk to the privacy of those using it. We consider the case of a neural network (NN) used to detect heart arrhythmia from electrocardiogram (ECG) data. Considering the sensitivity of such data, our goal is to perform this task without revealing user input data to other parties involved and while hiding the model parameters from the user. Thus, we aim to protect the privacy of the users of the service and, at the same time, partially the confidentiality of the NN model (parameters are only visible to whoever runs the model). We use homomorphic encryption (HE) and secure two party computation techniques, and implement the service as a client-server application for privacy-preserving NN inference.
Fig. 1: Sensor – Gateway – Cloud ecosystem.
METHODS: Figure 1 illustrates our use case, a Sensor-Gateway-Cloud ecosystem, where the sensor collects the ECG data and sends it to the gateway. On the gateway, the client side of the application encrypts the data and, together with the server side application on a cloud, runs the privacy- preserving NN inference algorithm.
As the basis of our work, we used a previously developed convolutional NN called DeepCardio1. It takes single-lead ECG segments of 30 seconds duration, cut into 25 equally sized windows, and classifies them to detect arrhythmia. The NN mainly consists of 6 convolutional layers with non-linear activation functions (ReLU and max-pooling). At each layer, the number of channels is doubled, and the signal length is halved. To implement a privacy- preserving version, we extended the approach described in GAZELLE2: for the linear parts (convolution) we use HE and for the non-linear parts (activation functions) secure two party computation, specifically garbled circuits (GC).
RESULTS: To evaluate our approach, we run the client side on a laptop and the server side on a powerful workstation. Figure 2 shows the average execution time per layer for one window. The time spent for the HE operations is dominated by actual computation time. The exponential growth is due to the doubling of the channels. The time spent for GC operations is dominated by data transmission time, similar for all layers because the number of GC evaluations stays constant. The evaluation of the classification accuracy is work in progress.
Fig. 2: Execution time and distribution (between GC and HE operations) per layer for one window, averaged over 25 windows (one classification).
DISCUSSION & CONCLUSIONS: The results show an achieved latency of ~7 seconds for one window, resulting in ~175 seconds for one classification. However, we were able to identify the important bottlenecks and we propose approaches to improve it. The heavy computations performed using HE are mainly done on the server side, further parallelization of the server code is thus a good starting point. By optimizing the GC implementation and its parameters, an important gain in transmission time should be possible as well.
To conclude, our work shows that performing NN inference in a privacy-preserving way is possible and that there is promising potential to improve its current performance.
REFERENCES:
1J. Van Zaen et al., “Classification of cardiac arrhythmias from single lead ECG with a convolutional recurrent neural network”, BIOSTEC 2019.
2C. Juvekar et al., “Gazelle: A low latency framework for secure neural network inference,”
USENIX Security 2018.
REVIEW OF WEARABLE ELECTROENCEPHALOGRAM SYSTEMS
Ku-young Chung1, Laura Tüshaus1 and Walter Karlen11 Mobile Health Systems Lab, Institute of Robotics and Intelligent Systems, Department of Health Sciences and Technology, ETH Zurich, Zurich, Switzerland INTRODUCTION: The electroencephalogram
(EEG) is a non-invasive measurement of electric potentials created by brain activity. The application of EEG using large numbers of wet electrodes is limited to laboratory settings due to its obtrusiveness, lack of portability and usability.
Thus, efforts have been made to develop wearable EEG devices that promise high reliability and usability. However, developing a wearable EEG system is still a challenge. This is due to the trade- off between high signal quality against the disposition of electrodes (design) and easy preparation of the system (type of electrodes) [1].
We performed a systematic search of existing wearable EEG devices and prototypes to identify the ideal approaches to develop EEG device designs.
METHODS: We systematically investigated the designs and electrodes used in wearable EEG.
Twelve different search criteria or keywords (C1 - C12) were systematically stacked for search engines, where each criteria is an ensemble of representative keyword and their synonyms or closely related keywords. Then, we searched for articles, patents, and official web sites on search engines such as, Google (News, Scholar and Patents), and PubMed Central (PMC). In order to identify the most recent approaches, selection of results were limited to publication dates between 01.01.2018 and 28.01.2019.
Figure 1: Systematic search method to identify wearable EEG design and electrodes.
Finally, we excluded results based on defined exclusion criteria (Figure 1). Multiple use of same device in different studies was included. From the
obtained results, the designs and types of electrodes sorted into classes and quantified in percentages.
RESULTS: The most popular design for wearable EEG devices were headbands (16, 39.0%), then followed by a headset design (11, 26.8%) (Table 1). Dry electrodes were by far the most prevalent (26, 63.4%) (Table 2).
Table 1. Different types of wearable EEG designs.
Wearable EEG design Number of results (%) Headband 16 (39.0 %)
Headphones 2 (4.9 %)
Headset 11 (26.8 %)
Ear EEG 7 (17.1 %) Eyewear 5 (12.2 %)
Total 41
Table 2. Different types of electrodes used in wearable EEG devices.
Electrode types Number of results (%)
Dry 26 (63.4 %)
Wet 8 (19.5 %)
Both 5 (12.2 %)
Not Identified 2 (4.9 %)
Total 41
DISCUSSION & CONCLUSIONS: We have reviewed different approaches to develop wearable EEG systems that eases usability while minimizing obtrusiveness of electrodes. Our analysis showed that more studies develop wearable EEG by using a headband design with dry electrodes.
However, the challenges to design a wearable EEG systems still remain due to dry electrodes’s trade- off between the comfort level and performance.
Therefore, we have focused in past wearable designs on wet electrode designs [2]. Further research must be conducted in order to obtain better designs for easy to use, wearable EEG systems.
REFERENCES: 1 A. Casson (2018) Biomed Eng Lett 9:2. 2 M. Ferster, C. Lustenberger and W.
Karlen (2019) IEEE Sens Lett 3:5.
OBSTACLE DETECTION FOR THE VISUALLY IMPAIRED AND BLIND WITH A MULTISENSOR SMARTCANE
Gabriela Dudnik1, Olivier Debicki2, Suzanne Lesecq2, and Ricard Delgado-Gonzalo1
1Embedded Software Group, CSEM, Switzerland
2Laboratoire d'électronique et de la technologie de l'information, CEA-LETI-DACLE, France
INTRODUCTION: For most visually impaired and blind (VIB) people, the main barrier to improve their quality of life is the inability to navigate across a large variety of environments. This difficulty gets aggravated under complex situations containing a large quantity of obstacles. INSPEX1 (GA 730953 – ICT3-2016 SSI) is an EU initiative, under the H2020 framework, whose objective is to develop personal and portable multi-sensor systems for obstacle detection. The developed technology contains a combination of embedded sensors such as short and long range LiDARs, RF UWB radars, or ultrasonic radars. This technology is intended to be used on multiple industries including high-end autonomous cars, smartcanes for the VIB, safer human navigation in reduced visibility conditions (smoke, dust, fog, heavy rain/snow, darkness or combinations of these), and drone navigation.
In the following, we present the results of the proof of concept of the smartcane developed as a Medical Device Class I.
Figure 1. Full navigation system for the VIB. It includes a smarcane with multiple sensors, a gateway in the form of a smartphone, an audio interphase for aiding during the navigation, and potentially input from public infrastructure.
METHODS: The system adds a miniaturized mobile detection device to a white cane for the VIB community. For this application, an Augmented Reality Audio Interface is used to provide spatial 3D sound feedback using extra-auricular earphones.
This feedback takes into account the head attitude, tracked by an Attitude and Heading Reference System (AHRS) in the headset, to provide 3D spatial sound feedback of an obstacle’s real
direction and range. Context aware communications integrate the user with wider smart environments such as smart traffic lights, navigation beacons and ID tags associated with IoT objects (Figure 1). The user’s smartphone allows an integration with localization by means of mapping apps (GPS or other means). The project followed the V-model design approach2, executing all the procedures and fulfilling in parallel with the prototype, the required documentation for a Medical Device Class I.
RESULTS: To illustrate the capabilities, in Figure 2 (left) we show the physical setup for a detection test. The cane is placed on a swiping module and a pole is positioned from 4 m to 0.5 m by steps of 50 cm. For each position of the pole, we acquired sensor data over a full rotation of the platform. The distance information of the cane is shown in Figure 2 (right) where the obstacle is clearly visible on the profile.
Figure 2. (left) Physical testing setup with a pole as an obstacle.
(right) Distance profile obtained from the smart cane in which the obstacle is clearly visible.
DISCUSSION & CONCLUSIONS: The initial results for obstacle detection in the shape of the smartcane are very promising. Small objects such as poles are visible and feedback can be given to the VIB. A second demonstrator is being integrated and system verification tests will soon start. Further validation tests will be carried out by the end users.
REFERENCES:
1INSPEX project official website (http://www.inspex- ssi.eu/).
2K. Forsberg et al. "The Relationship of System Engineering to the Project Cycle", Annual Symposium of National Council on System Engineering, 1991: 57–65.
ACKNOWLEDGEMENTS: This project received funding from the EU’s Horizon 2020 research and innovation program under grant agreement No.
730953. This work was supported in part by the Swiss Secretariat for Education, Research and Innovation (SERI) under Grant 16.0136 730953.
