Extraction of volitional EMG during dynamic hybrid contractions through wireless EMG-sensors

Politecnico di Milano Scuola di Ingegneria Industriale e dell’Informazione Corso di Laurea Magistrale in Ingegneria Biomedica Dipartimento di Elettronica, Informazione e Bioingegneria Acknowledgements   Abstract Introduction The Central Nervous System (CNS) is the hierarchical structure performing controls on the motor system. The physiological structure controlling the whole activities is the brain, which has the crucial role of planning, generating and controlling the signals directed to the muscles, which activate the effectors, namely the limbs. The effectiveness of the cerebral system can be damaged by Spinal Cord Injury (SCI), or by stroke. This work is placed in the framework of technological solutions for the rehabilitation of stroke survivors. Stroke has a high risk of death and an even greater risk of disability. The most common impairment after stroke is upper limb paresis, which is found in 77% of stroke survivors and regained in the latter six months only in half of stroke survivors.[1] Of course, the reduced mobility has a huge impact on patients independence in performing daily life activities (ADL). This makes the recovery of arm movements one of the main goals of stroke rehabilitation. Functional Electrical Stimulation (FES) performed on post-stroke patients, demonstrated its efficacy on both upper and lower limb. The success in enhancing motor relearning is confirmed if combined with goal-oriented repetitive therapy [2]. Indeed, combining FES with volitional effort has proven to maximize neural plasticity, which is the reason behind the goal of this work. The main aim of this work was in fact the development of a real-time algorithm for the estimation of volitional EMG during hybrid muscle contractions that can be therefore used in an EMG controlled FES system integrating FES and volitional effort. The EMG acquisition was carried out through an EMG wireless sensor [3] that additionally integrates inertial sensors to allow a contemporaneous estimation of the EMG signal and kinematic measures. This sensor is a versatile and low-cost solution, allowing the integration with a variety of devices. Methods This work involved the estimation of volitional EMG, during hybrid muscle contractions, of two muscles: the biceps brachialis and the medial deltoid. This aim was achievable using the following hardware instrumentation:  EMG-wireless sensor [3];  ReMove3™ stimulator from HASOMED; Torecord the EMG signals, five Ag/AgCl pre-gelled self-adhesive electrodes were used. Two electrodes were used for the recording of each muscle and one was the reference electrode. To deliver the stimulation, surface self-adhesive rectangular 3x5 cm electrodes, were used, two for each muscle. The experimental set up is shown in Figure 1 panel (a). The EMG signal recordings were performed with the EMG wireless sensor [3], shown in figure 1 panel (b). This sensor allows to contemporarily record the EMG signal from two different channels. In order to reach the aim of the thesis, the work was organised in three phases.  Set-up parameters definition;  Development and validation of the EMG filter algorithm: artefact identification and volitional estimation;  Real-time algorithm implementation. The first phase consisted in the identification of the optimal parameters of the EMG sensors in terms of sampling frequency and transmission frequency. These parameters were experimentally identified in order to minimize the number of missing samples due to the Bluetooth transmission. The protocol carried out during this step of the work was composed of three phases, repeated once for biceps brachialis and once for medial deltoid: 1. Three voluntary contractions without weight; 2. Three voluntary contractions holding a 1 kg weight; 3. Three voluntary contractions holding a 3 kg weight; Six healthy subjects (age 23-28) were involved in the trials: subjects were asked to perform some volitional contractions and the EMG signals of two channels were simultaneously recorded. The protocol was repeated in two different electrodes placement: EMG electrodes placed along the muscle fibre according to the SENIAM guidelines and EMG electrodes placed perpendicular to the muscle fibres in order to minimize the stimulation artefact [4]. Figure 2 shows the electrodes placement, while Table 1 reports the different combinations which were analysed for the configuration parameters. The first column is the sampling frequency of the sensor, the second column is the data transmission frequency. All the different configurations were compared in terms of percentage of missed samples, the coefficient of variation of samples and root mean square value of the signals recorded. The core of the work was the development of an algorithm to estimate the volitional component during hybrid muscle contractions in real-time. This estimate required the following steps:  Identification of the stimulation artefact  Removal of the stimulation artefact  Identification of the volitional component for each inter-pulse stimulus For the identification of the stimulation artefact, two different methods were initially compared in off-line, which are shown in diagrams in Figure 3.:  Wavelet-based algorithm [5], using continuous wavelet transform (CWT);  Differential function-based algorithm. After the artefact identification, for each inter-pulse period, a window of the EMG signal is selected. To assure that the windowed EMG does not include the stimulation artefact, 25 samples after the artefact position are discarded and the following 60 samples are selected. The estimation of the volitional EMG is performed using an adaptive filter [6], applied to the windowed EMG selected in the previous step. Figure 3 shows two flowcharts describing the steps of the estimation performed in the CWT based algorithm (a) and in the differential function-based algorithm (b). Some experimental trials were carried out in order to validate the filtering algorithm. The protocol involved the stimulation and simultaneous EMG recordings from two channels (1: biceps brachialis muscle; 2: medial deltoid) and can be summarized in six phases:  Phase 0) rest;  Phase 1) subject is asked to perform three voluntary contractions without weight and three with 1 kg weight, while no stimulation was delivered;  Phase 2) subject is asked to be relaxed while three different levels of stimulation current are delivered: 50% of maximum stimulation current, 100% of maximum stimulation current and 50% of maximum stimulation current delivered on both channels;  Phase 3) subject is asked to perform the same voluntary contractions as in phase 1 while 50% of maximum stimulation current is delivered;  Phase 4) subject is asked to perform the same voluntary contractions as in phase 1 while 100% of maximum stimulation current is delivered;  Phase 5) subject is asked to perform the same voluntary contractions as in phase 1 while 50% of maximum stimulation current is delivered on both channels. Subjects were asked to repeat the protocol once performing biceps brachialis muscle contractions and once performing medial deltoid muscle contractions. EMG signals were analysed with the algorithms described above with a double aim: 1. Identify the simplest algorithm to accurately identify the stimulation artefacts 2. Compare the volitional EMG estimate in the different condition of stimulation proposed in the experimental protocol. To reach the first aim the results, in terms of identified stimulation artefacts, of both algorithms developed were compared. While to reach the second aim, the volitional EMG estimates in the different phases of the protocol were compared using the Friedman test. To perform the Friedman test data was organised in five subsets: four subsets relative to each different phase of the protocol involving volitional contribution with the corresponding rest condition and one subset composed of all the rest conditions for each phase. For each of these subsets, a separate Friedman test and post hoc analysis was performed. The ultimate phase of the work was to implement the chosen algorithm in real-time in Simulink, allowing the estimation of the volitional EMG from hybrid contraction to be performed on-line. Results This thesis involved different steps of data acquisition, therefore the results of the work can be divided into different sections. The first outcome of this work was arising from the selection of the best configuration for set-up parameters. The trials to identify the configuration parameters of the EMG sensors found out that the best combination was represented by a sampling frequency of 2 kHz and a transmission frequency of 20 Hz, which was the configuration characterized by the lowest percentage of missed samples (0.1%), lowest coefficient of variation for the samples acquired at transmission frequency and highest RMS value of the volitional estimate. In the rest of the work, these parameters were used, and the stimulation frequency was set equal to the transmission frequency. In Figure 4, three different levels of volitional EMG can be clearly identified, the signal shown was acquired with the chosen configuration of set up parameters. The second fundamental result regards the comparison among the two algorithms for stimulation artefact detection and the estimate of the volitional EMG signal. This analysis demonstrated that both algorithms were successful in identifying the stimulation artefact and allow a correct volitional EMG estimation. The two algorithms were compared in terms of false positive and false negative, which were found to be zero in both cases. Given this result the choice was to implement the less complex algorithm in terms of the computational burden. The next step was to perform statistical analysis on the EMG signal of hybrid muscle contractions, extracted with the algorithm implemented. Figure 5 and figure 6, report the results of the Friedman test, where each histogram represents the median value, the vertical bar bounds the 25°and 75° percentiles and the ** refer to p<0.001, * to p<0.05. Conditions are codified by colours: blue for rest, green for volitional level 1 (no weight) and red for volitional level 2 (1 kg weightlifting). In figure 5 below it is possible to see that in the first phase of the protocol, when no stimulation is performed, all the phases are significantly different from each other as expected(p-value<0.001). However, in the phases in which the stimulation was delivered, the two volitional levels are not identified as significantly different (p-value<0.05). As ultimate validation, the algorithm implemented on-line was compared to the offline implementation in terms of artefact identification and root mean square error (RMSE) which are reported in Table 2 below. The volitional EMG estimated with the on-line and the off-line algorithm are reported in Figure 7. It is clearly visible in figure 7 that the on-line and the offline logics give very similar results, which means that the on-line logic implementation is valid. Conclusions and future works This thesis was aimed at implementing an online filtering methodology for the estimation of volitional EMG during hybrid muscle contractions, which was validated performing experimental trials involving six healthy subjects. The next steps of this work could be the validation of the algorithm on EMG signals acquired from stroke patients during hybrid muscle contractions and the development of a closed-loop EMG-controlled FES system based on the online estimate of the volitional component. This FES system might exploit also the inertial sensors included in the EMG sensors to estimate the arm kinematics.  Sommario Introduzione Il Sistema Nervoso Centrale (SNC) è la struttura gerarchica che esegue i controlli sul sistema motorio. La struttura fisiologica che controlla l'intera attività è il cervello, che ha un ruolo cruciale di pianificazione, controllo e generazione dei segnali diretti ai muscoli, che attivano gli effettori, ovvero gli arti. L'efficacia del sistema cerebrale può essere danneggiata da lesioni al midollo spinale (SCI), o causata da ictus. Questo lavoro si colloca nell'ambito delle soluzioni per la riabilitazione di pazienti post-ictus. L'ictus è caratterizzato da un alto rischio di morte e un rischio di disabilità ancora maggiore. Il danno più comune dopo l'ictus è la paresi degli arti superiori, che si riscontra nel 77% dei sopravvissuti e viene recuperata negli ultimi sei mesi solo nella metà dei pazienti[1]. Naturalmente, la mobilità ridotta ha un enorme impatto sull'indipendenza dei pazienti nello svolgimento delle attività quotidiane (ADL). Questo rende il recupero dei movimenti degli arti uno dei principali obiettivi della riabilitazione post-ictus. La Stimolazione Elettrica Funzionale (FES) eseguita su pazienti post-ictus, ha dimostrato la sua efficacia sia per gli arti superiori che inferiori. Il successo nel migliorare il riapprendimento motorio è confermato se combinato con una terapia ripetuta e mirata al raggiungimento di obiettivi [2]. Infatti, la combinazione di FES e sforzo volontario ha dimostrato di massimizzare la plasticità neurale. Lo scopo principale di questo lavoro è stato infatti lo sviluppo di un algoritmo in tempo reale per la stima dell'EMG volontario durante le contrazioni muscolari ibride, che possa essere utilizzato in un sistema controllato da EMG integrando FES e sforzo volontario. L'acquisizione dell'EMG è stata effettuata mediante sensore wireless EMG[3] che integra inoltre, sensori inerziali per consentire una stima contemporanea del segnale EMG e misure cinematiche. Questo sensore è una soluzione versatile e a basso costo, che consente l'integrazione con diversi dispositivi. Metodi Questo lavoro ha comportato la stima dell'EMG volontario, durante le contrazioni muscolari ibride, di due muscoli: il bicipite brachiale e il deltoide mediale. Questo obiettivo è stato raggiunto utilizzando la seguente strumentazione hardware:  Sensore EMG-wireless [3];  Stimolatore HASOMED ReMove3™; Per l'acquisizione del segnale EMG sono stati utilizzati cinque elettrodi autoadesivi con gel Ag/AgCl. Per la registrazione di ogni muscolo sono stati utilizzati due elettrodi il quinto è l'elettrodo di riferimento. Per la stimolazione, sono stati utilizzati elettrodi autoadesivi di superficie rettangolare 3x5 cm, due per ogni muscolo. Il set up sperimentale è mostrato in Figura 1 pannello (a). La registrazione del segnale EMG è stata eseguita con il sensore wireless EMG [3], rappresentato in figura 1 panello (b). Questo sensore permette di registrare contemporaneamente il segnale EMG da due canali differenti. I due canali sono rappresentati in Figura 1, i due elettrodi rossi e i due elettrodi neri in alto, mentre l'elettrodo nero al centro è l'elettrodo di riferimento. In order to reach the aim of the thesis, the work was organised in three phases.  Definizione dei parametri di set-up;  Sviluppo dell’algoritmo per individuazione dell’artefatto e stima del segnale volontario;  Implementazione real-time. La prima fase consiste nell'identificazione dei parametri ottimali in termini di frequenza di campionamento e di frequenza di trasmissione. Questi parametri sono stati identificati sperimentalmente al fine di ridurre al minimo il numero di campioni persi a causa della trasmissione Bluetooth. Il protocollo realizzato in questa fase del lavoro, si compone di tre fasi: 1. Tre contrazioni volontarie senza peso; 2. Tre contrazioni volontarie con peso da 1 kg ; 3. Tre contrazioni volontarie con peso da 3 kg ; Sei soggetti sani (23-28 anni) sono stati coinvolti nelle prove sperimentali: ai soggetti è stato chiesto di eseguire alcune contrazioni volontarie e sono stati registrati contemporaneamente i segnali EMG di due canali. Il protocollo è stato eseguito in due diverse configurazioni per quanto riguarda il posizionamento degli elettrodi: Elettrodi EMG posizionati lungo la fibra muscolare secondo le linee guida SENIAM ed elettrodi EMG perpendicolari alle fibre muscolari per minimizzare l'artefatto di stimolazione [4]. La Figura 2 mostra il posizionamento degli elettrodi, mentre la Tabella 1 riporta le diverse combinazioni che sono state analizzate. Le diverse configurazioni sono state confrontate in termini di numero di campioni persi, coefficiente di variazione dei campioni acquisiti e valore quadratico medio dei segnali acquisiti, al fine di scegliere la configurazione migliore. Il nocciolo del lavoro è stato lo sviluppo di un algoritmo per stimare la componente volontaria durante le contrazioni muscolari ibride in tempo reale. Questa stima è stata svolta attraverso le seguenti fasi: 1. Identificazione dell'artefatto di stimolazione; 2. Rimozione dell'artefatto di stimolazione; 3. Identificazione della componente volontaria per ogni periodo interstimolo. Per l'identificazione dell'artefatto di stimolazione, due diversi metodi sono stati inizialmente confrontati off-line, i due metodi sono mostrati nei diagrammi in Figura 3:  Algoritmo basato su trasformata wavelet[5], che implementa la trasformata wavelet continua (CWT);  Differential function based algorithm. Dopo l'identificazione dell'artefatto, per ogni periodo interstimolo, viene selezionata una finestra del segnale EMG. Per garantire che l'EMG finestrato non includa l'artefatto da stimolazione, vengono scartati i 25 campioni successivi all'artefatto e vengono selezionati i successivi 60 campioni. La stima dell'EMG volontario viene eseguita utilizzando un filtro adattivo[6], applicato all'EMG finestrato selezionato nella fase precedente. Il protocollo sviluppato per l'acquisizione dell'EMG può essere riassunto in sei fasi:  Fase 0) riposo;  Fase 1) al soggetto viene chiesto di effettuare tre contrazioni senza peso e tre contrazioni con peso da 1 kg, senza stimolazione;  Fase 2) viene chiesto di essere rilassato mentre vengono forniti tre diversi livelli di stimolazione: 50% della massima corrente definita, 100% della massima corrente e 50% della massima corrente erogata su entrambi i canali;  Fase 3) al soggetto viene chiesto di effettuare tre contrazioni senza peso e tre contrazioni con peso da 1 kg, con stimolazione contemporanea ad una corrente pari al 50% del valore massimo definito;  Fase 4) al soggetto viene chiesto di effettuare tre contrazioni senza peso e tre contrazioni con peso da 1 kg, con stimolazione contemporanea ad una corrente pari al 100% del valore massimo definito;  Fase 5) al soggetto viene chiesto di effettuare tre contrazioni senza peso e tre contrazioni con peso da 1 kg, con stimolazione contemporanea ad una corrente pari al 50% del valore massimo definito effettuata con entrambi i canali; Al soggetto è stato chiesto di ripetere il protocollo una volta eseguendo contrazioni muscolari con il bicipite brachiale e in seguito eseguendo contrazioni muscolari con il deltoide mediale. I segnali EMG sono stati analizzati con gli algoritmi sopra descritti con un duplice scopo: 1. Determinare l'algoritmo più semplice per identificare accuratamente gli artefatti da stimolazione 2. Confrontare la stima dell'EMG volontario nelle diverse condizioni di stimolazione effettuate nel protocollo sperimentale. Per raggiungere il primo obiettivo sono stati confrontati i risultati, in termini di artefatti da stimolazione identificati, di entrambi gli algoritmi. Mentre per raggiungere il secondo obiettivo, la stima dell'EMG volontario nelle diverse fasi del protocollo è stata confrontata con il test di Friedman. Per eseguire il test Friedman i dati sono stati organizzati in cinque subset: quattro subset uno per ogni diversa fase del protocollo con la relativa condizione di riposo e un subset composto da tutte le condizioni di riposo per ogni fase. Per ognuno di questi subset viene eseguito un test di Friedman per ogni fase e analisi post hoc. La fase finale del lavoro è stata quella di implementare l'algoritmo in tempo reale in Simulink. Risultati Il lavoro ha comportato diverse fasi di acquisizione dei dati, pertanto i principali risultati di questo lavoro possono essere suddivisi in diverse sezioni. Il primo risultato di questo lavoro è derivato dalla scelta della configurazione migliore per i parametri del sistema. Le prove per identificare i parametri di configurazione dei sensori EMG hanno evidenziato che la migliore combinazione è quella composta da una frequenza di campionamento di 2 kHz e una frequenza di trasmissione di 20 Hz, che è la configurazione caratterizzata dalla più bassa percentuale di campioni persi (0,1%). Nella Figura 4 si possono identificare chiaramente tre diversi livelli di EMG volontario, il segnale mostrato è stato acquisito con la configurazione scelta dei parametri di set up. Il secondo risultato fondamentale riguarda il confronto tra i due algoritmi. Questa analisi ha dimostrato che entrambi sono riescono nell'identificazione dell'artefatto da stimolazione e permettono una corretta stima dell'EMG volontario Il passo successivo è stato quello di effettuare analisi statistiche sul segnale EMG volontario estratto con l'algoritmo implementato. Le figure 5 e 6, riportano i risultati del test di Friedman, dove ogni istogramma rappresenta il valore mediano, la barra verticale il 25°-75° percentile e il ** si riferisce a p<0,001, * p<0,05. Le diverse condizioni sono codificate dai colori: blu per il riposo, verde per il livello volontario 1 (nessun peso) e rosso per il livello volontario 2 (sollevamento peso di 1 kg). Per l'implementazione in tempo reale è stato scelto l'algoritmo basato sul calcolo della funzione derivata, caratterizzata da minore costo computazionale. L'algoritmo implementato in tempo reale è stato confrontato con l'implementazione offline in termini di artefatti identificati ed errore quadratico medio (RMSE) tra l’EMG volontario stimato online e quello stimato offline. I risultati ottenuti in termini di RMSE sono riportati in Tabella 2. La prima colonna è relativa al canale 1, mentre la seconda colonna riporta i valori relativi al canale 2. L’ultima riga invece riporta il valore medio dell’RMSE di tutti i soggetti. Mentre la figura 7 sottostante riporta i risultati in termini di EMG volontario stimato. E' chiaramente visibile nella figura 7 che le logiche on-line e offline danno risultati molto simili, il che significa che l'implementazione della logica on-line fornisce una stima valida dell’EMG volontario. Conclusioni e sviluppi futuri Lo scopo del presente lavoro di tesi è stato l'implementazione di una metodologia di filtraggio online per la stima dell'EMG volontario delle contrazioni muscolari ibride, validata eseguendo prove sperimentali su sei soggetti sani. I passi successivi di questo lavoro potrebbero includere innanzitutto la validazione dell'algoritmo sui segnali EMG acquisiti da pazienti post-ictus durante la contrazione muscolare ibrida e lo sviluppo di un sistema FES controllato da EMG ad anello chiuso basato sulla stima online della componente volontaria. Questo sistema FES potrebbe sfruttare anche i sensori inerziali inclusi nei sensori EMG per stimare la cinematica del braccio.

Introduction The Central Nervous System (CNS) is the hierarchical structure performing controls on the motor system. The physiological structure controlling the whole activities is the brain, which has the crucial role of planning, generating and controlling the signals directed to the muscles, which activate the effectors, namely the limbs. The effectiveness of the cerebral system can be damaged by Spinal Cord Injury (SCI), or by stroke. This work is placed in the framework of technological solutions for the rehabilitation of stroke survivors. Stroke has a high risk of death and an even greater risk of disability. The most common impairment after stroke is upper limb paresis, which is found in 77% of stroke survivors and regained in the latter six months only in half of stroke survivors.[1] Of course, the reduced mobility has a huge impact on patients independence in performing daily life activities (ADL). This makes the recovery of arm movements one of the main goals of stroke rehabilitation. Functional Electrical Stimulation (FES) performed on post-stroke patients, demonstrated its efficacy on both upper and lower limb. The success in enhancing motor relearning is confirmed if combined with goal-oriented repetitive therapy [2]. Indeed, combining FES with volitional effort has proven to maximize neural plasticity, which is the reason behind the goal of this work. The main aim of this work was in fact the development of a real-time algorithm for the estimation of volitional EMG during hybrid muscle contractions that can be therefore used in an EMG controlled FES system integrating FES and volitional effort. The EMG acquisition was carried out through an EMG wireless sensor [3] that additionally integrates inertial sensors to allow a contemporaneous estimation of the EMG signal and kinematic measures. This sensor is a versatile and low-cost solution, allowing the integration with a variety of devices. Methods This work involved the estimation of volitional EMG, during hybrid muscle contractions, of two muscles: the biceps brachialis and the medial deltoid. This aim was achievable using the following hardware instrumentation:  EMG-wireless sensor [3];  ReMove3™ stimulator from HASOMED; Torecord the EMG signals, five Ag/AgCl pre-gelled self-adhesive electrodes were used. Two electrodes were used for the recording of each muscle and one was the reference electrode. To deliver the stimulation, surface self-adhesive rectangular 3x5 cm electrodes, were used, two for each muscle. The experimental set up is shown in Figure 1 panel (a). The EMG signal recordings were performed with the EMG wireless sensor [3], shown in figure 1 panel (b). This sensor allows to contemporarily record the EMG signal from two different channels. In order to reach the aim of the thesis, the work was organised in three phases.  Set-up parameters definition;  Development and validation of the EMG filter algorithm: artefact identification and volitional estimation;  Real-time algorithm implementation. The first phase consisted in the identification of the optimal parameters of the EMG sensors in terms of sampling frequency and transmission frequency. These parameters were experimentally identified in order to minimize the number of missing samples due to the Bluetooth transmission. The protocol carried out during this step of the work was composed of three phases, repeated once for biceps brachialis and once for medial deltoid: 1. Three voluntary contractions without weight; 2. Three voluntary contractions holding a 1 kg weight; 3. Three voluntary contractions holding a 3 kg weight; Six healthy subjects (age 23-28) were involved in the trials: subjects were asked to perform some volitional contractions and the EMG signals of two channels were simultaneously recorded. The protocol was repeated in two different electrodes placement: EMG electrodes placed along the muscle fibre according to the SENIAM guidelines and EMG electrodes placed perpendicular to the muscle fibres in order to minimize the stimulation artefact [4]. Figure 2 shows the electrodes placement, while Table 1 reports the different combinations which were analysed for the configuration parameters. The first column is the sampling frequency of the sensor, the second column is the data transmission frequency. All the different configurations were compared in terms of percentage of missed samples, the coefficient of variation of samples and root mean square value of the signals recorded. The core of the work was the development of an algorithm to estimate the volitional component during hybrid muscle contractions in real-time. This estimate required the following steps:  Identification of the stimulation artefact  Removal of the stimulation artefact  Identification of the volitional component for each inter-pulse stimulus For the identification of the stimulation artefact, two different methods were initially compared in off-line, which are shown in diagrams in Figure 3.:  Wavelet-based algorithm [5], using continuous wavelet transform (CWT);  Differential function-based algorithm. After the artefact identification, for each inter-pulse period, a window of the EMG signal is selected. To assure that the windowed EMG does not include the stimulation artefact, 25 samples after the artefact position are discarded and the following 60 samples are selected. The estimation of the volitional EMG is performed using an adaptive filter [6], applied to the windowed EMG selected in the previous step. Figure 3 shows two flowcharts describing the steps of the estimation performed in the CWT based algorithm (a) and in the differential function-based algorithm (b). Some experimental trials were carried out in order to validate the filtering algorithm. The protocol involved the stimulation and simultaneous EMG recordings from two channels (1: biceps brachialis muscle; 2: medial deltoid) and can be summarized in six phases:  Phase 0) rest;  Phase 1) subject is asked to perform three voluntary contractions without weight and three with 1 kg weight, while no stimulation was delivered;  Phase 2) subject is asked to be relaxed while three different levels of stimulation current are delivered: 50% of maximum stimulation current, 100% of maximum stimulation current and 50% of maximum stimulation current delivered on both channels;  Phase 3) subject is asked to perform the same voluntary contractions as in phase 1 while 50% of maximum stimulation current is delivered;  Phase 4) subject is asked to perform the same voluntary contractions as in phase 1 while 100% of maximum stimulation current is delivered;  Phase 5) subject is asked to perform the same voluntary contractions as in phase 1 while 50% of maximum stimulation current is delivered on both channels. Subjects were asked to repeat the protocol once performing biceps brachialis muscle contractions and once performing medial deltoid muscle contractions. EMG signals were analysed with the algorithms described above with a double aim: 1. Identify the simplest algorithm to accurately identify the stimulation artefacts 2. Compare the volitional EMG estimate in the different condition of stimulation proposed in the experimental protocol. To reach the first aim the results, in terms of identified stimulation artefacts, of both algorithms developed were compared. While to reach the second aim, the volitional EMG estimates in the different phases of the protocol were compared using the Friedman test. To perform the Friedman test data was organised in five subsets: four subsets relative to each different phase of the protocol involving volitional contribution with the corresponding rest condition and one subset composed of all the rest conditions for each phase. For each of these subsets, a separate Friedman test and post hoc analysis was performed. The ultimate phase of the work was to implement the chosen algorithm in real-time in Simulink, allowing the estimation of the volitional EMG from hybrid contraction to be performed on-line. Results This thesis involved different steps of data acquisition, therefore the results of the work can be divided into different sections. The first outcome of this work was arising from the selection of the best configuration for set-up parameters. The trials to identify the configuration parameters of the EMG sensors found out that the best combination was represented by a sampling frequency of 2 kHz and a transmission frequency of 20 Hz, which was the configuration characterized by the lowest percentage of missed samples (0.1%), lowest coefficient of variation for the samples acquired at transmission frequency and highest RMS value of the volitional estimate. In the rest of the work, these parameters were used, and the stimulation frequency was set equal to the transmission frequency. In Figure 4, three different levels of volitional EMG can be clearly identified, the signal shown was acquired with the chosen configuration of set up parameters. The second fundamental result regards the comparison among the two algorithms for stimulation artefact detection and the estimate of the volitional EMG signal. This analysis demonstrated that both algorithms were successful in identifying the stimulation artefact and allow a correct volitional EMG estimation. The two algorithms were compared in terms of false positive and false negative, which were found to be zero in both cases. Given this result the choice was to implement the less complex algorithm in terms of the computational burden. The next step was to perform statistical analysis on the EMG signal of hybrid muscle contractions, extracted with the algorithm implemented. Figure 5 and figure 6, report the results of the Friedman test, where each histogram represents the median value, the vertical bar bounds the 25°and 75° percentiles and the ** refer to p<0.001, * to p<0.05. Conditions are codified by colours: blue for rest, green for volitional level 1 (no weight) and red for volitional level 2 (1 kg weightlifting). In figure 5 below it is possible to see that in the first phase of the protocol, when no stimulation is performed, all the phases are significantly different from each other as expected(p-value<0.001). However, in the phases in which the stimulation was delivered, the two volitional levels are not identified as significantly different (p-value<0.05). As ultimate validation, the algorithm implemented on-line was compared to the offline implementation in terms of artefact identification and root mean square error (RMSE) which are reported in Table 2 below. The volitional EMG estimated with the on-line and the off-line algorithm are reported in Figure 7. It is clearly visible in figure 7 that the on-line and the offline logics give very similar results, which means that the on-line logic implementation is valid. Conclusions and future works This thesis was aimed at implementing an online filtering methodology for the estimation of volitional EMG during hybrid muscle contractions, which was validated performing experimental trials involving six healthy subjects. The next steps of this work could be the validation of the algorithm on EMG signals acquired from stroke patients during hybrid muscle contractions and the development of a closed-loop EMG-controlled FES system based on the online estimate of the volitional component. This FES system might exploit also the inertial sensors included in the EMG sensors to estimate the arm kinematics.