Tesi etd-03272020-114726
Link copiato negli appunti
Tipo di tesi
Dottorato
Autore
BANFI, TOMMASO
URN
etd-03272020-114726
Titolo
Predictive models and longitudinal sensing hardware for surgical risk mitigation in sleep deprived and fatigued conditions
Settore scientifico disciplinare
Istituto di Biorobotica
Corso di studi
Istituto di Biorobotica - BIOROBOTICS
Commissione
relatore Prof. CIUTI, GASTONE
Tutor Prof. FARAGUNA, UGO
Tutor Prof. SABATINI, ANGELO MARIA
Tutor Prof.ssa MENCIASSI, ARIANNA
Tutor Prof. DARIO, PAOLO
Tutor Prof. FARAGUNA, UGO
Tutor Prof. SABATINI, ANGELO MARIA
Tutor Prof.ssa MENCIASSI, ARIANNA
Tutor Prof. DARIO, PAOLO
Parole chiave
- actigraphy
- classification
- eeg
- electroencephalography
- neural network
- sleep
- sleep deprivation
- surgeon
- surgery
- wearable
Data inizio appello
22/05/2020;
Disponibilità
completa
Riassunto analitico
Sleep deprivation is a common problematic in the global society and its prevalence is increasing. Chronic and acute sleep deprivation have been linked to diabetes and heart diseases as well as depression and enhanced impulsive behaviours. Surgeons are often exposed to long hours on call and few hours of sleep in the previous days. Nevertheless, few studies have focused their attention on the effects of sleep deprivation on surgeons and more specifically on the effects of sleep deprivation on surgical dexterity. A better understanding of the consequences of sleep loss on the key surgical skill of dexterity can shed light on the possible risks associated to a sleepy surgeon. Overall, the available literature does not provide a clear understanding of the effect of sleep deprivation on surgical dexterity. Available research presents inconsistencies in the measures of surgical dexterity and in the experimental designs employed. Known confounding factors acting on surgical dexterity measures, such as surgical expertise level, use of stimulants and possibly circadian performance variations are not consistently and objectively measured and included in the analysis of the results. Moreover, current studies do not directly and objectively measure sleep quantity nor, also importantly, provide quality metrics, relying instead on subjective sleep measures that are known to be affected by subjective bias. These heterogeneities prevent the possibility to perform a meta analytical analysis of the available studies in the field, hence preventing us to achieve an objective understanding of the effects of sleep deprivation on surgical dexterity, thus encouraging further investigations.
To fill the gap in current literature and provide field deployable sleep deprivation tracking methodologies and instruments, the research activity of my PhD aims to develop an integrated system embedding/able to:
1. (aim 1) measure sleep longitudinally in an accurate, unobtrusive and scalable manner;
2. (aim 2) predict significantly detrimental exposure to sleep deprivation of individuals;
3. (aim 3) develop a miniaturized custom wearable EEG hardware embedding: i) a high performance low power analogue front end, ii) ad hoc real time motion artefact reduction system, and iii) secure real-time EEG data streaming;
4. (aim 4) collect EEG data in real time to objectively and unobtrusively monitor in near real time the behavioural state of an individual working in unstructured unconstrained environments.
These system requirements are also meant to be interpreted as main aims of this thesis.
The integrated system is meant to be used in the field by surgeons without disrupting their workflow while providing information useful to minimize their exposition to sleep deprivation, ultimately leading to a risk reduction for both patients and surgeons. The design of the system is built on the knowledge coming from the neurophysiology of sleep and the neurobehavioural consequences of sleep deprivation. The overall system is meant to work on two different time scales in order to take into account the effects of chronic sleep deprivation and ensure reliable predictions of behavioral states in real time. The first time scale encompasses days to weeks, while the second one observes the monitored subject on time windows spanning the very last 30 seconds.
To tackle aim 1, a sleep/wake detection algorithm for efficient on-device sleep tracking using wearable accelerometric devices was developed. The proposed approach is built upon a classification model which is learned end-to-end using a convolutional neural network. The convolutional neural network is fed using raw accelerometric signals recorded by an open-source wrist-worn actigraph. The developed classifier has the following main characteristics: i) is highly generalizable to heterogenous subjects, ii) does not require manual features’ extraction, iii) is computationally lightweight, embeddable on a sleep tracking device, and iv) is suitable for a wide assortment of actigraphs. Sleep parameters, such as total sleep time, waking after sleep onset and sleep efficiency, calculated using the developed approach were compared to the gold standard polysomnographic concurrent recordings. The relatively substantial agreement (Cohen’s kappa coefficient, median, equal to 0.78±0.07) and the low-computational cost (2727 floating-point operations) makes this solution suitable for an on-board sleep-detection approach.
To tackle aim 2, the sleep/wake schedule created using actigraphic data is used together with a biomathematical model developed by Ramakrishnan et al. . The mathematical model is fed with the classified sleep/wake time series produced by the actigraphy based sleep monitoring system. The system may also be used to predict if a subject will experience significant adverse neurocognitive impairments related to sleep deprivation in the next shift or days and can model the effects of caffeine administration as well as the adoption of different work schedules.
Aim 3 and Aim 4 were addressed through the development of a miniaturized wearable EEG system paired with a dedicated data collection infrastructure, suitable when dealing with restrictions imposed by unstructured deployment environments. To this end a system on a chip coupled with an integrated analogue front-end acquires EEG signals and streams them in real-time to an external device. To simplify system usage, EEG signals are collected using a disposable array of 10 microelectrodes to be placed around the ear. Using an artificial neural network, I developed a classification algorithm able to classify in real time behavioural states employing only raw EEG data while using relatively modest computational resources. The current implementation of the model enables transfer learning to other datasets for each specific application, electrode location and EEG instrumentation settings. The classification performance of the proposed model reached an appropriate sleep-wake classification performance: AUC, precision and F1 score of 0.95, 0.62 and 0.73 respectively. A hardware prototype of a wearable EEG data acquisition device was developed together with the necessary custom firmware.
Further attention was devoted to aim 3 and 4 through the development of a custom motion artifact reduction technique. Motion artifacts are among the most common artifacts affecting EEG recordings and have the potential to severely alter signal quality. To this end the WEEG system embeds an inertial movement platform on the main board hosting the analog front end and an array of IMUs that are embedded in a flexible circuit board to be mechanically joined with the cEEG grid electrodes (a type of flexible EEG electrodes developed by Bleichner et al. ). This method enables direct error measurement and the application of empirical mode decomposition that uses this information to mitigate the undesirable effects of motion on the recorded EEG signal.
The developed system proposes a novel integrated approach that enables the concurrent and longitudinal measure of both chronic and acute sleep deprivation in unstructured setting. The system embeds possible solutions to overcome traditional limitations in the use of EEG in wearable devices and other neurophysiological applications. The developed system may contribute to ease the adoption of rigorous sleep study methodologies in the surgical field, possibly mitigating some of the specific limitations found in the available studies of the field. Emphasis on the use of miniaturized, low cost, automatic and non-invasive measurement techniques was used as a basic design philosophy, thus making the adopted solution a viable candidate for technology transfer to less demanding applications or in other telemedicine frameworks.
------------
1 - Ramakrishnan, S. et al. A Unified Model of Performance for Predicting the Effects of Sleep and Caffeine. Sleep 39, 1827–1841 (2016).
2 - Bleichner, M. G. & Debener, S. Concealed, Unobtrusive Ear-Centered EEG Acquisition: cEEGrids for Transparent EEG. Front. Hum. Neurosci. 11, 1–14 (2017).
To fill the gap in current literature and provide field deployable sleep deprivation tracking methodologies and instruments, the research activity of my PhD aims to develop an integrated system embedding/able to:
1. (aim 1) measure sleep longitudinally in an accurate, unobtrusive and scalable manner;
2. (aim 2) predict significantly detrimental exposure to sleep deprivation of individuals;
3. (aim 3) develop a miniaturized custom wearable EEG hardware embedding: i) a high performance low power analogue front end, ii) ad hoc real time motion artefact reduction system, and iii) secure real-time EEG data streaming;
4. (aim 4) collect EEG data in real time to objectively and unobtrusively monitor in near real time the behavioural state of an individual working in unstructured unconstrained environments.
These system requirements are also meant to be interpreted as main aims of this thesis.
The integrated system is meant to be used in the field by surgeons without disrupting their workflow while providing information useful to minimize their exposition to sleep deprivation, ultimately leading to a risk reduction for both patients and surgeons. The design of the system is built on the knowledge coming from the neurophysiology of sleep and the neurobehavioural consequences of sleep deprivation. The overall system is meant to work on two different time scales in order to take into account the effects of chronic sleep deprivation and ensure reliable predictions of behavioral states in real time. The first time scale encompasses days to weeks, while the second one observes the monitored subject on time windows spanning the very last 30 seconds.
To tackle aim 1, a sleep/wake detection algorithm for efficient on-device sleep tracking using wearable accelerometric devices was developed. The proposed approach is built upon a classification model which is learned end-to-end using a convolutional neural network. The convolutional neural network is fed using raw accelerometric signals recorded by an open-source wrist-worn actigraph. The developed classifier has the following main characteristics: i) is highly generalizable to heterogenous subjects, ii) does not require manual features’ extraction, iii) is computationally lightweight, embeddable on a sleep tracking device, and iv) is suitable for a wide assortment of actigraphs. Sleep parameters, such as total sleep time, waking after sleep onset and sleep efficiency, calculated using the developed approach were compared to the gold standard polysomnographic concurrent recordings. The relatively substantial agreement (Cohen’s kappa coefficient, median, equal to 0.78±0.07) and the low-computational cost (2727 floating-point operations) makes this solution suitable for an on-board sleep-detection approach.
To tackle aim 2, the sleep/wake schedule created using actigraphic data is used together with a biomathematical model developed by Ramakrishnan et al. . The mathematical model is fed with the classified sleep/wake time series produced by the actigraphy based sleep monitoring system. The system may also be used to predict if a subject will experience significant adverse neurocognitive impairments related to sleep deprivation in the next shift or days and can model the effects of caffeine administration as well as the adoption of different work schedules.
Aim 3 and Aim 4 were addressed through the development of a miniaturized wearable EEG system paired with a dedicated data collection infrastructure, suitable when dealing with restrictions imposed by unstructured deployment environments. To this end a system on a chip coupled with an integrated analogue front-end acquires EEG signals and streams them in real-time to an external device. To simplify system usage, EEG signals are collected using a disposable array of 10 microelectrodes to be placed around the ear. Using an artificial neural network, I developed a classification algorithm able to classify in real time behavioural states employing only raw EEG data while using relatively modest computational resources. The current implementation of the model enables transfer learning to other datasets for each specific application, electrode location and EEG instrumentation settings. The classification performance of the proposed model reached an appropriate sleep-wake classification performance: AUC, precision and F1 score of 0.95, 0.62 and 0.73 respectively. A hardware prototype of a wearable EEG data acquisition device was developed together with the necessary custom firmware.
Further attention was devoted to aim 3 and 4 through the development of a custom motion artifact reduction technique. Motion artifacts are among the most common artifacts affecting EEG recordings and have the potential to severely alter signal quality. To this end the WEEG system embeds an inertial movement platform on the main board hosting the analog front end and an array of IMUs that are embedded in a flexible circuit board to be mechanically joined with the cEEG grid electrodes (a type of flexible EEG electrodes developed by Bleichner et al. ). This method enables direct error measurement and the application of empirical mode decomposition that uses this information to mitigate the undesirable effects of motion on the recorded EEG signal.
The developed system proposes a novel integrated approach that enables the concurrent and longitudinal measure of both chronic and acute sleep deprivation in unstructured setting. The system embeds possible solutions to overcome traditional limitations in the use of EEG in wearable devices and other neurophysiological applications. The developed system may contribute to ease the adoption of rigorous sleep study methodologies in the surgical field, possibly mitigating some of the specific limitations found in the available studies of the field. Emphasis on the use of miniaturized, low cost, automatic and non-invasive measurement techniques was used as a basic design philosophy, thus making the adopted solution a viable candidate for technology transfer to less demanding applications or in other telemedicine frameworks.
------------
1 - Ramakrishnan, S. et al. A Unified Model of Performance for Predicting the Effects of Sleep and Caffeine. Sleep 39, 1827–1841 (2016).
2 - Bleichner, M. G. & Debener, S. Concealed, Unobtrusive Ear-Centered EEG Acquisition: cEEGrids for Transparent EEG. Front. Hum. Neurosci. 11, 1–14 (2017).
File
| Nome file | Dimensione |
|---|---|
| phd_thes...banfi.pdf | 5.88 Mb |
Contatta l'autore |
|