Epilepsy is a neurological disorder that affects millions of people worldwide and is characterized by recurrent seizures. Managing epilepsy effectively remains a challenge, particularly for patients who do not respond to medication. Closed-loop neuromodulation systems have emerged as a promising approach for real-time seizure detection and intervention, offering the potential to improve the quality of life for individuals with epilepsy. These systems monitor brain activity and deliver therapeutic stimuli when abnormal patterns, indicative of an impending seizure, are detected. However, the effectiveness of these systems depends heavily on the accuracy and efficiency of the embedded seizure detection algorithms.

Implementing brain-computer interfaces (BCIs) for neuromodulation presents significant challenges, particularly in terms of power consumption, area constraints, and computational complexity. The use of machine learning (ML), especially lightweight classifiers, is critical for overcoming these challenges. ML models can be trained to identify complex seizure patterns with high sensitivity and specificity, but their deployment in implantable devices requires careful consideration of hardware limitations.

This thesis addresses these challenges by employing an oblique tree estimator to develop a low-power, low-area seizure detection classifier that relies solely on temporal features to minimize hardware costs. Specifically, the Sparse Projection Oblique Randomer Forest (SPORF) variant is utilized, to enable multiplier-less classifications, along with a multi-path inference approach thereby reducing memory requirements. The proposed model achieves state-of-the-art performance in both patient-specific (98.78%/ 99.89% and 98.52%/99.91% sensitivity/specificity in the CHB-MIT and ETHZ databases, respectively) and patient-independent (92.26%/98.29% in the CHB-MIT database) seizure detection frameworks. Additionally, post-layout simulations of the synthesized design reveal a power consumption of 1.22 µW and an area of 0.0076 mm², with an energy consumption of 7.58 nJ per classification using 40 nm CMOS technology.

Using independent vector analysis (IVA) to analyze and find subgroups in functional magnetic resonance imaging (fMRI) and functional ultrasound (fUS) data requires a lot of manual labour. Recently methods like subgroup identification using IVA (SI-IVA) and IVA for common subspace identification (IVA-CS) have tried to reduce this labour through automation. However, both methods did not test for accuracy. This thesis shows through simulations that these proposed methods are not accurate or robust enough to be trusted and that spectral clustering is a better alternative for automatic subgroup identification. Spectral clustering is then incorporated into the analysis of experimental fUS data of two groups of mice to try and identify these automatically. In this analysis, adaptive constrained IVA (acIVA) was used to incorporate references, further improving the interpretability of the results as components are directly linked to prior constraints.
However, applying subgroup analysis showed that the mice could not be clustered based on their response to the stimuli. Still, spectral clustering is more accurate in the simulations making it a promising alternative for automatic subgroup identification. Furthermore, combining spectral clustering with acIVA makes the results more interpretable due to constrained components not being subject to permutation ambiguity.

At the child brain facility at the Erasmus Medical Centre, multiple tests are performed with children who have one of several disorders. Two of these tests are done with electroencephalogram measurements and are called mismatch negativity and acoustic change complex. After a signal processing pipeline, the EEG measurements from these tests are shown as waveforms called event-related potentials. The goal of these measurements is to see if there is any relation between the waveforms and the disorder, age and other information about the subjects.\newline

These waveforms are measured for each subject, EEG electrode, and in the case of MMN, for different stimuli, which naturally results in a tensor data structure. Algorithms for discriminant analysis and regression that are developed to be applied to tensors are described and altered to take into account the properties of the EEG data. Discriminant analysis can be used to improve classification algorithms that distinguish disorders, while regression can be used to predict variables such as test scores based on the measured data. The algorithms are first tested on simulated data, which shows they can have some merit. Classification rates improve in most simulated cases when the discriminant analysis is applied to the data. Regression can also reliably predict variables when strong correlations are present between the input tensor and output variable. Based on the data from the child brain facility, the discriminant analysis still improves classification rates in some cases, but not as significantly as on the simulated data. Regression using the algorithms described in this thesis is not useful in predicting test scores from other experiments done with the subjects. \newline

The algorithms are also dissected to discover which specific features in the data tensor are weighted heavier by the algorithms. This is done to gain new insights into the differences between the disorders. When comparing the weights that are used for the simulated data with the features that are of importance, there is some relation, but not a very strong one. When the input tensor is however segmented in the time mode, the times of interest can be identified. The regression algorithm also resulted in weights that can be analysed to look at when and where the measurements relate to the output, but this did not show any interesting results. For real data, the segmented tensor resulted in some interesting insights about the differences between the various disorders.\newline

For the current dataset, the discriminant analysis algorithms do improve classification rates, but not by much. The features weighted the most by this algorithm in combination with a segmented tensor might give some insight into the disorders. The tensor regression methods do not work to predict a test score and do not give new insights into the disorders.

The thesis focuses on developing a brain-computer interface (BCI) aimed at differentiating between left-hand and right-hand motor imagery using EEG signals. Its primary objective was to create a scalable and user-specific model for accurately interpreting motor imagery tasks. The study involved the development of one SVM and two advanced deep learning models: the Advanced-EEG-MI-CNN and the Advanced-EEG-MI-CNN-LSTM-Transformer. These models were tested on the Physionet and BCIC IV 2a datasets, as well as on self-gathered data. Results indicated that the SVM model achieved an accuracy of 65% for all subjects and 78% for one subject, while the Advanced-EEG-MI-CNN model achieved an accuracy of 65% for all subjects and 78% for a single subject on average. The Advanced-
EEG-MI-CNN-LSTM-Transformer demonstrated promising results in capturing temporal dependencies, with an accuracy of 80% for all subjects and 92% for a single subject on average. Future work includes enhancing user-specific model implementation, applying Independent Component Analysis (ICA) for cleaner data, and improving data pre-processing to minimize noise.

This thesis details the theoretical background and development process of a classification model for electroencephalogram-based (EEG) motor imagery (MI) signals, to be used in a brain-computer interface (BCI) system. This project was undertaken in order to demonstrate the possibility of distinguishing MI-EEG signals acquired using the g.tec Unicorn Hybrid Black EEG measurement cap. The classification model devised in this thesis is a hybrid deep neural network model, which combines a convolutional neural network (CNN) and long short-term memory (LSTM) recurrent neural network (RNN) in parallel, closing with a fully-connected (FC) layer. Much experimentation and research was needed to create this model, and this is extensively discussed within the thesis. The classification model produced for this thesis is one part of a complete end-to-end BCI system, which also entails a measurement and data processing procedure, and the development of a graphical user interface which provides visual feedback to the user.

On publicly available datasets, the classification model produced promising performance, achieving 55% on the BCI Competition IV-2a (4 class) dataset and 78% on the BCI IV-2a (2 class) dataset. The model has not yet been tested on self-collected data.

This thesis investigates whether an EEG headset can be used to distinguish motor imagery signals in real time for a Brain Computer Interface (BCI).The specific EEG headset used for this project is the gtec Unicorn Hybrid Black. The aim of this subgroup is to stream the data in real time, preprocess the data, and create a training dataset using recordings from subjects. The sample from the EEG headset is streamed to a laptop using a live data streaming framework called Labstreaminglayer. The data is then filtered using frequency filters and ICA to remove noise and artifacts. Finally, ERDS plots are used to check the signal quality of the recordings. Recordings of sufficient quality should have (de)synchronisation peaks after the prompt is displayed. Before the recordings are made, an experimental setup is set up. This includes prompts with four different movements that the subject is asked to imagine. The data is sent to a GUI to be visualized with various graphs. It is also sent to a machine learning model to classify the movements. It was concluded that the subsystem could successfully stream data from the cap, process the data, and verify the quality of the data. Using ERDS plots, it was verified that some, but not all, MI actions are distinguishable for one individual. However, further verification of other individuals is required to conclude whether this is a systematic problem or is due to the individual.

The main goal of this project is to utilize a commercially available OpenBCI Ultracortex IV for the measurement of Electroencephalogram(EEG) signals. A pipeline consisting of preprocessing, classification and extraction is employed to transform the motor execution EEG signal into a singular Left or Right output. This output is then further displayed on an Interface that offers the option to either calibrate or play a simple game.

The Ultracortex and relevant software were used to determine the sensor layout, with the placement of the sensors focused on areas which exhibited high cortical activity during motor execution. Experiments were strategically designed to optimize our chance of successful readings and OpenVIBE was used in conjecture with preprocessing filters to save the raw and filtered data which was further sent to the Machine Learning group.

The collected data was analyzed through Spectrograms, Power Spectral Density(PSD) and Event-Related Desynchronization/Synchronization(ERDS) plots. The analysis aimed to confirm whether the desired activity occurred and whether the observed patterns resemble those documented in other research papers.

The data from the headset is live-streamed to the interface via Lab Streaming Layer(LSL) where it undergoes further filtering before being sent to the Machine learning group. This process was done through python libraries which then allowed for efficient and effective communication between the other groups.

In this thesis, the development of a graphical user interface for a brain computer interface (BCI) system is discussed. This system is based on electroencephalographic (EEG) signals from motor imagery (MI). BCI applications for low consumer-grade are very limited, since most of the applications are towards medical use or games. This limits the potential BCI can have, since it only applies to one specific environment. The goal of this thesis is to show that a low consumer-grade BCI can be made, ranging from educational to personal use, with the needed quality and usability. This is achieved by showing the possibility of distinguishing MI signals. Specifically in this thesis, left hand, right hand, tongue and feet. This can be done and utilized through a machine learning algorithm. Presentation methods for this are displaying EEG signals and using demos that show the purpose of the BCI system. An example of such a demo is moving a simulated computer cursor. In order to have this system for a wider public, the interface should be able to adapt to each individual user using personalized machine learning models. Due to the distinct, personal EEG patterns, the accuracy of classifying the MI-EEG signals using personalized ML models is much higher than with an ML model trained solely on a public dataset. For more advanced personal use of the system, the interface contains everything needed regarding EEG signals, including real-time data streaming in order to directly see change in EEG signals and to help the user detect potential errors in the measurement setup

The purpose of this thesis is to study the possibility of developing a computer interface that makes use of electroencephalogram (EEG) signals in order to improve the interaction between humans and computers. The major objective is to develop a system that is able to read brain activity in real time and then transform that information into instructions that can be executed on a computer. This study combines electroencephalogram (EEG) technology with sophisticated machine learning algorithms in order to develop an interface that is fluid, responsive, and user-friendly. Among the most important goals are the development of methods for robust signal processing, the design of an interface that is easy to use, and the installation of tools for real-time data presentation and analysis. A comprehensive set of tests was performed on the system, which revealed considerable improvements in terms of accuracy, responsiveness, and the overall user experience. This research makes a significant contribution to the expanding field of neurotechnology by providing an interface that is simple to use and may be used as an inspiration for a broad variety of purposes, including medical diagnosis and treatment, as well as entertainment and other games.

The brain stands as the most powerful processor in the known universe. It generates a continuous stream of electrical and chemical signals that underpin every thought, sensation, and action. Our past efforts in decoding these signals have made it possible to diagnose and treat many neurological disorders, helped us gain a deeper understanding of cognitive processes and consciousness, and paved the way for brain-computer interfaces. To take another step forward in our long but rewarding journey of discovering the brain's complex organization, we rely on advances in imaging technologies and signal processing.

Functional ultrasound is a neuroimaging technique that has emerged in the last decade, and has gained remarkable attention since then. The popularity of this technique stems from its portability, high resolution and affordability. Functional ultrasound can detect subtle fluctuations in local blood dynamics, which serve as delayed indicators of the underlying changes in neuronal activity. The goal of this thesis is to develop novel signal models and processing algorithms that can reveal the spatial and temporal characteristics of hemodynamic activity induced by external stimuli using functional ultrasound.

Existing techniques that explore how the brain reacts in response to stimuli model the design variables using a linear time-invariant system with binarized input representations, marking when a stimulus is on or off. However, experimental evidence suggests that the brain reacts in a more intricate manner. While some regions exhibit consistent responses to repeated stimuli, others can show substantial variation even when exposed to the same stimulus seconds apart. In our in-vivo experiments, we particularly focus on key regions within the mouse visual processing pathway, which are analogous to those in the human brain. We track how visual information flows across these areas, and propose methods that can incorporate the spatiotemporal variability of brain responses when identifying evoked activity. Using these methods, we show that functional ultrasound can capture the dynamic nature of brain responses with high spatial and temporal resolution, and provide us with further insights into the functional organization of the brain. Future directions of this dissertation include multimodal processing of the functional ultrasound signal together with neuronal activity, aiming to enhance our understanding of neurovascular coupling.@en

Brain disorders in children pose significant challenges to their development, impacting cognition, speech, movement, and behavior. The uncertainty surrounding prognostic information at the time of diagnosis leaves families with numerous questions about the future. The Child Brain Lab at Erasmus MC Sophia Children's Hospital conducts IQ, electroencephalogram (EEG), speech, and movement tests in playful environments, enhancing scientific research and healthcare practices for a better understanding of disease progression.

The Otolaryngology department at the Child Brain Lab focuses on auditory-related potentials (ERPs) obtained from EEG measurements to predict the future development of children with brain disorders. Analyzing ERP data from experiments like Mismatch Negativity (MMN) and Acoustic Change Complex (ACC) yields insights into developmental trajectories and connections between hearing, language, and brain development.

This thesis aims to explore alternative methodologies for extracting comprehensive information from ERPs, overcoming limitations of the commonly used peak amplitude and latency analysis. Tensor decompositions are employed to exploit structural information present in the data, using data fusion methods to combine multiple datasets for improved classification and deeper insights into group differences.

Simulations on artificial ERP data demonstrate that data fusion methods perform better on two ERP tensors compared to single tensor decomposition when group differences are shared between datasets. On a real dataset, tensor decompositions show promise for classifying subjects based on auditory event-related potentials while giving more insights into the neurological sources.

This report proposes an alternative method for analyzing ERP data, highlighting the potential of tensor decompositions and data fusion techniques.

Micro-Doppler (µDoppler) ultrasound imaging is a high frame rate ultrasound imaging modality that provides high spatiotemporal resolution ultrasound images of blood flow. It is sensitive to slow blood flow and particularly suitable for capturing fast-changing phenomena like rapid blood flow. Clutter filtering is an essential step in µDoppler data processing to reject tissue clutter signals and keep blood flow information as much as possible. 3D freehand µDoppler imaging is an emerging ultrasound technique that can construct full spatial vasculature images with a panoramic view that conventional 2D ultrasound is not able to provide. As freehand implies the continuous and nonuniform movement of the probe, it becomes more challenging for clutter filtering to acquire high-quality images.

This thesis explores and compares different state-of-art clutter filtering techniques on freehand in-vivo µDoppler imaging of the human brain. Specifically, Singular Value Decomposition (SVD), Robust Principle Component Analysis (Robust PCA), and Independent Component Analysis (ICA) clutter filtering techniques have been investigated. The aim is to test and compare their performance on in-vivo µDoppler ultrasound data with freehand probe movement and understand how freehand motion affects the threshold selection criteria. Besides that, a newly proposed method that combines ICA clutter filtering and clustering is included in this thesis to bring another perspective for sorting independent components corresponding to blood flow and rejecting unwanted ones consisting mostly of tissue clutter signals.

Purpose
The main purpose of this report is to find out whether the OpenBCI "Ultracortex Mark IV" Electroencephalogram (EEG) headset is capable of differentiating EEG-signals of motor execution from neutral state with recorded data and to find out whether it can differ motor executions between left and right hand. Next to that, it is to be determined whether the OpenBCI headset was the optimal one for this purpose.

Method
First, the specifications of different headsets were compared. Afterwards, a montage of the electrodes was designed to detect motor execution and motor imagery, mainly centered around the locations C3, Cz and C4, on the top of the scalp. The software "Openvibe" was used to extract data from the headset during experiments and to record it in a csv file. A subject was asked to follow a video with a sound cue followed by a visual cue instructing to move either its left hand or right hand.


Result
Merging the left and right hand trial data together, the result is that the headset shows in the alpha band (7-12 Hz) mostly a decrease (ERD) in magnitude around the visual cue, sometimes followed by a bigger increase in magnitude (ERS). Looking at the extremes after the cue, it is seen that mostly the difference in magnitude is around a factor 1.5 compared to the average magnitude of before the visual cue. Splitting the trial data between left and right hand, similar results can be seen, but one hand produces slightly more ERD or ERS than the other hand depending on the position of the electrode on the left or right hemisphere of the brain.

Conclusion
The OpenBCI headset can in fact detect a difference between movement of the hands and the neutral state. Differentiating between the movements of left and right hands seems possible from the results, but the difference in the signal of left and right hand is minimal. It is recommended to repeat the experiment with more trials and different subjects to get a more solid conclusion.

In the context of designing a real-time brain-computer interface for playing a game using the OpenBCI Ultracortex "Mark IV" headset, this paper focuses on the work of the decoding subgroup. The primary responsibility is to analyse EEG data retrieved from the OpenBCI headset and classify the intention of the user. Our objective is to achieve a high-accuracy classification of the EEG signals. The paper is structured into three main sections: preprocessing, feature extraction, and classification. Multiple methods for preprocessing and classification of motor execution EEG signals will be analysed, striving to contribute to the real-time implementation of the project. The results of our work provides valuable insights for future research and development in this field.

This document presents the development of a user interface for an EEG motor imagery based Brain-Computer Interface (BCI) as the interface subgroup. The aim of this subgroup in the project was to design and implement a graphical user interface (GUI) incorporating visual neurofeedback to enhance the accuracy of the decode algorithm, developed by the other subgroup, during the calibration/training stage for the user such that the user will have better motor imagery control using the GUI. The GUI features two interactive games, namely pong and breakout, and incorporates topographic maps displaying the user’s brain activity. The primary goal of these maps was to improve the training process of the algorithm for each individual user. Additionally, the thesis explores the feasibility of incorporating steady-state visually evoked potential (SSVEP) elements in the games or for creating a new game that combines motor imagery and SSVEP elements allowing for more complex games to be played thus positively affecting the user experience.

To better understand how brain signals are processed and even how the human mind works, analyzing the hemodynamic signal model is one of the most essential steps. In the CUBE group of Erasmus MC, functional ultrasound (fUS) data of a mouse’s brain is recorded. By using this fUS dataset, this thesis will solve the problem regarding the joint estimation of hemodynamic response function (HRF) and the underlying stimulus. Usually, hemodynamic responses are investigated in the time domain, while this thesis provides another perspective
from frequency domain signal processing.

We consider the hemodynamic response as a convolutive signal mixture, then try to transform it into an instantaneous mixing model by converting the context into the frequency domain. By applying independent vector analysis (IVA), this estimation problem can be solved without facing permutation ambiguity which is a well-unknown problem regarding independent component analysis (ICA). Additional steps before and after IVA are also discussed so that a whole estimation road map is formed.

Both simulation and experimental analysis are provided to validate this estimation algorithm. Results show that by using this method, both stimulus and HRF estimation can be achieved satisfyingly in a suitable experimental setting. This thesis provides insights and future potentials for IVA to be further investigated in neural signal processing problems.

In recent years, the increase in brain research led to the development of large-scale brain imaging techniques. With large-scale brain imaging techniques, such as functional magnetic resonance imaging (fMRI), functional connectivity analyses have shown altered connectivity patterns in humans and mice with neurobiological disorders, such as autism spectrum disorder (ASD). To further investigate different mutations that contribute to ASD, a behavioral neuroscientific experiment has been performed at the neuroscientific department of Erasmus MC. During the trial, brain activity in groups of wild-type and homozygous mice is measured using functional ultrasound (fUS) while allowing for spontaneous behaviors of mice. The homozygous mice lack the SHANK2 protein, resulting in hyperactivity and autistic-like behavioral alterations associated with ASD in humans. Understanding the origin of ASD is key to providing effective treatment. However, with the introduction of novel large-scale brain imaging techniques such as fUS, new methods have to be developed that enable functional connectivity analyses. Furthermore, new insights might also be acquired from dynamic functional connectivity analyses, in which changes in functional connectivity over time are evaluated. Subsequently, the question remains if it is possible to unravel differences in brain dynamics between wild-type and homozygous mice using a dynamic functional connectivity analysis.
First, an fUS data model is developed to model how fUS signals arise from a generative perspective. This model comprises a combination of a convolutive and a state-space model. Subsequently, inference of functional networks and their temporal dynamics can be performed. Also, a pre-processing pipeline for experimental fUS data is designed to reduce problem complexity and data cleaning. The performance of the developed methods is evaluated on the experimental data set, where a difference in brain dynamics between wild-type and homozygous mice is investigated.
It is found that a deconvolution procedure using the non-negative least absolute shrinkage and selection operator (NNLASSO) is necessary to reconstruct the underlying activity of neural populations. After that, using the hidden Markov model (HMM) as a state-space model, it is found that functional networks and their temporal dynamics can be learned from fUS data using expectation maximization (EM). It has been discovered that the developed methods consistently decompose reconstructed neural activity into biologically plausible functional networks from experimental fUS data. Also, with 96% certainty, a difference in brain dynamics between wild-type and homozygous mice is found using this method.
In summary, in this thesis, novel methods are developed to perform a dynamic functional connectivity analysis on experimental fUS data. Also, by performing such dynamic functional connectivity analysis for the first time on fUS data, a consistent decomposition of reconstructed neural activity into biologically plausible functional networks and a possible difference in brain dynamics between wild-type and homozygous mice are found. This research highlights the potential of fUS as a large-scale brain imaging technique in the quest to understand the origin of ASD and other neurobiological disorders.

Background & Objective
Cardiovascular diseases (CVDs) are the leading cause for death globally nowadays. Pulse wave velocity (PWV), a marker of arterial stiffness, is an important predictor of CVD risk. In precedent work, carotid artery data was collected with ultrasound to estimate the PWV with a digital signal processing (DSP) pipeline. As a potential alternative to the DSP-based approach, this thesis studies the applicability of machine learning(ML) for the estimation of carotid artery motion (diameter, distension, etc.) and explores to what extent neural networks can exploit the ultrasound data to extract relevant biomarker information.
Methods:
This thesis proposes a ML pipeline that processes the ultrasound data in a different perspective than the DSP approach. The ML pipeline consists of four modules (neural networks & post-processing) to: 1) segmentation based on cardiac cycle (CC), 2) detect the region of interest (ROI) of artery in the ultrasound data, 3) tracking the artery diameter and 4) post processing to estimate cardiac parameters e.g. pulse arrival time (PAT), an essential part of PWV estimation. Exploiting the features of the artery-lumen structure and time-evolving characteristics of collected ultrasound signal, the designed ML pipeline can acquire cardiac markers spatially and temporally with irregular kernels and sliding mechanism, decompose the complicated estimation into compact sub-modules.
Results:
The results show that the ML approach can successfully estimate the artery diameters and reserve important waveform features (max-slope moment) of the artery diameter, and can infer the CC markers without ECG data as a segmenting event for heart cycles. Thus, the PAT can be computed as the time difference of the max-slope moments of inferred artery diameter and detected CC markers. According to the numerical results, the PAT can be estimated on the average for an ultrasound data recording (120s), and the correlation coefficient of label PAT (computed from estimated parameters of DSP pipeline and ECG data) and estimated PAT (ML pipeline) is 0.8250. This indicates a good correlation and hence proves the effectiveness of the mean PAT estimation.
Conclusion:
In conclusion, the proposed ML pipeline can effectively estimate mean PAT, and demonstrate the feasibility to estimate PWV as a relevant cardiovascular marker using only ultrasound data of carotid arteries. Apart from the PAT, the heart rate can also be possibly tracked via intermediate results of the ML pipeline (CC markers). From the future perspective, the potential of phase information in the raw ultrasound data and further optimization are worth exploring, and the extension to hardware (e.g. chips, embedded system) can be implemented as a practical application.

Background and Objectives: Wearable devices (WDs) capable of recording electrocardiograms (ECGs) for prolonged periods in ambulatory settings offer the possibility of detecting non-predictable events such as epileptic seizures and atrial fibrillation. Nevertheless, these systems suffer from additional noise sources, such as movement artifacts (MA).
Several adaptive (AF) algorithms have been proposed in the literature to suppress movement artifacts from ECG without consistent results. Adaptive algorithms for signal enhancement require a reference signal correlated with the noise and not correlated with the signal of interest. This correlation can change significantly depending on the absolute and relative location of the electrodes and the location of the reference sensor (i.e., accelerometer, gyroscope); objectively measuring the correlation is not a trivial problem.
For this reason, first, we used an algorithm to obtain a rough estimate of the movement artifacts from the recorded ECG to calculate the correlation between them and the available reference signals (three-axis accelerometer and three-axis gyroscope) and selected the one with the highest correlation. Then, we compare three adaptive filter algorithms using the signal-to-noise ratio (SNR) coefficient as the evaluation parameter.
Methods: For testing the implemented adaptive filters, first, we used a simulated signal, then data from an openonline database, and last, a single lead ECG wearable device called AFi® (Praxa Sense™, The Netherlands) with an embedded IMU. To induce the movement artifacts in a controlled setting, participants performed a set of predefined movements within three intensities; high (running, jumping), moderate (torso rotations, pushups), and low (walking). Then we analyzed the recorded data offline as follows:
1. Test the correlation between noise and the IMU components, and select the component with the highest correlation to be used as a reference input for the adaptive filters.
2. Compare three adaptive filters in terms of SNR improvement; the Least means squares (LMS), the Normalized least means squares (NLMS), and the Recursive least squares RLS.
3. Filter the selected reference input with wavelet decomposition, and test if there is a filter performance improvement in SNR.
Results: The implemented adaptive filters performed as expected with the simulated signals, but they showed very poor results once we used them on real data.
The RLS filter showed superior performance than the least mean squares-based filters in terms of convergence speed and the root mean squared error minimization. Nevertheless, it requires a high correlation (ρ) above ρ>0.8 between the reference input and the undesired signal or noise to provide a proper signal enhancement and morphology recovery.
The low correlation between the movement artifacts and the components of the IMU used as a reference input for the adaptive filters affected the filter performance heavily. Filtering the reference input with the wavelet decomposition did not improve the correlation or the filter performance.
Conclusions The correlation between ECG motion artifacts and movement recorded with inertial sensors appeared to be low and inconsistent. Given this, adaptive filters using inertial sensors as reference input are unsuitable for removing ECG movement artifacts.