Spinal cord injury (SCI) refers to damage to the spinal cord that results in partial or total loss of motor and/or sensory functions below the injury level. Affecting over 15 million people worldwide as of 2024, SCI imposes significant limitations on quality of life by restricting individuals' ability to carry out fundamental activities of daily living, such as walking. Aside from conventional therapies, including physical therapy and the use of assistive devices, epidural electrical stimulation (EES) is emerging as a new approach for promoting motor recovery after injury. Although EES has shown effectiveness in restoring motor function in individuals with SCI, key challenges remain. The patients' differing anatomies and responsiveness to stimulation generate the need to identify patient-specific stimulation patterns. Therefore, it is necessary to develop reliable metrics for the assessment of efficient stimulation parameters, while enabling longitudinal evaluation of the therapy. Force exerted by the muscles can represent one of these metrics. This project investigates the relationship between force and surface electromyography (sEMG) to develop an accurate force estimation model using sEMG data. Initial exploration of mean frequency led to its exclusion as a sole estimator, due to an inconsistent relationship with force. Machine learning and deep learning models were thus identified as possible methods for uncovering a potential hidden relationship between sEMG and exerted force. Three healthy subjects were requested to execute single joint movements in order to collect sEMG from the tibialis anterior and vastus lateralis while recording the force with a dynamometer. Algorithms including support vector regressor (SVR), light gradient-boosting machine (LGBM) and convolutional neural network (CNN) were trained with concatenations of different EMG features, and their performances were compared to linear regression based on sEMG amplitude. Since an overall reliable metric would need to work consistently across therapy sessions, the experiments incorporated multiple electrode placements to account for consequent variation in amplitude. The linear regression model generally showed the highest error rates, whereas LGBM consistently demonstrated the best performance across both muscles. While DL models achieved promising accuracy on certain datasets, their performance was inconsistent, often reverting to average predictions and lacking precision. These findings suggest that there could be an underlying force-sEMG relationship. However, the substantial error of even the best-performing models highlights a need for larger datasets, improved feature extraction or more advanced DL architectures to achieve reliable and clinically applicable results.
La lesione del midollo spinale (SCI) rappresenta un danno al midollo spinale che comporta una perdita parziale o totale delle funzioni motorie e/o sensoriali al di sotto del livello della lesione. Nel 2024, ne sono affette 15 milioni di persone in tutto il mondo. La SCI impone limitazioni significative alla qualità della vita, limitando la capacità di svolgere attività fondamentali della vita quotidiana, come camminare. Oltre alle terapie convenzionali, tra cui la fisioterapia e l'uso di dispositivi assistivi, la stimolazione elettrica epidurale (EES) sta emergendo come un nuovo approccio per promuovere il recupero motorio dopo una lesione. Sebbene l'EES abbia dimostrato efficacia nel ripristinare le funzioni motorie in individui con SCI, rimangono sfide aperte. Le differenze anatomiche e la diversa risposta alla stimolazione nei pazienti rendono necessario identificare pattern di stimolazione specifici per ciascun paziente. È quindi necessario sviluppare metriche affidabili per la valutazione dei parametri di stimolazione efficaci, consentendo al contempo una valutazione longitudinale della terapia. La forza esercitata dai muscoli può rappresentare una di queste metriche. Questo progetto indaga la relazione tra forza ed elettromiografia di superficie (sEMG) per sviluppare un modello di stima della forza utilizzando i dati sEMG. Un'esplorazione iniziale della frequenza media ha portato alla sua esclusione come unico parametro, a causa di una relazione inconsistente con la forza. I modelli di Machine Learning e Deep Learning sono stati quindi identificati come possibili metodi per esplorare una potenziale relazione nascosta tra sEMG e forza. A tre soggetti sani è stato chiesto di eseguire movimenti semplici, caratterizzati dal coinvolgimento di una singola articolazione, per raccogliere sEMG dal tibiale anteriore e dal vasto laterale. La forza esercitata dal muscolo target è stata acquisita attraverso un dinamometro. Algoritmi come Support Vector Regression (SVR), Light Gradient-boosting Machine (LGBM), e Convolutional Neural Networks (CNN) sono stati addestrati con concatenazioni di diverse features estratte dall’EMG, e le loro prestazioni sono state confrontate con una regressione lineare basata sull'ampiezza del sEMG. Poiché una metrica affidabile dovrebbe funzionare in modo coerente tra le sessioni di terapia, gli esperimenti hanno incorporato più posizionamenti degli elettrodi per tenere conto delle variazioni di ampiezza. Il modello di regressione lineare ha generalmente mostrato tassi di errore più alti, mentre LGBM ha dimostrato costantemente le migliori prestazioni in entrambi i muscoli. Sebbene i modelli di DL abbiano raggiunto sufficiente precisione su alcuni dataset, le loro prestazioni sono risultate incoerenti, tendendo spesso a predizioni vicine alla media del segnale e a mancanza di precisione. Questi risultati suggeriscono che potrebbe esserci una relazione complessa tra forza e sEMG. Tuttavia, l’errore sostanziale anche dei modelli con le migliori prestazioni evidenzia la necessità di dataset più ampi, un miglioramento dell’estrazione delle caratteristiche o architetture di DL più avanzate per ottenere risultati affidabili e clinicamente applicabili.
Investigation of the relationship between surface electromyography and force in single joint movements
Nevidal, Dora
2023/2024
Abstract
Spinal cord injury (SCI) refers to damage to the spinal cord that results in partial or total loss of motor and/or sensory functions below the injury level. Affecting over 15 million people worldwide as of 2024, SCI imposes significant limitations on quality of life by restricting individuals' ability to carry out fundamental activities of daily living, such as walking. Aside from conventional therapies, including physical therapy and the use of assistive devices, epidural electrical stimulation (EES) is emerging as a new approach for promoting motor recovery after injury. Although EES has shown effectiveness in restoring motor function in individuals with SCI, key challenges remain. The patients' differing anatomies and responsiveness to stimulation generate the need to identify patient-specific stimulation patterns. Therefore, it is necessary to develop reliable metrics for the assessment of efficient stimulation parameters, while enabling longitudinal evaluation of the therapy. Force exerted by the muscles can represent one of these metrics. This project investigates the relationship between force and surface electromyography (sEMG) to develop an accurate force estimation model using sEMG data. Initial exploration of mean frequency led to its exclusion as a sole estimator, due to an inconsistent relationship with force. Machine learning and deep learning models were thus identified as possible methods for uncovering a potential hidden relationship between sEMG and exerted force. Three healthy subjects were requested to execute single joint movements in order to collect sEMG from the tibialis anterior and vastus lateralis while recording the force with a dynamometer. Algorithms including support vector regressor (SVR), light gradient-boosting machine (LGBM) and convolutional neural network (CNN) were trained with concatenations of different EMG features, and their performances were compared to linear regression based on sEMG amplitude. Since an overall reliable metric would need to work consistently across therapy sessions, the experiments incorporated multiple electrode placements to account for consequent variation in amplitude. The linear regression model generally showed the highest error rates, whereas LGBM consistently demonstrated the best performance across both muscles. While DL models achieved promising accuracy on certain datasets, their performance was inconsistent, often reverting to average predictions and lacking precision. These findings suggest that there could be an underlying force-sEMG relationship. However, the substantial error of even the best-performing models highlights a need for larger datasets, improved feature extraction or more advanced DL architectures to achieve reliable and clinically applicable results.File | Dimensione | Formato | |
---|---|---|---|
2024_12_Nevidal_Executive_Summary_02.pdf
non accessibile
Descrizione: executive summary
Dimensione
1.85 MB
Formato
Adobe PDF
|
1.85 MB | Adobe PDF | Visualizza/Apri |
2024_12_Nevidal_Thesis_01.pdf
non accessibile
Descrizione: thesis text
Dimensione
6.21 MB
Formato
Adobe PDF
|
6.21 MB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/231504