Brain-Computer Interfaces (BCIs) open avenues for communication among individuals unable to use voice or gestures. Silent speech interfaces are one such approach for BCIs that could offer a trans- formative means of connecting with the external world. Performance on imagined speech decoding however is rather low due to, amongst others, data scarcity and the lack of a clear starting point of the imagined speech in the brain signal. We investigate whether using electroencephalography (EEG) signals from articulated speech can be used to improve imagined speech decoding in two ways: we investigate whether articulated speech EEG signals can be used to predict the end point of the imagined speech and use the articulated speech EEG as extra training data for speaker-independent imagined vowel classification. Our results show that using EEG data from articulated speech did not improve classification of vowels in imagined speech, probably due to high variability in EEG signals amongst speakers.

With the aging population, the demand for healthcare and related services is increasing and, for this reason, technologies for remote patient monitoring are developing, aiming at indoor scenarios. Remote patient monitoring can help capture the clinical data of patients at home, which can save time and money, specifically reducing the need for hospitalization by potentially detecting health-related issues before they become too serious.
The non-contact radar-based technology can be applied in the remote patient monitoring system for detecting vital signs. Radars are suitable for applications at home because they are non-invasive, robust in changing lighting and temperature, and suitable for patients with skin irritation.
Heartbeat and respiration are critical clinical data for the diagnosis of the disease. The study of respiration frequency estimation was explored by previous work, such as the MSc thesis in \cite{Maxthesis}. Building on that work, this project proposes a pipeline to measure the heartbeat frequency and cancel the random body movement. The impact of different orientations is also studied. The phase history difference of the chest displacement due to vital signs is extracted, and the wavelet transform is used to separate heartbeat and respiration signals. Different methods are tested to calculate the heartbeat frequency in the time and frequency domain. The RBM is detected by the energy threshold of the phase difference, and the intervals with the RBM are discarded.
The simulation and experimental results indicate that the proposed processing pipeline can work on the radar data.

Blind Source Separation (BSS), the separation of latent source components from observed mixtures, is relevant to many fields of expertise such as neuro-imaging, economics and machine learning. Reliable estimates of the sources can be obtained through diagonalization of the cumulant tensor, i.e., a fourth-order symmetric multi-linear array containing the cross-kurtosis values of observed mixtures. The downside of such diagonalization methods is that they scale quartically with the increase of the amount of source components to estimate due to the tensor’s quartic size. Tensor decomposition can simultaneously diagonalize the cumulant tensor and address its size. However, it does not resolve the scalability issue due to the restriction of having to first explicitly compute the tensor.

It is studied how decomposing the cumulant tensor in implicit fashion can be used to solve the BSS problem while simultaneously addressing its scalablity issue. A class of implicit cumulant tensor decomposition algorithms is derived which scale more favorably than their explicit counterparts in terms of either computational cost, storage cost or both. Firstly, a novel QR-Tensor algorithm (QRT) is introduced which allows for the simultaneous diagonalization of a tensor’s outer-slices. It is theoretically shown how an implicit version of the QRT algorithm can be used to solve the BSS problem at a linearly scaling computational cost. Secondly, a fixed-point Canonical Polyadic Decomposition (CPD) iteration method is presented. It reduces the computational complexity from a quartic dependence to a linear dependence on the amount of signals to estimate. The source estimation performance of the devised implicit decomposition methods is compared to that of the state-of-the-art FastICA for an artificial linear BSS problem.

Results show that both fixed-point CPD and QRT are superior to FastICA when it comes to the computation time needed to reach convergence, while producing estimated sources of similar quality. It is shown that when the amount of sources to estimate is increased both QRT and FastICA struggle to converge. In contrast, the fixed-point CPD method converges within a consistent amount of iterations, suggesting a method more suitable for the estimation of a large amount of sources.

The challenge of dealing with patients suffering from chronic diseases and an aging population requires evolving from traditional hospital-based healthcare systems into a person-centered approach, where patients can be monitored remotely via modern technologies by cost-effective and reliable solutions based on emerging technologies in the healthcare domain.

Due to its contactless capabilities, radio-frequency technologies can lead to proactive monitoring of conditions directly related to health statuses. These technologies can include the tracking and monitoring of vital signs or the identification of abnormalities and critical life-threatening events, such as strokes or falls, in order to react before more complex scenarios and non-treatable conditions can appear over time.

This thesis project explores developing, evaluating, and verifying a processing pipeline based on radar sensing technology, jointly exploring human activity recognition and breathing frequency estimation, two of the most immediate capabilities to detect and monitor the general health conditions of a human being.

Through Doppler-Time and Range-Time data domains, the differentiation between translational and in-place activities, namely walking and sitting, is addressed, aiming to successfully identify and locate the segments where the test subject is not moving. This then triggers a proposed pipeline for the continuous estimation of the breathing frequency for the in-place scenario based on a sequential estimator, specifically the extended Kalman filter.

Absence seizures have a real-life impact on epileptic subjects, as day-to-day tasks can by suddenly interrupted making for dangerous situations. Though a lot of work has been done on seizure detection, to limit the impact on epileptic patients, the true necessity lies in timely prediction of seizures before they manifest. Various attempts have been made using conventional algorithms to accurately predict seizures however, so far, results are not that encouraging. In this work, we applied various machine-learning algorithms, in an attempt to identify complex, multi-dimensional epileptic precursors in brain recordings. Three types of neural networks are used in this feasibility study, namely Multi-Layer Perceptron (MLP) networks, Long Short-Term Memory (LSTM) networks and Gated Recurrent Unit (GRU) networks. The used input data was annotated Electrocorticography (ECoG) data, recorded in living mutant rodents, containing epileptic events at an interval of about one minute. The data was pre-processed for better learning performance by data normalisation and by generating distinctive training features. The neural networks were configured as three-class classifiers, distinguishing among inter-ictal, pre-ictal and ictal periods. A grid-search approach was applied to determine the best set of parameters for the neural networks. Despite our best efforts, the relation between the input data and output data could not be learned in a reliable way. The maximum reached Average Prediction Rate (APR) was 0.57 with a prediction time of 3.1s when using the normalised data as input and 0.65 with a prediction time of 6.1s when using the distinctive features as input. These results essentially signify good detection but virtually no prediction of upcoming seizures. The evaluation of the experimental findings has revealed that the employed ECoG recordings were ill-selected for training our various neural-network models. Also, a non-conclusive exploratory experiment is performed by applying a Weibull Time-To-Event Recurrent Neural Network (WTTE-RNN) on a sub-set of the normalised input data. The experiment has yielded some positive results, a short-notice prediction of the upcoming seizure in some cases, encouraging for further exploration of this approach. Despite the limited success of this work, however, through its extended forensics analysis, it has paved the crucial, initial steps in the direction of seizure prediction.

Fluid accumulation in the human body is, in many cases, a symptom of some underlying pathological condition. Most common examples include edema, ascites, pleural effusion, and internal bleeding. Currently available non-invasive methods to assess fluid accumulation in the body are mainly imaging-based (i.e. ultrasound, computed tomography). This means that monitoring is limited to spot-checks and is only performed when the presence of a particular condition is already suspected from prior clinical evaluation of the patient. There is high clinical interest in the development of a system allowing such monitoring to be performed automatically, continuously over a prolonged period of time, and independently of prior clinical evaluation. For instance, in the case of internal bleeding, hemorrhage indicators such as variation in heart rate and blood pressure are only observable after significant blood loss (up to more than 1 liter) has already taken place, therefore an early warning system can save lives, decrease hospital stay length and significantly reduce complication-related costs. In this thesis, a pioneer solution is proposed for development of such a monitoring system. The hypothesized solution is based on the observation of specific energy-describing features of the Ballistocardiogram (BCG) signal, a measure of the periodic displacements generated on the body as a result of ballistic forces produced by the heart during each cardiac cycle. Because of additional damping generated by the presence of internally accumulated fluid, the energy of this signal is expected to decrease compared to its baseline value. The BCG can be recorded unobtrusively with different sensing modalities on the surface of the body. In order to validate the research hypothesis and assess the feasibility of this novel technique, a custom experimental set-up exploiting several different BCG sensing options has been used to carry out a study on 15 human volunteers. Localized fluid accumulation along the GI tract was induced in a controlled, safe and simple fashion, by means of water intake by the participants, and the BCG signal was recorded before and after intake. Signal feature exploration and performance analysis has been used to develop an optimized feature able to accurately capture the decrease in signal energy due to fluid accumulation, as well as to identify the most suitable sensor type and monitoring body location for the present application. The developed energy-describing feature shows a significant decrease in energy value from baseline to after-intake condition with a p-value<0.001. Moreover, the selected feature was able to correctly identify presence of fluid accumulation with high sensitivity (90% in bed-based, and 100% in standing-position monitoring). Given the promising results of the present study, further research towards improvement and development of the proposed technique is highly encouraged.

Physical activity and mobility are important indicators of the recovery process of patients in the general ward of the hospital. Currently, monitoring mobility of hospitalized patients relies largely on direct observation from the caregivers. Accelerometers have the potential to quantify physical activity of patients objectively and without obstructing their daily routines. Human Activity Recognition (HAR) is a technique used to assess the type of activity an individual subject is carrying out based on sensor readings and has been extensively studied. However, the literature shows that HAR methodologies have been largely developed to recognize activities typical for healthy subjects. This means that activities performed at a slow and irregular pace, such as by a symptomatic patient or an elder, are scarcely considered to design HAR methods. Using HAR for patient monitoring would allow clinically meaningful metrics, such as time spent ambulating or in sedentary behaviour each day, to be obtained automatically. This may offer a convenient solution to enable caregivers automatically monitoring the recovery process of patients.

The aim of this work was to develop an accurate HAR model to recognize activities typical for hospitalized patients. A data collection study was conducted with 20 healthy volunteers in a simulated hospital environment. A single triaxial accelerometer placed on the trunk was used to measure body movement and recognize seven activity types: Lying in bed, upright posture, walking, walking with walking aid, wheelchair transport, stair ascent and stair descent. Features from both time and frequency domain were extracted and used to train three machine learning (ML) classifiers (Naïve Bayes, Random Forest, Support Vector Machine). Additionally, a deep neural network (DNN) consisting of a three convolutional layers and a Long Short-Term Memory layer was developed.

The performance of the DNN model was evaluated on holdout data and compared to the performance of the feature-based ML classifiers. The DNN model reached a higher classification accuracy than the latter approaches (F1-score= 0.902 vs. 0.821). All the models showed a large number of misclassification between the walking with or without walking aid class. By combining these two classes the DNN model reached an F1-score 0.946, compared to F1-score 0.856 of the best feature-based ML approach represented by a support vector machine classifier.

This work shows for the first time the value of applying deep-learning techniques to improve the accuracy of feature-based ML classifiers for addressing the problem of HAR using a single triaxial accelerometer in simulated hospital conditions.

The goal of this bachelor thesis is to develop a control system that controls the temperature of a microthruster. This system also needs to acquire data for research. The microthruster contains a resistor that is used both as a heater and as a sensor. To facilitate the acquisition of data and the testing of the control system, a lab setup with a Keithley 2450 SourceMeter power supply was used. LabVIEW was used to control the power supply and to execute the control algorithms. The final system consists of a microcontroller that runs the control algorithms based on proportional-integral-derivative (PID) control developed in this thesis. The PID values can be adapted with use of the graphical user interface (GUI). A read-out circuit and current supply will be part of the integrated system. These circuits will be developed by other groups that are part of this bachelor project.

Stroke is a cerebrovascular disorder with 15 million cases every year worldwide. The most common symptom is motor deficits. In order to overcome such symptoms, the motor brain either repairs the damaged tissue or reorganises to compensate for the injured brain region. To stimulate this reorganisation transcranial Direct Current Stimulation (tDCS) is considered to be a promising thera- peutic intervention. Simulations of electric field distributions generated by tDCS currently entail individualised volume conduction models to improve tDCS. A volume conduction model includes geometry and conductivity properties of tissue types in healthy subjects. When applying existing models to chronic stroke subjects, electric field distribution patterns differ substantially compared to healthy subject distribution patterns. In current models, the lesion is not identified and acknowledged as a distinctive tissue type, as it is yet unclear what the lesion influence is. However, the lesion is a potential source of variability in desired electric field distribution which could result in different motor recovery. A volume conduction model is designed by combining the software SimNIBS, which can segment the head of healthy subjects and LINDA, able to distinguish lesion tissue of chronic stroke subjects. The location and the conductivity value of the lesion seem to influence the electric field distribution of tDCS where this individualised model is preferred. Including the lesion is an important advance towards the use of volume conduction models for chronic stroke subjects to prospectively find optimal electrode configurations, keep the safety margins and to prospectively analyse the results of tDCS.

In the Netherlands the preterm birth rate is around 7%. The problematic part is that these preterm infants make up for 72% of all deaths during or shortly after birth. To provide a way to aid these infants the following question was stated by the doctor: "turn on a light in the Neonatal Intensive Care Unit when the preterm infant’s development is not as expected". This question was interpreted as following: first of all, find an informative signal present for preterm infants which can provide knowledge on the state of development. Secondly, find a way to quantify this signal through measurement and provide a way to automate the detection. Based on literature the delta brush was selected as a promising waveform in the neonatal EEG, which can be used to follow the maturation of the brain. The spectral behavior of the EEG is used to identify the specific waveforms, but this analysis of the EEG is still mainly based on visual analysis, which is a subjective and time-consuming task. A way to quantify and automate this process was found in the Continuous Wavelet Transform (CWT). Using the Morlet mother wavelet, a time-frequency distribution of the EEG was calculated. The CWT’s redundant behavior was found to provide a finer sampling which could be used to reinforce specific waveform behavior. A third goal was to investigate the troubles that new innovations face in being accepted by the medical field. This was found to be related to the building of trust. The key elements for building trust in this work were identified as: interpretability of the results, transparency of the process, and ease of operation. These elements will be taken into account in the proposed solution. A set of high-density EEG measurements of preterm infants was acquired through collaboration with the the Academic Hospital Antwerp and the Erasmus Medical Center. This high-density consists of 19 electrode channels instead of the more commonly used 11 electrode channels. The measurement was done at 30 weeks postmenstrual age (PMA). These measurements were accompanied with annotation for the waveform of interest. An interactive application for visualization of the EEG data and wavelet coefficients was designed in MATLAB. This application can run standalone and can be used to visualize the detector’s results for the end user. Based on the inherent characteristics of the target waveform two detection features were chosen. The first feature is a measure of the ratio between peak in the high-frequency region (3.3-40 Hz) and the low-frequency region (0.1-3.3 Hz). The second feature is an energy calculation based on a windowed Squared Energy Operator (SEO). A triple-threshold multi-channel detector, based on these features, was initialized and performance was tested over the full threshold range. A Precision-Recall (PRC) curve was provided showing the total solution space for both feature thresholds. Based on the point closest to (100,100) in the training set PRC, a single threshold was selected for validation with the validation datasets. This has resulted in a detector with a sensitivity of 73.66% and a precision of 21.71%. This performance is expected to increase with cross-validation, due to lower quality EEG datasets in the validation set. The differences in performance can be explained by the uncertainties in the exact delta brush behavior and the incomplete annotation set, which causes possible detected delta brush activity to be labeled as false positives. In the end an informative waveform was found and a suitable quantification method has been proposed. A delta brush detector has been designed based on the inherent characteristics of the waveform. A standalone interactive application has been developed to visualize the EEG signal in combination with its time-frequency behavior. The application can also be used in combination with the detector to provide insight into the decision-making process and to visualize the detector’s output.