CC BY
3
Visnyk N'l'UU KP1 Seriia Radiolekhnika tiadioaparatobuduummia, "2021, Iss. 85, pp. 48—52
УДК 004.89
Method for Blink Detection in Single Channel of Invasive Electromyogram Signal
Bobrov A. L.1, Borysenko 0. M.\ Popov A. 0?
1 Institute of otolaryngology named after prof. O.S. Kolomijchenko NAMS of Ukraine
2 National Technical University of Ukraine "Igor Sikorsky Kyiv Polytechnic Institute"
E-mail: ncurol.ol-og&grnaU. com
Problem statement. Facial nerve damage is the cranial nervous system disorder often leading to facial muscle paralysis, which might be effectively restored using functional electrical stimulation of the fully or partially denervat.ed circular muscle of the eye to achieve muscle contraction to close the eyelids. To control the invasive stimulation system, the automated detection of the blink event in the intact eye is used as a trigger. To achieve this, the new approach to single channel invasive electromyogram (EMG) signal analysis is proposed. Materials and Methods. The combined time-spectral approach to blink detection consists of the two stages, starting from the thresholding of filtered EMG signals in the sliding window, which is followed by comparing the total spectral power in the Fourier domain to minimum and maximum thresholds. If both conditions are met, the EMG in the current window is considered to contain the blink event. In the experiment, the EMG data recorded from the one male adult healthy volunteer is used, the signal contained an acceptable amount of artefacts and was recognized as reflecting the usual EMG. The true positive rate (TPR), positive predictive value (PPV), False Discovery Rate (FDR), and False Negative Rate (FNR) is used as a performance metrics. Results. In the result of applying the proposed blink detection algorithm with 500 ms duration of the time window and 100 ms overlap, the following performance metrics are obtained: TPR = 93%, PPV = 63%, FDR = 7%, FNR = 37%. Impact. Acceptable true positive rate of blink detection suggests the method is promising for wider applications in the clinical settings and might be incorporated in the prototypes of implanted systems for facial muscle paralysis restoration using functional electrical stimulation for further development.
Key words: electromyography, facial nerve paralysis, blink detection DOI: 10.20535/RADAP. 2021.85.48-52
Introduction
Facial nerve (FN) damage is one of the most common disorders of the cranial nervous system, which often leads to facial muscle paralysis [1. 2]. The consequence is inability to close the eye. resulting in corneal atrophy and vision loss. This often cannot be effectively restored with surgical interventions, neither with mechanical springs, weight of the upper eyelid increase, muscle transpositions. In several studies [3 6]. the usability of functional electrical stimulation (FES) of the fully or partially denervated circular muscle of the eye (CMO) to achieve near-physiological muscle contraction is proposed [7.8] as an promising alternative possibility of closing the eyelids.
To drive the automated implantable system for CMO stimulation, the blink in the intact eye is used as a trigger. To that aim. the closed-loop system should continuously monitor the neural or myographic activity on the undamaged side, and in case of blink event it should be detected and the appropriate control signal is to be sent to trigger the CMO stimulation
for eye closure on the damaged side. To implement this operation, the electromyogram (EMG) signal from either the implanted or surface electrode should be measured and continuously processed in near-to-real time with blink detection algorithm.
The real-time algorithms for blink detection are not well developed at the moment. In the work [9] the simple peak detection is applied to the EMG signal, and the increasing of the static threshold serves as a trigger for the stimulation, which is very prone to false triggering due to noise and artefacts. No metrics of the algorithm performance are reported. In [10]. the multichannel surface EMG is analyzed with a custom amplitude-based or slope-based suprathreshold activity detector, applied to each channel independently. Despite promising, the direct application to the singlechannel EMG from implanted electrodes is not straightforward. Combination of the time-domain EMG amplitude analysis with the spectral-domain power analysis in the predefined frequency range is proposed in [11] using the constant thresholds, demonstrating the acceptable performance in the implantable system.
Parallel to the methods relied solely on the invasive EMG, there are other approaches to eye blink detection, such as based on combined analysis of electroencephalogram (EEG) and electrooculogram (EOG) for assistive communication tools [12]. for controlling the wheelchair [13]. and driver drowsiness detection [14]. Despite quite elaborated algorithms which are proposed therein, they do use surface signals for blink detection, which are not directly transferrable to invasive EMG due to different signal properties.
In our work, the combined time-spectral approach to blink events finding in single-channel invasive EMG is further developed, and the new method for blink detection is proposed and tested on recordings from healthy human subject.
1 Materials and Methods
1.1 Experimental data
In the experiment, the EMG data is recorded from the one male adult healthy volunteer (23 y.o.) in the Institute of Neurosurgery of Ukraine. The sampling frequency is 20 kHz: EMG spectral range from 0 to 300 Hz was selected for further analysis. In total, ^15 minutes of EMG was recorded, the signal contained an acceptable amount of artefacts and was recognised as reflecting the usual EMG. Before applying the proposed algorithm, EMG was filtered with 5th order Butterworth high-pass filter with 3 Hz cut-off frequency.
1.2 Method for blink detection
The combined time-spectral approach is grounded in the proposed algorithm for blink detection. The algorithm is presented in Fig. 1. First, the current time window i is selected in the filtered EMG signal Xj [n] of N samples length. Standard deviation of the EMG is calculated as
STD
1
N
N
¿ (x¿[n] - Xi[n]^
n=l
blinks), then the blink is absent in the current time window, and the next time window is being analyzed.
Since we adopt the subject-specific approach, the thresholds were experimentally tuned to reach the best performance. Three thresholds are to be empirically optimized in the laboratory settings to yield the highest performance of the implantable device for each particular subject. In clinical applications they should either be computed from population studies, or adaptively selected for each subject in the patient-specific settings.
where Xj [n] is the mean value of the EMG in the current window. Then it is compared to the threshold STDth.
If the standard deviation of the EMG in the current window is over the threshold, then the power spectral
total power is calculated and tested if it fits between two thresholds: minimum total power (SPth iow) and maximum (SPih high). If both conditions are satisfied, then the blink is detected. This event will trigger the system to generate stimulation pulses. If either standard deviation is lower than the threshold (due to absence of the excessive activity with respect to background), or the spectral power is larger (e.g. due to artefacts) or lower (due to less powerful events than
Fig. 1. Proposed algorithm for blink detection
To evaluate the performance of the proposed algorithm in the experiment, true positive rate (TPR) and positive predictive value (PPV) were used. True positive rate (Sensitivity) is the ratio of correctly identified blink to all blinks presented in the EMG. Positive predictive value is the ratio of correctly identified blinks to all signal fragments identified as blinks (both correctly and incorrectly). Based on these two metrics, False Discovery Rate (FDR) is calculated as 1-PPV, and False Negative Rate (FNR) is calculated as 1-TPR, to additionally describe the performance of the algorithm.
2 Results
The example of the EMG signal with marked blink events is presented in Fig. 2. A typical blink lasts approximately 300 ms, so 500 ms duration of the time window is selected with 100 ms overlap.
In the result of applying the proposed blink detection algorithm, the performance metrics are obtained for a range of the algorithm thresholds. After tuning the
50
tíobrov A. L., tíorvsonko O. M., Popov A. O.
-1-1-1-1-1-1-1-1-r
300 320 340 360 380 400 420 440 460 480 500
time, sec
Fig. 2. Example of the EMG signal with marked events of the blinks (stems with red squares) and results of the blink detection using the proposed algorithms (stems with green stars) [12]
algorithm parameters, the following best metrics are obtained (see Table 1).
Table 1 Blink detection performance metrics
Metric Value, %
Truc positivo rato (TPR) 93
Positivo prodictive valuó (PPV) 63
Falso discovory rato (FDR) 37
Falso negativo rato (FNR) 7
True positive rate is acceptable and false discovery rate of the erroneous blinks is low. which is promising for real-life applications and natural conditions for triggering the blinks synchronously for both sides of the face. At the same time, quite decent positive predictive value suggests that there is the room for improvement of the algorithm. First of all. more experimental data is required to understand the approach to selection of thresholds both in time and spectral domains. Also, the preprocessing of EMG signal will potentially improve the blink detection quality by removing the artefacts (such as head movement, chewing, swallowing, etc.)
Conclusions and Outlook
The new method for detecting the blinks in singlechannel EMG is proposed, based on the combined time and frequency domain thresholding approaches. Its application on the EMG collected from the human subject is presented, resulting in an acceptable true positive rate of detection, which is promising for wider applications in the clinical settings.
The method was tested in one single subject due to the limitations in the collection of clinical data, hence this work should be considered clS ct feasibility study, which prepares the road to wider studies with recruiting more participants. Nevertheless, current results could be considered as encouraging: the proposed
algorithm is able to detect the eye blinks in close-to-real settings with acceptable accuracy.
On the other hand, our current idea is the patient-specific settings, where the parameters of the algorithm are tuned manually by trials and errors for each user. The main reason for this is highly variable appearance of the signals from circular muscle of the eye. which depends both on the implanted electrode positioning and the anatomy of the face muscles. In favor of the current experiment setup with subject-specific algorithm parameter tuning speaks the fact that in implantable systems all parameters are being tuned for each user. Hence the future system should be fine-tuned for each subject, and the current findings are the example of its possibility. From the clinical perspective. such tuning of the parameters is acceptable, since in practice parameters of the implantable devices are selected for each user by the team of medical doctors and engineers.
In future, after the substantial database of the annotated signals is collected, we will work on the non-subject specific algorithms, if we can catch the common patterns in invasive EMG. including those based on the machine learning.
References
[1] Samii M.. Matthios C. (1994). Indication, technique and results of facial nerve reconstruction. Acl.a N eurochirurgi-ca, Vol. 130(1-4), pp. 1-25 139. DOl: 10.1007/b«)14055r2.
[2] Manni .1. (2012). Atlas of Surgery of the Facial Nerve. Chapter-18 Facial-Hypoglossal Nerve .lump Anastomosis for Reanimation of the Paralyzed Face. Jaypee Brothers Medical Publishers (P) Ltd2nd ed„ pp. 177 177. DOl: 10.5005/jp/books/11709_18.
[3] Tretiak LB. (2007). Prolonged electrical stimulation of peripheral nerves and plexus damage. Ukrainian Neurosurgical Journal, Vol. 2 (2007), pp. 58-61. DOl: 10.25305/unj.130686. [In Russian],
[4] Willand M. P. ("2015). Electrical stimulation enhances reinnervation after nerve injury. European .Journal of Translational Myology, Vol. "25. Iss. 4. DOl: 10.4081/ejtm.2015.5243.
[5] Bigard Л.-Х. et al. (1993). Ellects of surface electrostimulation on the structure and metabolic properties in monkey skeletal muscle. Medicine в Science in Sports &"Exercise, Vol. 25. Iss. 3. p. 355-362. DOl: 10.1249/00005768-199303000-00010.
[6] Oittins .1.. Martin R.. Sheldrick .1.. Roddy A.. Thean L. (1999). Electrical stimulation as a therapeutic option to improve eyelid function in chronic facial nerve disorders. Investigative ophthalmology в visual science, Vol. 40. Iss. 3. pp. 547-554.
[7] Mokrusch T.. Neundorfer B. (1994). Electrotherapy of permanently denervated muscle: long-term experience with a new method. European journal of physical medicine в rehabilitation, Vol. 4. Iss. 5. pp. 166-173.
[8] Salerno G., McClellan G. A.. Bleicher .1. N.. Stromberg B. V.. Cheng S.-C. (1991). Electrical Stimulation Treatment of Dog:s Denervated Orbicularis Oculi Muscle. Annals of Plastic Surgery, Vol. 26. Iss. 5. pp. 431 440. DOl: 10.1097/00000637-199105000-00004.
[9] Kalivaraprasad В.. Prasad V. M. D.. Harshavardhan L. (2021). Development of Blink Restoration Model for Facial Paralysis Detection. .Journal of Physics: Conference Series, 1804. 012175. DOl: 10.1088/1742-6596/1804/1/012175.
[10] Frigerio A.. Cavallari P.. Frigeni M.. Pedrocchi A.. Sarasola A.. Ferrante S. (2014). Surface Electromyographic Mapping of the Orbicularis Oculi Muscle for Real-Time Blink Detection. JAMA Facial Plastic Surgery, Vol. 16. Iss. 5. pp. 335 342. DOl: 10.1001/jamafacial.2014.283.
[11] .lia .1.. Yi X.. Wang M. Wang G.. Deng S.. Shen G. (2012). Л blink restoration system with contralateral EMG triggered stimulation and real-time software based artifact blanking. 2012 IEEE International Symposium on Circuits and Systems, pp. 1536-1539. DOl: 10.1109/iscas.2012.6271543.
[12] Yu Y.. Liu Y.. Yin E.. Jiang .1.. Zhou Z.. Hu D. (2019). An Asynchronous Hybrid Spelling Approach Based on EEC EOG Signals for Chinese Character Input. IEEE 'lYansactions on Neural Systems and Réhabilitation Engineering, Vol. 27. No. 6. pp. 1292-1302. doi: 10.1109/TNSRE.2019.2914916.
[131 Huang Q.. Zhang Z.. Yu Т.. He S.. Li Y. (2019). An EEG-/EOG-Based Hybrid Brain-Computer Interface: Application on Controlling an Integrated Wheelchair Robotic Arm System. Frontiers in neuroscience, 13: 1243. DOl: 10.3389"/fnins.2019.01243
[14] Baruaa S.. Ahmed M. U.. Ahlstromb C.. BegumaS. (2019). Automatic driver sleepiness detection using EEC. EOG and contextual information. Expert systems with applications, Vol. 115. pp. 121-135. DOl: 10.1016/j.eswa.2018.07.054
[15] Batulin D.. Popov A.. Bobrov A.. Tretiakova A. (2017). Eye blink detection for the implantable system for functional restoration of orbicularis oculi muscle. 2017 Signal Processing Symposium (SPSympo), pp. 1-4. doi: 10.1109/SPS.2017.8053650.
Метод детекцп моргания в однока-нальному сигнал! швазивноТ електро-мюграми
Бобров А. Л., Борисенко О. М., Попов А. О.
Пошкоджеппя лицьового перва - це розлад череппо! первово! системи. який часто призводить до парал!чу м'яз!в обличчя. який можпа ефективпо в1дповити за допомогою фупкцюпалыю! електричпо! стимуляцп пов-шстю або частково депервовапого кругового м'яза ока, щоб досягти скорочешш м'яз!в для закрпття повши. Для управлпшя системою пгоазивпо! стимуляцп в яко-ст! пускового мехашзму використовуеться автоматично ВИЯВЛС1ШЯ подп моргаппя в пеушкоджепому оц1. Для цього в робот! запропоповапо повий шднд до апаль зу одпокапалыюго сигналу пгоазивпо! електромюграми (ЕМГ).
Комбшовапий часово-спектралышй шднд до вияв-леппя моргаппя складаеться з двох еташв. почипаючи з пор1впяппя з порогом в1дф1льтровапих ЕМГ-сигпал1в у ковзпому в!кш. за яким сл!дуе пор1впяппя загалыго! спектрально! потужпост в облает! Фур'е з мшмалышм та максималышм порогами. Якщо викопуються дв! умови. вважаеться. що ЕМГ у поточному в!кш м!стить подпо моргаппя.
В експериммт використовуються даш ЕМГ. записан! в!д одного дорослого здорового чоловша-добровольця: сигнал м!стив прийнятиу шльшеть ар-тефактав i визпапий таким, що в!дображае звичайпу ЕМГ. Чутлшмсть (TPR). позитивпе прогпозпе зпачеп-пя (PPV), коефщ!епт помилкового виявлеппя (FDR) та хибпо-пегативпий коефкцепт (FNR) використовуються як показпики ефективпоста.
В результат! застосуваппя запропоповапого алгоритму виявлмшя моргаппя з тривал!стю часового в!кпа 500 мс та перекриттям 100 мс отримаш так! показпи-ки ефективпостк TPR = 93%, PPV = 63%, FDR = 7%, FNR = 37%.
Прийпятпа чутлшмсть виявлешш моргаппя св!дчить про те. що метод е перспективпим для бглын широкого застосуваппя в клипчпих умовах i може бути застосо-вапий у прототипах 1мплаптовапих систем вщповлеппя парал!чу м'яз!в обличчя за допомогою фупкгцопально! електрпчпо! стимуляцп для !х подальшого розвптку.
Ключоег слова: електромюграф!я. парал!ч лицьового нерву, детектуваппя моргашш
Метод детекции моргания в однока-нальном сигнале инвазивной электро-миограммы
Бобров А. Л., Борисенко О. П., Попов А. А.
Повреждение лицевого перва - это расстройство черепной нервной системы, которое часто приводит к параличу мышц лица, который можно эффективно восстановить с помощью функциональной электрической стимуляции полностью или частично депервировашюй круговой мышцы глаза, чтобы добиться сокращения мышц для закрытия века. Для управления системой ипвазивпой стимуляции в качестве пускового механизма используется автоматическое обнаружение события моргания в неповрежденном глазу. Для этого в работе предложен новый подход к анализу одпокапалыюго сигнала ипвазивпой электромиограммы (ЭМГ).
52
Бобров А. Л., Борисенко О. М., Попов А. О.
Комбинированный временно-спектральный подход к выявлению моргания состоит из двух этапов: сравнение с порогом отфильтрованных ЭМГ-сигналов в скользящем окне, за которым следует сравнение общей спектральной мощности в области Фурье с минимальным и максимальным порогами. Если выполняются три условия, считается, что ЭМГ в текущем окне содержит событие моргания.
В эксперименте используются данные ЭМГ, записанные от взрослого здорового мужчины-добровольца; сигнал содержал приемлемое количество артефактов и признан отражающим обычную ЭМГ. Чувствительность (TPR), положительное прогнозное значение (PPV), коэффициент ложного обнаружения (FDR) и ложно-отрицательный коэффициент (FNR) используются как показатели эффективности.
В результате применения предложенного алгоритма обнаружения моргания с продолжительностью временного окна 500 мс и перекрытием 100 мс получены следующие показатели эффективности: ТРИ, = 93%, РР\' = 63%, РБ11 = 7%, К.ХН 37'/.
Приемлемая чувствительность обнаружения моргания свидетельствует о том, что метод является перспективным для более широкого применения в клинических условиях и может быть применен в прототипах имплантированных систем восстановления паралича мышц лица с помощью функциональной электрической стимуляции для их дальнейшего развития.
Ключевые слова: электромиография, паралич лицевого нерва, детектирование моргания
