The laboratory set-up limits have promoted the development of new technologies able to monitor patients in their daily routine. In particular, human motion analysis development focuses on measuring the kinematics of body segments and gait-related descriptive parameters through wearable inertial sensors. These new technologies slowly entered the motion analysis field and thus also new algorithms started to be developed. Gait analysis had its major clinical impact as a test for patients with central nervous disorders associated with spasticity (Simon, 2004); cerebral palsy is the most common neuromotor disability in children, with an impact between 1.5 and 2.5 per 1000 births (Brégou Bourgeois et al., 2013). The majority of them could walk independently, with spastic hemiplegia or diplegia. Even in environments where standard gait analysis laboratories were available, they were often used such as pre- and post- operative assessments; moreover, a child natural gait usually was affected by the short distance walkways of the laboratory settings. For example, in order to quantify CP symptoms, the foot angle in correspondence to the heel-strike and the one in correspondence to the toe-off were computed. To do so, the initial and final contacts of the foot on the ground must be identified (Bötzel et al., 2015). By now, the research increasingly moved to sophisticated sensors and efficient algorithms to be implemented on large dataset. Simultaneously, the increasing spread in using machine learning to generate data-driven solutions influenced this science field. The functionality in using machine learning to analyze real time data and in providing prediction responses was the reason that moved this research. This work focused on the analysis and classification of human gait events and on the prediction of them during a gait cycle. The main goal was using a machine learning algorithm to predict heel-strikes and toe-offs events (and concurrently stance and swing phases) by extracting gait signals from a wearable inertial sensor. The data was compared and synchronized within a gold-standard gait analysis laboratory. To the best of our knowledge, no previous works adopted a similar approach; in addition, the majority of literature studies combined IMU and machine learning for the discrimination of different human activities, but no studies focused on the differentiation of different phases of the walking action. We hypothesize that such data-driven approach to be independent on pre-determined physical models (i.e. inverted pendulum), thus potentially embracing multiple activity conditions. If this will prove to be a reliable solution, this work would point the way for embedded wearable applications: on-chip microcontrollers nowadays can efficiently run semi-real time application featured by machine learning models, thus opening interesting application pathways for this approach. This study involved the simultaneous collection of gait data with a commercial IMU platform and within an instrumented motion analysis laboratory, serving as a reference for the supervised learning training. Forty healthy patients were acquired in the “Luigi Divieti” Posture and Movement Analysis Laboratory at the Politecnico di Milano. The total amount of steps recording was 500. The first procedure was to emulate the pre-existing methods about the gait events recognition through the wavelet decomposition. Subsequently, three classification models were used and tested: a multiclass classification model by using four labels in order to discriminate gait phases, and two binary classification models, by using first two labels for the classification of heel-strikes and toe-offs, then by classifying between stance phase and swing phase. The methods were then trained with a decision tree algorithm, and finally, the outputs were corrected through an implemented deterministic algorithm, aimed to improve the performances of the prediction models. The first binary classification model was implemented by creating two different prediction models and then by combining the responses. The first one was constructed to detect only the heel-strike events among others, whereas the second one was implemented to detect only toe- offs. On the contrary, the second binary classification model was implemented by using only one prediction learning to classify between stance phase and swing phase. In this situation, only one learning was chosen because the two phases were exclusive, such that the foot could be on the ground or lifted up, and any other situations were possible. The results from the binary classification method based on the identification of stance and swing times to evaluate gait events showed the most promising and realistic results. Using this implementation, the detection of heel-strikes and toe-offs occurred automatically, by imposing the first heel-strike at the beginning of the stance phase, the toe-off at the end of this last, and, finally, the second heel-strike at the end of the swing phase. Although by disposing of a workstation also other classifiers could be tried to choose the most accurate one, the decision tree achieved valuable performances (95.7% of accuracy). The previous considerations are supported also by the values resulted by the construction of the confusion matrices and ROC Curves. The confusion matrix for the binary classification method based on the differentiation of stance and swing phases showed that the stance phase was correctly predicted 93% of times, whereas the swing phase was exactly predicted 90% of times among the whole swing phase predictions. The binary classification method based on the differentiation of stance and swing phases also reported the higher area under the curve values (0.97 for both the phases considered). Thus, the results were then compared with those obtained by using the reference motion capture system. In particular, the statistical analysis was conducted by applying paired t-tests on the gold standard values and the estimation ones. In addition, also the mean absolute error was computed to evaluate the distance between the gold standard method and the estimation one. Finally, also the Bland-Altman plot was displayed, in order to analyze the agreement between the two methods outcomes. Concerning the classification of stance and swing phases, the data showed that the distance between the gold standard method and the binary one in classifying the two phases was equal to the values obtained by using the wavelet transform. In particular, the mean absolute error for stance phase identification was equal to 0.075 s, whereas the mean absolute error for swing phase identification was equal to 0.069 s, both for the wavelet transform and the binary classification method. In addition, the confidence interval of the binary stance and swing phases classification is centered around zero: the range for the stance phase identification is included between -0.57 s and 0.004 s, whereas for the swing phase identification the confidence interval was included between -0.030 s and 0.036 s. Moreover, t-tests returned nonsignificant p-values. In particular, for the stance phase identification, the p-value was 0.091, for the swing phase classification it was equal to 0.842. That meant that there were no significative differences between the two methods considered. In conclusion, the Bland-Altman plots confirmed these findings. The distribution of data by comparing the considered binary classification method with the gold standard resulted in a well-localized area of the graph, both for stance and swing phases. There were no outliers, as for the wavelet decomposition method with respect to the gold standard, but with the advantage that the developed classification model does not depend on the inverted pendulum model, and so that it can be applied to a wider range of activities and also to unphysiological conditions. Promising research pathways originate from the results of this work. The experiments with a commercial inertial unit and the machine learning approach with the correction algorithm leaves room for improvements and future developments in the field of IMUs and daily-life monitoring. The using of a IMU for detecting and monitoring human movement exploits the tendency of these days about the use of sensors during daily life activities. Clearly, the academic development of this work needs to be expanded for real-time applications, that is the main interesting field of application: the instantaneous evaluation of the gait pattern of an individual while he is doing activities in uncontrolled environments is the goal that inspires this research.
L’analisi del movimento umano consiste nella valutazione quantitativa del movimento e tratta la quantificazione di parametri cinematici, cinetici e variabili biomeccaniche, quali angoli articolari, momenti articolari e parametri spazio-temporali. I campi di applicazione sono numerosi. Tra questi senza dubbio il più comune è la valutazione clinica del cammino, finalizzata allo studio di condizioni patologiche o al supporto di decisioni riguardo trattamenti riabilitativi. L’impatto più rilevante dell’analisi del movimento è sui pazienti affetti da disordini del sistema nervoso centrale; tra questi, i bambini affetti da paralisi cerebrale infantile sono spesso sottoposti a tale analisi: per diagnosticare la patologia, si misura l’angolo del piede rispetto al suolo nei momenti di impatto del tallone e di distacco. Tradizionalmente, l’analisi del cammino avveniva esclusivamente in centri di ricerca o strutture sanitarie, all’interno di laboratori specializzati, dotati di sistemi optoelettronici e pedane di forza; questi sistemi sono dotati di alta precisione e accuratezza, tuttavia presentano limitazioni in termini di costi e di ristretto campo di applicazione. Lo sviluppo di nuove tecnologie ha permesso l’introduzione di sensori indossabili in grado di estendere l’analisi del movimento ad ambienti non circoscritti e sfruttando strumenti più accessibili; ciò ha permesso di monitorare l’attività umana durante la quotidianità, fornendo ai clinici strumenti di valutazione addizionali. Attualmente, esistono numerosi algoritmi in grado di classificare le fasi del cammino umano acquisendo dati da dispositivi indossabili. In particolare, numerosi studi scientifici trattano metodologie per la determinazione degli istanti che caratterizzano il passo: l’istante di appoggio del tallone al suolo, la fase di sostegno, l’istante in cui il piede si distacca dal suolo e, infine, la fase di volo del piede, che precede un nuovo ciclo del passo. Lo sviluppo dell’intelligenza artificiale applicato a sensori indossabili ha esteso l’approccio ingegneristico all’analisi del movimento. Infatti, l’intelligenza artificiale offre nuove tecniche per studiare le strategie utilizzate dall’uomo per camminare, procurando output in tempo reale tramite algoritmi di regressione o di classificazione avanzati. Il presente elaborato si concentra sull’analisi del movimento umano e, in particolare, sulla classificazione degli eventi del cammino che caratterizzano il ciclo del passo. L’obiettivo principale è lo sviluppo di un algoritmo di classificazione automatico per la previsione e dunque la classificazione delle fasi di appoggio del piede al terreno e di volo (e conseguentemente, degli istanti di appoggio del tallone al terreno e di distacco dal suolo), estraendo i segnali da un dispositivo inerziale indossabile. I dati sono stati sincronizzati e poi confrontati con quelli ottenuti utilizzando un sistema di monitoraggio del movimento di riferimento, in laboratorio. Attualmente, per quello che è stato possibile verificare, non esistono algoritmi che sfruttano l’intelligenza artificiale per la classificazione del cammino; in letteratura, sono presenti approcci simili per distinguere attività motorie da altre, ovvero per distinguere il cammino dalla corsa, o dalla salita o discesa dalle scale, acquisendo segnali da accelerometri e giroscopi. Per l’acquisizione dei dati sul cammino sono stati raccolti 500 passi, acquisiti da 40 soggetti sani, all’interno del laboratorio “Luigi Divieti” (Politecnico di Milano): ad ogni soggetto è stato chiesto di camminare a velocità preferenziale, e i dati relativi sono stati acquisiti simultaneamente da un’unità inerziale indossabile e da un sistema di riferimento optoelettronico e pedane di forza, presenti in laboratorio. Inizialmente, come suggerito da studi esistenti, è stata applicata la trasformata wavelet per il riconoscimento degli eventi e delle fasi del ciclo del passo. Successivamente, sono stati testati tre modelli di classificazione: un modello di classificazione multiclasse, utilizzando quattro marcatori per discriminare le fasi del cammino, e due modelli di classificazione binaria. In particolare, per il primo modello binario sono stati utilizzati due marcatori, uno per classificare l’istante di appoggio del tallone a terra, e uno per l’istante di distacco dal suolo, mentre per il secondo modello di classificazione binario, è stato scelto un marcatore per la fase di appoggio e uno per la fase di volo. Il primo modello di classificazione binario è stato implementato creando due differenti modelli di predizioni, poi combinati: il primo in grado di classificare solo l’evento di appoggio del tallone al terreno, il secondo in grado di classificare solo l’evento di distacco del piede dal suolo. Al contrario, il secondo modello di classificazione binaria è stato ideato in modo da classificare la posizione del piede rispetto al suolo, per cui un unico marcatore che cambia valore in base alla posizione del piede rispetto al suolo. Quest’ultima situazione ha permesso di utilizzare solo un modello di classificazione, in quanto le due fasi del ciclo del passo sono esclusive: il piede può trovarsi o in fase di appoggio, o in fase di volo. Successivamente, è stato “allenato” l’albero decisionale sui tre metodi di classificazione. I tre modelli sono stati poi applicati al dataset di test per ottenere una predizione; per aumentare le performance del modello, è stato implementato un algoritmo deterministico di correzione dell’output fornito dal modello di classificazione. I risultati ottenuti applicando il modello di classificazione binaria basato su fase di appoggio e di volo del piede sono i più promettenti e realistici. Utilizzando questo modello, l’identificazione dell’istante di appoggio e del momento di distacco del piede avvengono automaticamente, individuando l’istante di inizio della fase di appoggio e l’inizio della fase di volo, corrispondentemente. L’albero decisionale applicato ha mostrato una buona performance (95.7% di accuratezza), e anche i valori ottenuti dall’analisi delle matrici di confusione e delle curve ROC confermano l’efficienza del modello. Analizzando la matrice di confusione per il modello di classificazione binario in questione, risulta che la fase di appoggio viene classificata correttamente il 93% delle volte sulle predizioni totali della suddetta fase, mentre quella di volo il 90%. Inoltre, dalle curve ROC risulta un’area sottesa alla curva pari a 0.97 per entrambe le fasi del ciclo. Infine, i risultati sono stati poi confrontati con quelli ottenuti utilizzando il sistema di riferimento optoelettronico: l’analisi statistica è stata eseguita applicando T-Test e analizzando il grafico di Bland-Altman per comparare i valori stimati e quelli ottenuti dal sistema di riferimento. I dati ottenuti hanno mostrato che l’errore medio assoluto del metodo di classificazione binario rispetto al sistema di riferimento è equivalente all’errore risultato dall’applicazione della trasformata wavelet. In particolare, per entrambi i metodi, l’errore nella fase di appoggio è di 0.075 s, mentre quello nella fase di volo è di 0.069 s. Inoltre, l’intervallo di confidenza al 95% è incluso tra -0.057 s e 0.004 s per la prima fase, e tra -0.030 s e 0.036 s per la seconda. Il p-value ottenuto (t-test) è maggiore di 0.05 per entrambe le fasi (per la fase di appoggio risulta 0.091, per la fase di volo 0.842). Infine, anche l’analisi del grafico di Bland-Altman ha confermato che il modello di classificazione scelto è il più accurato e promettente: come la trasformata wavelet, ha riportato valori ben localizzati nel grafico e senza outliers, con il vantaggio di essere indipendente da modelli fisici, quali il modello del pendolo inverso, e quindi essere applicabile a più attività motorie e a più situazioni, anche non fisiologiche. I valori ottenuti dimostrano che l’approccio ideato utilizzando il metodo di classificazione binario basato su fase di appoggio e fase di volo, con seguente correzione tramite algoritmo deterministico, è una valida soluzione e alternativa all’utilizzo della trasformata wavelet per l’identificazione degli eventi del ciclo del passo. È tendenza sempre più diffusa utilizzare sensori indossabili per il monitoraggio delle attività motorie, e il più interessante campo di applicazione, nonché il fine ultimo di questa ricerca, è il monitoraggio del movimento in real-time. Il presente lavoro potrebbe dunque avere sviluppi futuri promettenti nell’ambito dell’intelligenza artificiale applicata a IMU per il monitoraggio del movimento umano quotidiano, all’aria aperta e in real-time.
A classification model to identify human gait events : a machine learning approach exploiting a wearable inertial measurement unit
DELFINO SPIGA, MARTINA
2018/2019
Abstract
The laboratory set-up limits have promoted the development of new technologies able to monitor patients in their daily routine. In particular, human motion analysis development focuses on measuring the kinematics of body segments and gait-related descriptive parameters through wearable inertial sensors. These new technologies slowly entered the motion analysis field and thus also new algorithms started to be developed. Gait analysis had its major clinical impact as a test for patients with central nervous disorders associated with spasticity (Simon, 2004); cerebral palsy is the most common neuromotor disability in children, with an impact between 1.5 and 2.5 per 1000 births (Brégou Bourgeois et al., 2013). The majority of them could walk independently, with spastic hemiplegia or diplegia. Even in environments where standard gait analysis laboratories were available, they were often used such as pre- and post- operative assessments; moreover, a child natural gait usually was affected by the short distance walkways of the laboratory settings. For example, in order to quantify CP symptoms, the foot angle in correspondence to the heel-strike and the one in correspondence to the toe-off were computed. To do so, the initial and final contacts of the foot on the ground must be identified (Bötzel et al., 2015). By now, the research increasingly moved to sophisticated sensors and efficient algorithms to be implemented on large dataset. Simultaneously, the increasing spread in using machine learning to generate data-driven solutions influenced this science field. The functionality in using machine learning to analyze real time data and in providing prediction responses was the reason that moved this research. This work focused on the analysis and classification of human gait events and on the prediction of them during a gait cycle. The main goal was using a machine learning algorithm to predict heel-strikes and toe-offs events (and concurrently stance and swing phases) by extracting gait signals from a wearable inertial sensor. The data was compared and synchronized within a gold-standard gait analysis laboratory. To the best of our knowledge, no previous works adopted a similar approach; in addition, the majority of literature studies combined IMU and machine learning for the discrimination of different human activities, but no studies focused on the differentiation of different phases of the walking action. We hypothesize that such data-driven approach to be independent on pre-determined physical models (i.e. inverted pendulum), thus potentially embracing multiple activity conditions. If this will prove to be a reliable solution, this work would point the way for embedded wearable applications: on-chip microcontrollers nowadays can efficiently run semi-real time application featured by machine learning models, thus opening interesting application pathways for this approach. This study involved the simultaneous collection of gait data with a commercial IMU platform and within an instrumented motion analysis laboratory, serving as a reference for the supervised learning training. Forty healthy patients were acquired in the “Luigi Divieti” Posture and Movement Analysis Laboratory at the Politecnico di Milano. The total amount of steps recording was 500. The first procedure was to emulate the pre-existing methods about the gait events recognition through the wavelet decomposition. Subsequently, three classification models were used and tested: a multiclass classification model by using four labels in order to discriminate gait phases, and two binary classification models, by using first two labels for the classification of heel-strikes and toe-offs, then by classifying between stance phase and swing phase. The methods were then trained with a decision tree algorithm, and finally, the outputs were corrected through an implemented deterministic algorithm, aimed to improve the performances of the prediction models. The first binary classification model was implemented by creating two different prediction models and then by combining the responses. The first one was constructed to detect only the heel-strike events among others, whereas the second one was implemented to detect only toe- offs. On the contrary, the second binary classification model was implemented by using only one prediction learning to classify between stance phase and swing phase. In this situation, only one learning was chosen because the two phases were exclusive, such that the foot could be on the ground or lifted up, and any other situations were possible. The results from the binary classification method based on the identification of stance and swing times to evaluate gait events showed the most promising and realistic results. Using this implementation, the detection of heel-strikes and toe-offs occurred automatically, by imposing the first heel-strike at the beginning of the stance phase, the toe-off at the end of this last, and, finally, the second heel-strike at the end of the swing phase. Although by disposing of a workstation also other classifiers could be tried to choose the most accurate one, the decision tree achieved valuable performances (95.7% of accuracy). The previous considerations are supported also by the values resulted by the construction of the confusion matrices and ROC Curves. The confusion matrix for the binary classification method based on the differentiation of stance and swing phases showed that the stance phase was correctly predicted 93% of times, whereas the swing phase was exactly predicted 90% of times among the whole swing phase predictions. The binary classification method based on the differentiation of stance and swing phases also reported the higher area under the curve values (0.97 for both the phases considered). Thus, the results were then compared with those obtained by using the reference motion capture system. In particular, the statistical analysis was conducted by applying paired t-tests on the gold standard values and the estimation ones. In addition, also the mean absolute error was computed to evaluate the distance between the gold standard method and the estimation one. Finally, also the Bland-Altman plot was displayed, in order to analyze the agreement between the two methods outcomes. Concerning the classification of stance and swing phases, the data showed that the distance between the gold standard method and the binary one in classifying the two phases was equal to the values obtained by using the wavelet transform. In particular, the mean absolute error for stance phase identification was equal to 0.075 s, whereas the mean absolute error for swing phase identification was equal to 0.069 s, both for the wavelet transform and the binary classification method. In addition, the confidence interval of the binary stance and swing phases classification is centered around zero: the range for the stance phase identification is included between -0.57 s and 0.004 s, whereas for the swing phase identification the confidence interval was included between -0.030 s and 0.036 s. Moreover, t-tests returned nonsignificant p-values. In particular, for the stance phase identification, the p-value was 0.091, for the swing phase classification it was equal to 0.842. That meant that there were no significative differences between the two methods considered. In conclusion, the Bland-Altman plots confirmed these findings. The distribution of data by comparing the considered binary classification method with the gold standard resulted in a well-localized area of the graph, both for stance and swing phases. There were no outliers, as for the wavelet decomposition method with respect to the gold standard, but with the advantage that the developed classification model does not depend on the inverted pendulum model, and so that it can be applied to a wider range of activities and also to unphysiological conditions. Promising research pathways originate from the results of this work. The experiments with a commercial inertial unit and the machine learning approach with the correction algorithm leaves room for improvements and future developments in the field of IMUs and daily-life monitoring. The using of a IMU for detecting and monitoring human movement exploits the tendency of these days about the use of sensors during daily life activities. Clearly, the academic development of this work needs to be expanded for real-time applications, that is the main interesting field of application: the instantaneous evaluation of the gait pattern of an individual while he is doing activities in uncontrolled environments is the goal that inspires this research.| File | Dimensione | Formato | |
|---|---|---|---|
|
DelfinoSpigaMartina_TESI.pdf
non accessibile
Descrizione: Testo della tesi
Dimensione
17.19 MB
Formato
Adobe PDF
|
17.19 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/10589/154306