6 thesis projects within the wearable data observatory
Mitra Baratchi
Title: Sensing movements with accelerometers (tremor, gait, multi-sensor measurements)
Subtitle: Quantification of Parkinson’s disease severity using wearable accelerometry
Intro:
Acceleration data has the potential to reform diagnostics and prognosis through their accuracy and cost-effectiveness for several diseases such as Parkinson’s disease. The objective of this project is to develop an accelerometer-based algorithm that will reflect the severity of Parkinson’s disease symptomatology. In particular, the research will focus on quantifying the severity of tremor as well as detecting the presence/absence of bradykinesia and dyskinesia. The dataset is available for different body positions, therefore a preliminary analysis needs to be implemented to detect the body position in which the mounted sensor gives a higher accuracy in the result.
Sensors available for this research: Accelerometers
Datasets: Publicly available and CHDR trial datasets
** The data that will be used for this project are from devices located in several body locations (forearm, shank, back, wrist, waist).
General research questions:
This internship focuses on the development of accelerometer-based algorithms that can identify and quantify the severity of tremor as well as detecting the presence/absence of bradykinesia and dyskinesia.
Furthermore, the internship investigates which sensor gives more accurate results on identifying the different characteristics of Parkinson’s disease (e.g., resting tremor, postural tremor, action tremor, gait).
Evaluation approach: Standard machine learning performance metrics
Steps of the research project:
This internship starts with a literature review focusing on existing studies that use accelerometer data for quantifying tremor levels. Second, the data provided by an online available database as well as data from CHDR trials will be analyzed using (machine learning) algorithms for development and optimization of the system. Furthermore, the model will be evaluated by the different positional sensors (forearm, shank, back, wrist and waist) to see which sensors result in a higher accuracy. Findings of the experiments will be summarized in a final report and presented at the CHDR.
Profile of the candidate:
- A strong and demonstrable interest in programming
- A strong commitment to research: investigate the state-of-the-art/literature; evaluate ideas and solutions against relevant dataset; be patient in working with real data
- Independent, self-motivated, reliable, and eager to learn.
Related papers:
- “Analysis of Subtle Movements related to neurodegenerative diseases using wearable inertial sensors: a study in healthy subjects”, Martinez-Mendez Rigoberto, 2013, 10.1109/EMBC.2013.6610949.
- “The role of accelerometer and gyroscope sensors in identification of mild cognitive impairment”, Migyeong Gwak, University of California, 2019, 10.1109/GlobalSIP.2018.8646622.
Contact info:
Eleni Kapousidou (ekapo...@chdr.nl) and azhuparris (azhup...@chdr.nl)
Supervisors:
Daily CHDR supervision: Eleni Kapousidou (ekapo...@chdr.nl) and Vasileios Exadaktylos (VExada...@chdr.nl)
Daily LIACS supervision: Mitra Baratchi, Maedeh Nasri
___________
Title: Motion sensor calibration and compensation algorithm based on physical-aware modeling.
Intro: Sensors in wearables for the consumer market have specific calibration needs if they are used in a research setting where reproducibility of the findings based on these sensor values are of importance. Most sensors suffer from individual differences in offset and spread which on top of this also depends on the orientation of the sensor compared to the surface plain of the earth. Further, most wearables use time-sliced operating systems which cause irregular timestamps of their samples. Most practitioners in the field use standard resampling techniques or apply algorithms originally designed for similar problems found in astronomy. Others apply high - and low pass filters without any resampling nor correction. The focus of this study is to find a compensation algorithm to rule out all of these errors mentioned.
As a suggestion, an approach for compensation could be to use physical-aware modeling, make a physical model of acceleration and angle velocity of well-known movements and use them together with the real motion sensor values in the wearable to come to an AI model for compensation.
Remark: this MSc project proposal can be combined with the next one since this proposal can use the developed calibration method.
Figures: One idea is to see if it is possible to use a robot (if a suitable one is available) to implement a well-designed set of pre-defined movements of the wearables. Because the movements are well known, a physical model of the robot can be developed to calculate the expected values of all motion sensors, i.e. A,G and M, regularly sampled to be compared to the measured ones A’, G’ and M’. The challenge is to find a nice transformation function AI() with AI(A’,G’,M’) = (A, G, M).
See below a robot arm with a black bar B (the inlay shows what is in the bar) of wearables which are attached to its outer moveable point. By rotating several joints of the robot, you can design the movements of the wearables in all possible positions.
Datasets/sensors available for this research:
● Wearables with motion sensors are available to get raw sensor recordings.
● You have to arrange a robot to start with the Media Technology Lab at LIACS if you feel the need for this study.
General research questions:
● Is there a way to compensate for individual and structural distortions (i.e. calibration algorithms) in motion sensors found in wearables meant for the consumer market?
● Can this compensation be done without explicit zero base measurements?
● How important is the correction of irregular timestamps of samples of the sensors for human activity research / reproducibility?
● At what extent will calibration influence reproducibility of research in the human activity domain?
Steps of the research project:
- Search for a benchmark data set or if not there, create a data set such that it can potentially become a benchmark.
- Search for similar attempts in literature, compare and set new research directions about the following subjects:
- What are the benchmark data sets available in the literature for the same purpose?What are the baseline methods that have been implemented and can be used in this project for comparison?
- Develop an improved algorithm compared to the baseline found to correct the measured motion sensors using the benchmark found. If you created the data set, you need to run these baseline algorithms also on this data set.
Evaluation approach:
Profile of the candidate:
● Interest in programming in Python.
● Interest in investigating the state-of-the-art in scientific literature.
● Interest in creating a data set which could become a benchmark data set.
● Interest in doing experiments with a robot arm or similar.
● Enthusiasm and a wish to be trained in getting an academic mindset.
Related papers:
- Here is a recent paper with calibration with a neural net, “Two-axis accelerometer calibration and nonlinear correction using neural networks ...”, Mario A. Soriano, University of Alberta, 10.1109/TIM.2020.2978568.
- Zimmermann, T., Taetz, B., & Bleser, G. (2018), “IMU-to-segment assignment and orientation alignment for the lower body using deep learning.”, Sensors, 18(1), 302,
- Xiao, X., & Zarar, S. (2018, May). Machine learning for placement-insensitive inertial motion capture. In 2018 IEEE International Conference on Robotics and Automation (ICRA) (pp. 6716-6721). IEEE.
- For the stability of robot arms, calibration of its sensors are of importance: “Automatic calibration of a motion capture system based on inertial sensors for tele-manipulation”, Jorg Hoffmann et al., ICINCO 2010.
- Most techniques for calibration of motion sensors involve the use of high-low pass filters. This paper claims to be twice as accurate and does this with an auto calibration technique, “Inclination measurement of human movement using a 3-D accelerometer with auto calibration”, H.J. Luinge, University of Twente, 2004, 10.1109/TNSRE.2003.822759.
- Calibration with a robot as shown above with five methods, “Separating Movement and Gravity Components in an Acceleration Signal and Implications for the Assessment of Human Daily Physical Activity”, Vincent T. van Hees, Institute of Metabolic Science, Cambridge, 2013, doi:10.1371/journal.pone.0061691
- An alternative for a robot, a 3-D printed polyhedron, “A low-cost calibration method for low-cost MEMS microelectromechanical system accelerometers ...”, Jesus A. Garcia, 2020, Journal of Sensors, 10.3390/s20226454.
<MSc student name>, Maedeh Nasri, Richard van Dijk, Mitra Baratchi, “Benchmark data set with baseline algorithm for calibration of consumer wearables”, 2022, Journal of Sensors, ...
Contact info:
● Mitra Baratchi m.bar...@liacs.leidenuniv.nl, assistant professor at LIACS
● Richard van Dijk, m.k.va...@liacs.leidenuniv.nl, postdoc research software engineer LIACS
● Maedeh Nasri, m.n...@fsw.leidenuniv.nl, PhD candidate at FSW
Supervisors:
● Joost Visser, full professor at LIACS, co-supervisor, software engineering and data science
● Mitra Baratchi, supervisor, artificial intelligence and machine learning
● Richard van Dijk, daily supervisor, software engineering and data science
● Maedeh Nasri, daily supervisor, researcher of SchoolYard project, center of BOLD cities.
_______________
Title: GPS refinement measured with wearables with support of its motion sensors and map references in small scale environments.
Intro: In recent urban social analysis studies, wearable technology such as smart watches has been widely used in different small scale environments such as offices, healthcare hostels, sport centres and schools. These wearables are equipped with motion sensors such as the triaxial accelerometer sensor (force on the body), triaxial gyroscope sensor (angular velocity of the body) and triaxial magnetometer sensor (move direction of the body) and GPS sensors for outdoor positioning of real life subjects.
These GPS sensors suffer from low accuracy in these small scaled spatial environments, see the figure below. The question is how to use motion- and GPS sensor recordings together with known locations of map references to obtain a more accurate positioning. Map references can be physical boundaries of the small environment such as pavements and fences or places where the real life subjects frequently gather around, like a bench in a park.
For this study a new data set needs to be created if a benchmark data set is not available.
Figure shows inaccuracies in small scale environments when the edge of a circle with 10 and 100 meter radius and center point positioned on a map (the green flag), is crossed with a GPS device moving on a well known grid (the red points show the measured cross points, which should laying on the edge of the virtual circle). Because the radius and position of the circle is known, you can predict where the red points should be measured and so corrected.
Datasets/sensors available for this research:
● Wearables with GPS and Motion sensors are available to get raw sensor recordings.
● A calibrated location measuring system with high accuracy (~10 cm, ultrasound or ultra wideband) can be made available around the beginning of 2022 to get the ground truth, but a rope pinned on the ground will do as well.
General research questions:
● Is it possible to correct GPS coordinates to get a higher accuracy in the range of 50 cm with the use of the motion sensors accelerometer, gyroscope, magnetometer and use of these map references as described above?
● Motion sensors can be used to get the GPS error in direction and distance when the real life subject is not moving and resides in a reference point. Could you apply this GPS error knowledge to other moving subjects close by in order to improve their correct positions? For more exploration of this thought, see the figure below. The non-moving real life subject (the Beacon GPS) records its GPS coordinates at certain times, x0, x1, x2, …, x6. The vector between this Beacon GPS and the correct, well known GPS coordinates of this Beacon are used for a trajectory colored in black, to have it corrected with the same vector. With motion data, sub trajectories in green around each GPS measurement point are determined and entirely moved to its proper position found on the real black trajectory the real life subject was moving on.
● What motion sensors are most important, have the strongest influence? And how about the placement of the wearable on the real life subject, wrist or waist?
Bonus research questions:
● What is the influence of calibration of the motion sensors for the GPS refinement?
● Are outdoor map references with known GPS coordinates detectable, to what extent, and so useful for the aforementioned GPS refinement? For example by use of a stop-move detection in GPS trajectories.
Steps of the research project:
● Search for a benchmark data set or if not there, create a data set such that it can potentially become a benchmark.
● Search for similar attempts in literature, compare and set new research directions about the following subjects:
○ What are the benchmark data sets available in the literature for the same purpose?
○ What are the baseline methods that have been implemented and can be used in this project for comparison?
● Develop an improved algorithm compared to the baseline found to correct the measured GPS points based on motion sensors and map references using the benchmark found. If you created the data set, you need to run some of these baseline algorithms also on this data set.
Evaluation approach:
● What are the evaluation metrics/methods in the literature that have been used for the same purpose?
Profile of the candidate:
● Interest in programming in Python.
● Interest in investigating the state-of-the-art in scientific literature.
● Interest in creating a data set which could become a benchmark data set.
● Interest in doing data collections outdoors with the help of real life subjects.
● Enthusiasm and a wish to be trained in getting an academic mindset.
Related papers:
For inspiration see an excellent MSc thesis made by students of the School of Health and Society, Department Computer Science Embedded Systems in Malmoe, Sweden, in 2011 with as title “Indoor positioning using sensor-fusion in Android devices” by Ubejd Shala and Angel Rodriguez to get warm.
For applying variations of Kalman filters, see the MSc thesis of M. de Vries Delft University of Technology published in 2019 with the title “Multi-rate unscented Kalman filtering for pose estimation”, or have a look at the book written by his peer Manon Kok, “Using inertial sensors for position and orientation estimation”, arXiv:1704.06053v2, 1704.06053.pdf (arxiv.org).
For an example of a map matching algorithm evaluated on existing benchmarks including software, see a recent paper of Adam Millard-Ball from the University of California published in 2019 with title “Map-matching poor-quality GPS data in urban environments: the pgMapMatch package”, https://doi.org/10.1080/03081060.2019.1622249. The corrected version can be found online https://doi.org/10.1080/03081060.2019.1631576.
Or take the paper of Nabil M. Drawing from the University of Waterloo, published in 2013 with the title “GPS localization accuracy classification: a contact-based approach”, https://doi.org/10.1109/TITS.2012.2213815.
<MSc student name>, Maedeh Nasri, Richard van Dijk, Mitra Baratchi, “Benchmark data set with baseline algorithm for GPS refinement in small scale environments”, 2022, Journal of Sensors, ...
Contact info:
● Mitra Baratchi m.bar...@liacs.leidenuniv.nl, assistant professor at LIACS
● Richard van Dijk, m.k.va...@liacs.leidenuniv.nl, postdoc research software engineer LIACS
● Maedeh Nasri, m.n...@fsw.leidenuniv.nl, PhD candidate at FSW
Supervisors:
● Joost Visser, full professor at LIACS, co-supervisor, software engineering and data science
● Mitra Baratchi, supervisor, artificial intelligence and machine learning
● Richard van Dijk, daily supervisor, software engineering and data science
● Maedeh Nasri, daily supervisor, researcher of SchoolYard project, center of BOLD cities.
__________________
CHDR internship
Daily supervision: Eleni Kapousidou (ekapo...@chdr.nl) and Vasileios Exadaktylos (VExada...@chdr.nl)
Daily LIACS supervision: TBD
Responsible RD: Geert Jan Groeneveld
Education Director: Jeroen van Smeden
Period: 6-9 months; Starts immediately
Fulltime: yes
Financial compensation: yes
Title: Skin Disorder Detection System
Research description:
CHDR MORE is a highly customizable platform which allows remote monitoring of patients and trial subjects, data ingestion, and data management. The MORE app enables unobtrusive data collection from multiple smartphone sensors (e.g., location data, accelerometers, and ambient light) as well as meta-data from phone usage logs (e.g., app usage, calls, and texts).
Being able to quantify skin disorder conditions objectively and reliably would allow for the assessment of pathology severity and treatment results among patients suffering from a dermatologic disorder. The objective is to develop a computer vision-based machine learning algorithm that can distinguish skin disorders and identify changes after treatment (e.g., size). Specifically, the system will rely on image data to define the most significant image-based features for the creation of a reliable system.
This internship focuses on the development of a computer vision-based skin disorder detection method for use in clinical trials. This internship starts with a literature review of existing methods to analyze skin abnormalities. Second, existing data will be analyzed using (machine learning) algorithms for the development of a skin disorder detection algorithm. Findings of the experiments will be summarized in a final report and presented at the CHDR.
Products:
Literature review, report/thesis, two presentations (one after a few weeks, one at the end of the internship)
_______________________
CHDR internship
Daily supervision: Eleni Kapousidou (ekapo...@chdr.nl) and Vasileios Exadaktylos (VExada...@chdr.nl)
Daily LIACS supervision: TBD
Responsible RD: Geert Jan Groeneveld
Education Director: Jeroen van Smeden
Period: 6-9 months; Starts immediately
Fulltime: yes
Financial compensation: yes
Title: Gait analysis by accelerometers
Research description:
The walking pattern (gait) of an individual is a variable that can change if a person suffers from some disease, had a surgery, takes medication etc. Alterations in step length, step velocity, step time and many more gait features can be easily detected and indicate changes from normal values. Thus, gait analysis has the potential to be an objective and reliable tool used in clinical research to detect changes in one’s gait pattern.
This project focuses on the use of accelerometers, located in different parts of the body (upper limb and back of the pelvis), to estimate gait features. To validate the accelerometer-based gait features, the features will be compared to insole pressure sensor-based gait features (the gold standard). The objective and challenge of such a project is to create a reliable and clinically relevant system with accelerometer data as input that in the end will be able to detect and recognize gait patterns accurately and reliably.
This internship focuses on the development of gait features using accelerometers for use in clinical trials. It starts with a literature review focusing on designing a pilot-study that evaluates the optimal configuration to obtain and process data from sensors (accelerometer, pressure sensor). Second, the collected data will be analyzed using (machine learning) algorithms for the development of an algorithm. Findings of the experiments will be summarized in a final report and presented at the CHDR.
Products:
Literature review, report/thesis, two presentations (one after a few weeks, one at the end of the internship)
_________________CHDR internship
Daily supervision: Ahnjili ZhuParris (azhup...@chdr.nl) and Robert-Jan Doll (rjd...@chdr.nl)
Daily LIACS supervision: TBD
Responsible RD: Geert Jan Groeneveld
Education Director: Jeroen van Smeden
Period: 6-9 months; Starts immediately
Fulltime: yes
Financial compensation: yes
Title: Prediction of Brain Age Using EEG Data
Research description:
With the increasing life expectancy, an expanding number of individuals suffers from age-related
neurodegenerative diseases such as dementia, Alzheimer disease and Parkinson disease. Despite this increase in prevalence and incidence, the development of new clinical therapies and drug treatment options remains low. Evaluation of new treatment options, such as new candidate drugs, require an informative and robust tool. Neuroimaging technologies, such as MRI and EEG, can be used for diagnosing and tracking progression of neurodegenerative diseases via a biomarker called ‘Brain Age’(BA). BA functions as a biomarker for aging, and an increased predicted BA is associated with numerous neurodegenerative diseases such as dementia, depression and cognitive decline. Many studies show changing EEG signals to be able to indicate aging as well as cognitive decline and dementia.
This internship focuses on the development of a machine learning model that predicts the BA of healthy individuals based on resting state EEG features. These models can be used for evaluation of new compounds and other interventions targeting neurodegenerative diseases. This internship will start with a literature review focusing on existing studies that use EEG and machine learning to predict BA. Next, EEG data from CHDR trials will be used to train and optimize a BA machine learning model. Within this study, we start with an exploratory phase to determine which EEG characteristics improve the prediction of brain age in healthy subjects, and which ML algorithm produces the best prediction performance. Finally, findings of the experiments will be summarized in a final report and presented at the CHDR.
Products:
Literature review, report/thesis, two presentations (one after a few weeks, one at the end of the internship