Sleep is a recurring daily state essential for a person’s health and wellbeing. Nowadays, sleep is being researched and monitored due to all the functional and economic implications of sleep deprivation. Other than sleep disorders, sleep can be disturbed by psychosocial issues such as stress. That is why there is a growing interest in finding new ways to monitor sleep, e.g., using wearable sensors that capture physiological signals like electrodermal activity (EDA). EDA is considered a sensitive index for emotional processing since it reflects sympathetic activity which is activated by the stress response of the body. However, despite the major impact of stress on sleep, there is a lack of EDA studies in the context of sleep quality studies. Therefore, the first aim of this master thesis is to research the possible relationship between EDA signal and sleep quality. EDA is the result of the activation of the eccrine sweat glands which activate due to thermoregulation or ‘emotional’ sweating. The second objective of this thesis is to explore the relationship between EDA and skin temperature during both day and night. Physiological data from wearable sensors (such as the electrocardiogram, skin conductance, skin temperature, activity...) have been collected within IMEC to research their use for sleep analysis. Therefore, a sleep detection algorithm was already developed before the master thesis. In this master thesis, research was executed with as goal to adapt this algorithm so it would accurately classify sleep without sleep diaries as input. Also, the newly adapted algorithm was validated against other open-source sleep detection algorithms to evaluate sleep classification efficiency. The research questions, previously mentioned, were investigated using data exploration and statistical analysis (e.g., Pearson correlation), machine learning (k-nearest neighbours, decision trees, random forest, support vector machine and multi-layer perceptron), principal component analysis (PCA) and multi-level models (linear mixed regression models). For the first research question, a relationship between EDA and sleep quality was not really observed. 3 out of the 6 tested machine learning models (k-nearest neighbours, support vector machine, and multi-layer perceptron) did not perform well (i.e., area under ROC < 0.5). However, the best performing machine learning model was a decision tree that had an accuracy of 71% and f1 score of 72% for the classification of sleep quality using EDA. Also, PCA was applied to the feature dataset, but there weren’t additional findings. For the multi-level models, a significant relationship (p-value<0.05) between EDA and sleep quality was not found. However, there were some limitations to this work su the as absence of EDA dynamics in the model (since only the mean EDA was considered for the whole night), a small dataset, lack of confounders in the model, such as temperature and mean motion (which influence the EDA signal), and possible incorrectness of the sleep quality parameters since the sleep algorithm used was not validated with polysomnography (PSG). Therefore, further research is needed to conclude these findings. Regarding the second objective, a significant relationship between EDA and skin temperature was found for using both data from the day and the night. Moreover, as expected, thermoregulation was not the only process causing the EDA signal since the variance explained by skin temperature was less than 1%. Therefore, the effect size of this interaction is very small. It was also confirmed that motion and day/night status significantly affect EDA signal. When motion and day/night variables were included in the model, the explained variance by the fixed effects increased to 8%. Moreover, a different ‘interaction’ between EDA and skin temperature during day and night was also observed. However, additional data is needed in order to confirm these results. In future work, data related to general arousal should be included in the models in order to effectively say if the EDA signal from the wrist is related to emotional sweating.
Il sonno è uno stato che ricorre quotidianamente, essenziale per la salute e il benessere di una persona. Al giorno d'oggi, il sonno è oggetto di ricerca e monitoraggio a causa di tutte le implicazioni funzionali ed economiche derivanti dalla privazione del sonno. Oltre ai disturbi specifici del sonno, questo può essere disturbato anche da problemi psico-sociali come lo stress. Per questo motivo, c'è un crescente interesse nel trovare nuovi metodi per monitorare il sonno, ad esempio, utilizzando sensori indossabili che catturano segnali fisiologici come l'attività elettro-dermica (EDA). L'EDA è considerato un indicatore sensibile dell'elaborazione emotiva poiché riflette l'attività simpatica attivata dalla risposta allo stress. Tuttavia, nonostante il grande impatto dello stress sul sonno, mancano studi EDA nel contesto degli studi sulla qualità del sonno. Pertanto, il primo obiettivo di questo lavoro di tesi è quello di ricercare la possibile relazione tra il segnale EDA e la qualità del sonno. L'EDA è il risultato dell'attivazione delle ghiandole sudoripare eccrine che si attivano per termoregolazione o sudorazione "emotiva". Il secondo obiettivo di questa tesi è esplorare la relazione tra il segnale EDA e la temperatura cutanea, sia di giorno che di notte. I dati fisiologici registrati da sensori indossabili (come l'elettrocardiogramma, la conduttanza cutanea, la temperatura cutanea, l'attività...) sono stati raccolti all'interno di IMEC per ricercarne il loro uso nell'analisi del sonno. Pertanto, un algoritmo di rilevamento del sonno è già stato sviluppato prima della suddetta tesi di laurea. In questo lavoro si è studiata la possibilità di adattare questo algoritmo già esistente in modo da classificare accuratamente il sonno senza diari del sonno come input. Inoltre, l'algoritmo appena adattato è stato validato rispetto ad altri algoritmi di rilevamento del sonno open source per valutarne l'efficienza di classificazione del sonno. Le domande di ricerca, precedentemente menzionate, sono state studiate utilizzando l'esplorazione dei dati e l'analisi statistica (ad esempio l’indice di correlazione di Pearson), il machine learning (k-nearest neighbours, alberi decisionali, random forest, support vector machine, multi-layer perceptron, ecc.), PCA (principal component analysis) e modelli multilivello (modelli di regressione lineare misti). Per la prima domanda di ricerca, non è stata realmente osservata una relazione tra EDA e qualità del sonno. 3 dei 6 modelli testati (k-nearest neighbours, support vector machine, multi-layer perceptron) non hanno funzionato bene (area sotto la curva ROC <0,5). Tuttavia, il modello con le migliori prestazioni è stato decision tree con un'accuratezza del 71% e un punteggio f1 del 72% per la classificazione della qualità del sonno mediante EDA. Anche la PCA è stata applicata al set di dati, ma non sono stati rilevati ulteriori risultati. Per i modelli multilivello non è stata trovata una relazione significativa (p-value < 0,05) tra EDA e qualità del sonno. Tuttavia, esistono alcune limitazioni a questo lavoro come un set di dati di piccole dimensioni, la mancanza di fattori confondenti nel modello come la temperatura e il movimento medio (che potrebbero influenzare il segnale EDA) e la possibile scorrettezza dei parametri della qualità del sonno poiché l'algoritmo del sonno utilizzato non è stato convalidato con polisonnografia (PSG). Pertanto, sono necessarie ulteriori ricerche per concludere questi risultati. Per quanto riguarda il secondo obiettivo, è stata trovata una relazione significativa tra EDA e temperatura cutanea sia durante il giorno che la notte. Inoltre, come previsto, la termoregolazione non era l'unico processo che causava il segnale EDA poiché la varianza spiegata dalla temperatura cutanea era inferiore all'1%. Pertanto, l'effetto dimensionale di questa interazione è molto piccola. È stato riscontrato che anche il movimento e lo stato giorno/notte influiscono significativamente sul segnale EDA. Quando nel modello vengono incluse le variabili movimento e giorno/notte, la varianza spiegata dagli effetti fissi aumenta all'8%. Inoltre, è stata osservata anche una diversa "interazione" durante il giorno e la notte tra l'EDA e la temperatura cutanea. Tuttavia, sono necessari ulteriori dati per confermare questi risultati. In lavori futuri, i dati relativi all’arousal generale dovrebbero essere inclusi nei modelli per dire se effettivamente c’e’ una relazione significativa tra l’EDA del polso e la sudorazione emotiva.
Exploration of multi modal datasets towards sleep quality analysis
Mufali, Xhesjana
2021/2022
Abstract
Sleep is a recurring daily state essential for a person’s health and wellbeing. Nowadays, sleep is being researched and monitored due to all the functional and economic implications of sleep deprivation. Other than sleep disorders, sleep can be disturbed by psychosocial issues such as stress. That is why there is a growing interest in finding new ways to monitor sleep, e.g., using wearable sensors that capture physiological signals like electrodermal activity (EDA). EDA is considered a sensitive index for emotional processing since it reflects sympathetic activity which is activated by the stress response of the body. However, despite the major impact of stress on sleep, there is a lack of EDA studies in the context of sleep quality studies. Therefore, the first aim of this master thesis is to research the possible relationship between EDA signal and sleep quality. EDA is the result of the activation of the eccrine sweat glands which activate due to thermoregulation or ‘emotional’ sweating. The second objective of this thesis is to explore the relationship between EDA and skin temperature during both day and night. Physiological data from wearable sensors (such as the electrocardiogram, skin conductance, skin temperature, activity...) have been collected within IMEC to research their use for sleep analysis. Therefore, a sleep detection algorithm was already developed before the master thesis. In this master thesis, research was executed with as goal to adapt this algorithm so it would accurately classify sleep without sleep diaries as input. Also, the newly adapted algorithm was validated against other open-source sleep detection algorithms to evaluate sleep classification efficiency. The research questions, previously mentioned, were investigated using data exploration and statistical analysis (e.g., Pearson correlation), machine learning (k-nearest neighbours, decision trees, random forest, support vector machine and multi-layer perceptron), principal component analysis (PCA) and multi-level models (linear mixed regression models). For the first research question, a relationship between EDA and sleep quality was not really observed. 3 out of the 6 tested machine learning models (k-nearest neighbours, support vector machine, and multi-layer perceptron) did not perform well (i.e., area under ROC < 0.5). However, the best performing machine learning model was a decision tree that had an accuracy of 71% and f1 score of 72% for the classification of sleep quality using EDA. Also, PCA was applied to the feature dataset, but there weren’t additional findings. For the multi-level models, a significant relationship (p-value<0.05) between EDA and sleep quality was not found. However, there were some limitations to this work su the as absence of EDA dynamics in the model (since only the mean EDA was considered for the whole night), a small dataset, lack of confounders in the model, such as temperature and mean motion (which influence the EDA signal), and possible incorrectness of the sleep quality parameters since the sleep algorithm used was not validated with polysomnography (PSG). Therefore, further research is needed to conclude these findings. Regarding the second objective, a significant relationship between EDA and skin temperature was found for using both data from the day and the night. Moreover, as expected, thermoregulation was not the only process causing the EDA signal since the variance explained by skin temperature was less than 1%. Therefore, the effect size of this interaction is very small. It was also confirmed that motion and day/night status significantly affect EDA signal. When motion and day/night variables were included in the model, the explained variance by the fixed effects increased to 8%. Moreover, a different ‘interaction’ between EDA and skin temperature during day and night was also observed. However, additional data is needed in order to confirm these results. In future work, data related to general arousal should be included in the models in order to effectively say if the EDA signal from the wrist is related to emotional sweating.| File | Dimensione | Formato | |
|---|---|---|---|
|
21_22_luglio_ Mufali.pdf
accessibile in internet per tutti
Dimensione
3.67 MB
Formato
Adobe PDF
|
3.67 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/10589/189575