FMCW radar (Frequency Modulated Continuous Wave) enables contactless vital sign monitoring without the need to attach electrodes or body-worn sensors to the patient. The periodic movements of the chest wall surface caused by breathing (displacement typically in the millimeter range) and heartbeat (typically in the sub-millimeter range) modulate the phase of the reflected radar signal. Respiratory rate and heart rate can be estimated from these phase changes.

Compared to contact-based methods such as ECG or pulse oximetry, no electrodes or clips need to be applied. This increases comfort during long-term monitoring, avoids skin irritation, and is particularly suitable for patients in whom contact sensors are problematic (e.g., preterm and newborn infants or burn patients). Compared to camera-based approaches, radar additionally offers a privacy advantage, as no image of the person is recorded. Potential application areas include inpatient and home care as well as intensive care and sleep medicine.

Our research focuses on robust statistical signal processing methods for the joint localization of persons and estimation of their vital signs (respiratory rate and heart rate) using FMCW radar. Key challenges include the suppression of static and dynamic clutter, the separation of the weak cardiac signal from the dominant breathing signal (whose harmonics overlap with the cardiac frequency band), robustness against unintentional body movements and other interference, and the attribution of signals to multiple persons within the measurement area.

Possible thesis topics:

• Development and evaluation of robust methods for person localization using FMCW radar
• Robust estimation of respiratory rate and heart rate from radar signals
• Signal separation and vital sign estimation for multiple persons in the measurement area
• Analysis of the impact of interference sources (clutter, body motion) on estimation accuracy and development of compensation methods
• Simulation and verification of algorithms for joint localization and vital sign estimation
• Comparison of model-based signal processing with data-driven (machine learning) approaches

Knowledge of Python and/or C++ and a background in signal processing are recommended. Familiarity with statistics or optimization is advantageous but not required.

Christoph Löser (contact)

Stroke is one of the most common causes of permanent impairments of arm and hand function. Conventional therapy approaches often provide limited means to objectively track movement progress or to individually adapt the level of support. Our research focuses on an end-effector rehabilitation robot combined with camera-based, markerless motion capture to quantitatively analyze therapeutic movements and adaptively tailor support to each individual patient. Robotic sensor data and image-based information are processed jointly to enable precise motion analysis and to align therapy with real-world, everyday movement demands.

Possible research topics include:

• Analysis and modeling of assistance and compensation strategies in robotic systems
• Fusion of robot and camera data for precise arm position estimation
• Quantitative analysis of movement quality
• Analysis of multimodal sensor data (e.g., EMG, EEG) in neurorehabilitation
• Machine learning and computer vision for rehabilitative applications

Knowledge of Python (or alternatively MATLAB/Simulink) as well as signal processing is recommended. Basic knowledge of robotics or control engineering is an advantage.

Friederike Hicking (Contact)

Within the LOEWE focus area MultiDrug-TDM, an innovative point-of-care sensing system is being developed to determine the concentrations of anticancer drugs directly at the patient’s bedside in pediatric oncology. The goal is a personalized, data-driven dose adjustment in real time without the need for specialized laboratory analysis.

Raman spectroscopy plays a central role, as it captures the molecular composition of a blood sample in a single spectrum. The signal contributions of individual substances overlap, requiring robust spectral unmixing algorithms to determine the concentrations of the respective drugs. In addition, a Virtual Raman Device is being developed, a simulation model of the measurement system that enables the generation of synthetic training data and can be used to optimize measurement parameters.

Possible research topics include:

• Identification of physical prior knowledge for Raman spectra
• Integration of physical constraints into unmixing algorithms
• Comparison of physics-informed and classical unmixing approaches
• Development and training of a Virtual Raman Device

Knowledge in Python is recommended. Prior experiences in machine learning can be beneficial for certain topics.

Malte Drees (Contact)

Camera-based vital signs estimation is becoming an increasingly popular topic in research, with great potential for future applications. It enables contactless monitoring of patients' vital signs, which can help reduce patient stress and shift focus to treatment in emergency situations, rather than attaching sensors, finger clips, etc.

Many vital sign estimation methods, such as PPGI (Photoplethysmography Imaging), rely on model parameters that must be determined through machine learning. A major challenge here is the lack of sufficient training data and data diversity, which is further complicated by data protection regulations.

The solution: generating artificial and simulated training data, consisting of both synthetic videos and simulated vital parameters. The goal is also to simulate potential diseases and dysfunctions so that the resulting algorithms are robust to (pathological) abnormalities.

Possible topics include:

• Modeling of a realistic human body model
• Generation of a stochastic skin and hair model
• Simulation of vital signs for healthy individuals and individuals with illnesses
• Comparison of real and synthetic training data by evaluating the performance of trained algorithms, etc.

Recommended skills: Knowledge of Python. Experience with rendering engines (preferably Blender) is beneficial but not required.Theoretical knowledge of machine learning may be advantageous for certain topics

Tobias Reinhardt (contact)

With Electrical Impedance Tomography (EIT), cross-sectional images of the human body can be generated by injecting small amounts of electrical current and measuring the resulting voltages. Unlike Computed Tomography (CT), no harmful X-rays are used, and the system is also significantly less expensive than an MRI. The applications of EIT are diverse: it can be used for monitoring respiration, examining the musculoskeletal system, and, in a miniaturized form, for monitoring cell growth.

In EIT, both simulated measurements and measurements performed on experimental tanks or on humans are used. We are researching how to modify our measurement setup and our measurement protocol so that we obtain the most informative measurements possible. We are also investigating how to adapt our simulation environment so that the simulated measurements resemble the real measurements as closely as possible, allowing us to transfer insights gained from simulations more effectively to reality. The main application area we focus on is the musculoskeletal system, for example the monitoring of fracture healing.

Possible topics for research include:

• Further development of the existing measurement setup
• Development and implementation of various methods to improve the measurement protocol, e.g., using Machine Learning
• Simulation of EIT measurements in MATLAB, e.g., of fractures
• Comparison of simulation and reality and reduction of the differences

Knowledge of Python and/or MATLAB is recommended, but not strictly required. Prior knowledge of Machine Learning/Deep Learning may be helpful for some topics.

Theresa Nolte (contact)

The quality of treatment for various diseases (such as epilepsy and cardiovascular diseases) has been significantly improved by their early detection using medical measurement methods (such as ECG, EEG, EMG, PPG, CT, MRI, etc.). Unfortunately, these measurement methods often suffer from artifacts and missing sensors, so that they only provide noisy and incomplete data. We are therefore researching robust algorithms that can deal with these artifacts and supplement missing information, for example by fusing with other modalities.

Possible topics for processing are

Robust machine learning & interpretability

Methods for sensor fusion (cameras, wearables, …)

Unsupervised / self-supervised learning for time series

Transfer learning between different modalities

Knowledge of Python and signal processing is recommended. Theoretical knowledge of machine learning / deep learning can be an advantage for some topics.

Jad Haidamous (Kontakt)

Vital parameters such as pulse and respiration are important indicators of various diseases and the general state of health of a person. The gold standard methods, such as electrocardiography for monitoring the heart, usually require contact with the patient, e.g. in the form of adhesive electrodes. We are researching ways to replace these with cameras where possible.

Possible topics include

• Machine Learning & Interpretability
• Multiphysics simulations and rendering of synthetic video data
• Methods for sensor fusion (cameras, wearables, …)
• Photoplethymography imaging
• Robust estimation of heart rate variability

Knowledge of Python (or alternatively Matlab) and signal processing is recommended. Knowledge of Blender or Unreal Engine may be an advantage for some topics.

Our doctoral candidate Maurice Rohr is close to completing his PhD. If you are interested in this topic, please contact Tobias Reinhard.

Tobias Reinhardt (Contact)

Stress is a common cause of sleep problems, concentration disorders, depression or even cardiovascular disease. While stress is often assessed purely subjectively, we are working on recording vital signs in order to make a quantitative statement about a person's well-being. The vital signs are primarily to be recorded non-invasively in order to enable the end user to be as comfortable as possible. In addition, stress can also be recorded in people who are unable to articulate this independently, such as newborn babies or people with disabilities.

While research is mainly focussed on the development of stress caused by acoustic stimuli such as noise, it is not limited to this. Possible areas of application include contactless parameter recording at the workplace, in the car or in bed. The parameters are to be analysed primarily with the help of machine learning algorithms.

Possible topics:

• Non-invasive vital parameter estimation based on cECG, thermographic camera and PPGI
• Sensor fusion for robust parameter estimation
• Quantitative analysis of stress
• Application of machine learning with a focus on computer vision and signal processing
• Implementation of virtual reality environments for test setups

Programming skills in Python or Matlab are recommended. If you are looking for a topic in the field of VR, it is recommended that you have knowledge of C# and Unity programming.

Tizian Dege (Contact)

The analysis of human movement is an important component of many applications in medicine and sports. A classic area of application is physiotherapy, where exercises are ideally performed under the supervision of a healthcare professional who can provide individualized and immediate feedback. However, performing exercises at home without direct supervision can also support the recovery process. At the same time, incorrect execution, misjudgment of one’s fitness level, and overexertion can lead to inefficient training or, in the worst case, serious injuries.

Beyond physiotherapy, automated movement analysis is becoming increasingly important in other fields as well. One example is cycling biomechanics, where individual movement patterns while riding a bicycle can be analyzed, for instance to evaluate riding posture or to investigate biomechanical loads during movement.

To address such challenges, automated assessment systems can be used to evaluate the quality of movement patterns, provide personalized feedback, and reduce the need for human supervision. Building on advances in computer vision, we are researching the development of such support systems using video-based human pose estimation.

Possible research areas include:

• Natural Language Processing (NLP) for the automated generation of training data from videos with subtitles
• Analysis of sensor data (EMG, force plates, IMUs) and its relationship to camera recordings
• Development of biomechanical models of human movement
• Development of methods for the quantitative assessment of human movement
• Machine Learning and Computer Vision
• Movement analysis in the context of physiotherapy, sports, and cycling biomechanics

Recommended skills: Knowledge of Python, computer vision, and signal processing is recommended. Experience with Blender and NLP may be advantageous for some topics.

Sebastian Dill (Contact)

Video-based Human State Analysis refers to the use of video footage to assess and interpret the physical, emotional, or psychological states of individuals. This involves analyzing visual data captured by cameras to derive insights about a person's behavior, emotions, actions, or overall well-being.

Possible topics:

• Methods for detecting/tracking human limb movements to correlate with the individual's current state
• Methods for human reflex analyses
• Quantitative analysis of human emotional state
• Machine learning and computer vision

Using tools such as Python, Matlab, classical signal, and image processing

Our doctoral candidate Gökhan Güney is close to completing his PhD. If you are interested in this topic, please contact Tizian Dege or Friederike Hicking.

Tizian Dege (Contact), Friederike Hicking (Contact)