Skip to main content

Topics & projects

Every CACoM project answers a research question with a clinical or biomedical motivation. You can take one of the projects listed below or propose a custom topic.

Either way, you turn the project into a specific, testable question, write a short proposal and get it approved by the topic approval deadline. The Refinement guide and the Proposal guide explain how, and the Schedule has the dates.

How listed projects work

Each listed project goes to one group. Many projects are broad enough to be split into subprojects, so ask us if you want to do that. Unlike custom topics, listed projects need no mini-pitch.

Projects this semester​

Fetal heart rate and CTG​

CTG (cardiotocography) is the standard way to monitor the fetal heart rate, yet its accuracy, interpretation and clinical value are still debated. Most CTG monitors estimate the heart rate from Doppler ultrasound, and signal quality, device validation and the transparency of the algorithms remain open questions. Our in-house fetal heartbeat simulator produces realistic heartbeat signals, which makes both experimental and computational projects possible.

CTG monitors on the test bench

Set up our fetal heartbeat simulator as a standard test tool for CTG monitors from different manufacturers, then use it to start on tests related to short-term variation (STV), a measure computed from the fetal heart rate. On the simulator you set the heartbeat yourself, so you know what a monitor should report. Setting up the rig is part of the project and can take a while.

The project is large and difficult but has high potential. It suits ambitious students with a knack for physics experiments.

Three-tier vs five-tier CTG classification

Japan has long used a five-tier CTG classification, while the FIGO system has three tiers. How would you set up a study that compares the two? There is renewed interest in the question.

We have a computerized system that classifies traces in both systems, so you can concentrate on the study design, for example on what to compare the two classifications against. You deliver the study design and a pilot comparison on CTG data we provide.

Machine learning for predicting fetal distress

Many machine learning methods for predicting fetal distress from CTG have been published. We provide CTG data with umbilical pH as the outcome.

Start with a literature review and choose your methods: we must approve both within two months of the start of the course. Then implement the methods, evaluate them and draw your conclusions. Compare against a simple baseline, and report sensitivity and specificity along with accuracy.

For a custom topic in this area, you could for example:

  • study how ambient noise, acoustic interference or probe placement affect Doppler signal quality;
  • design or benchmark algorithms for beat detection or heart rate variability;
  • use public CTG datasets or simulated Doppler signals for reproducible machine learning studies;
  • explore transformations between ECG- and Doppler-derived series (for example smoothing or downsampling to mimic CTG resolution), if you can show that you have access to the data;
  • develop new measures of variability, noise robustness or clinical interpretability;
  • build or test research-grade Doppler devices for open validation (not medical devices).

Some of these need hardware or specialized datasets, so your proposal must say how you will obtain, simulate or share the data responsibly. Useful skills: signal processing, machine learning, time-series analysis, experimental design, numerical modeling and statistics.

Movement and wearable sensors​

Wearable sensors capture movement, including gait and posture, in the lab and in daily life. We work with body-worn IMUs (inertial measurement units) such as the actibelt®, with a foot-sole pressure insole and with selected actibelt datasets, some of them recorded in extreme conditions such as microgravity. Some datasets are unlabeled and call for creative preprocessing or self-supervised methods.

Weekly activity summaries for lifestyle coaching

A medical lifestyle-change unit offers check-ups with doctors and lifestyle recommendations, for example to managers or to people with depression or burnout. In a pilot study, we want to couple its program with the actibelt. Participants wear it for a week, and the data are analyzed and shown as a weekly graphical summary on a website or phone. A coach then discusses with each participant whether their activity is in the normal range for their age, sets goals (for example 10% more steps) and checks them at a follow-up a month later.

The missing piece is a semi-automatic route from an actibelt download to a clear weekly summary, with activity levels week by week, averages and graphics. Such a summary would also be useful in clinical studies and as a consumer or medical-grade tool for diagnosis or for supervising treatment.

This semester you build that pipeline and validate it on recordings from volunteers, for example by testing how reliable each weekly measure is. You also draft the protocol for the pilot study with real participants.

Movement data from the International Space Station

We have a video of an astronaut on the ISS who wears an actibelt while riding a stationary bike and doing some basic activities around that time. The recording covers one session. Analyze the accelerometer and gyroscope data against the video, for example by aligning the two in time, labeling the activities from the video and testing how well the sensor data alone recognize them. In orbit the accelerometer measures no static gravity, so methods that rely on it (for posture or tilt) need rethinking and the gyroscope matters more.

The video and data are shared with the project team only.

A wireless PCB for the FullSoul insole

The FullSoul is a thin-leather barefoot shoe with a foot-sole pressure sensor, built by a previous CACoM team. In Summer 2026 another CACoM project built a first Bluetooth Low Energy (BLE) module for it. Build a better implementation: design a custom PCB around the BLE module we have selected and have it manufactured. The data go to your own laptop. Unlike most CACoM projects, this one centers on building a device, so it runs in full alignment with us and under closer supervision.

You must evaluate the result quantitatively, for example by measuring lost frames, sampling stability, latency and battery life.

Foot sensitivity in barefoot shoes

Barefoot (minimalist) shoes have thin, flexible soles that are meant to keep the sensory feedback from the ground, while thick-soled trainers cushion the foot and may change how we walk. A previous CACoM team built the FullSoul, a thin-leather barefoot shoe with a foot-sole pressure sensor, and compared its pressure distribution with that of a sneaker carrying a similar sensor in its thicker midsole.

How much sensory feedback does the foot actually receive? Compare four conditions: bare foot, the FullSoul, the FullSoul with its velostat pressure sensor in place (to check that the sensor doesn't blunt sensitivity) and a barefoot shoe with a thin rubber sole, the most common alternative on the market.

Start from the tuning-fork test (courtesy of Prof. Dr. Walther, Schön Klinik): a vibrating tuning fork is pressed against the foot, and the subject marks on a foot diagram where they felt it. Then find or develop further methods, because several independent measures make a stronger comparison. Daniel Lieberman's work on barefoot locomotion used tactile tests with subjects lying on the floor, but lying down may itself reduce sensitivity, so any adaptation has to consider posture. Further benchmarking of the pressure sensor fits into this project too.

Cross-device clock synchronization

Every recording device with a free-running clock drifts, and the longer the recording, the larger the drift. The actibelt is calibrated at the start of each session but can record for up to two months, so drift gets in the way of any analysis that combines devices, such as two actibelts on one person, or an actibelt and a phone.

Build an automatic synchronization method for devices without a central time source, and test it the way the devices are really used. Mount two actibelts rigidly together (a 3D-printed jig works well) and perform a known motion pattern, such as rotations and taps. Then take them out of the mount and use them separately for a longer period. Afterwards, put them back in the mount and repeat the pattern. Your method has to work with this kind of data: estimate the clock offset and drift from the sessions in the mount and keep the separate recordings aligned in between, when the devices no longer share any motion.

Measure how well the signals agree after alignment, and how large the error is at the second session in the mount when you predict it from the first. Two actibelts are the natural starting point; an actibelt and a smartphone is the more useful long-term target, because the phone can then serve as a general-purpose reference.

Automated gold-standard annotation of walking videos

Better movement models need gold-standard walking data: recordings with the exact time of every heel strike and toe-off, plus a reliable speed reference. Today one such recording takes four people: the subject wearing an actibelt, a second person walking alongside with a measuring wheel (and another actibelt) for ground-truth speed, a third filming the steps and a fourth who later goes through the video frame by frame to mark every heel strike and toe-off.

Replace the fourth person with an automated pipeline built on standard computer vision. One route is pose estimation (for example with the ultralytics Python package) followed by algorithms that find the gait events in the joint trajectories. The other is a model trained or fine-tuned to detect heel strikes and toe-offs directly. Manual frame-by-frame annotation is the natural reference for validation. A good pipeline can later grow into automatic activity labeling (walking, sitting, lying, transitions) for less controlled recordings.

A previous CACoM project tried markerless pose estimation on archival footage from 1973 of Skylab astronauts running around the inside of the station's cylindrical workshop. Generic pose models failed on the low-resolution footage, and the feet (ankles, toes and heels), which matter most for gait events, stayed hard to track. Redo that analysis with your pipeline and do better, especially on the feet.

For a custom topic in this area, you could for example:

  • develop or validate algorithms that estimate speed, step length or distance from foot- or waist-mounted IMUs, on a treadmill or in free walking;
  • recognize posture transitions or activity segments in IMU data;
  • explore drift correction, sensor fusion or context-aware constraints for better position estimates;
  • compare balance when walking barefoot and in minimalist shoes.

If you collect data, describe your setup (sensor type, sampling rate, participant safety). If you use actibelt or other clinical data, discuss data access and labeling limits with us early. Useful skills: signal processing, sensor fusion, time-series modeling, experimental design and algorithm validation.

Physiological signals and body sounds​

Electronic stethoscopes and respiratory microphones record heart and lung sounds, and optical sensors (PPG) track the pulse. These signals are non-invasive but hard to interpret reliably.

Heartwise: validating the actibelt with heart patients

Heartwise is a funded study in Singapore that aims to build a wearable early-warning system for patients whose heart failure may recur. Patients wear an actibelt almost around the clock, weigh themselves daily and record heart and lung sounds with an electronic stethoscope. This project is part of the Heartwise study.

It continues a CACoM project from Summer 2026 that piloted a short assessment protocol with heart patients at TUM Klinikum Rechts der Isar, using a handheld single-lead ECG and SpO2 device, digital stethoscopes and the actibelt during short functional tests such as the six-minute walk test. You repeat the pilot there with a stronger design and validate the actibelt measurements; clinical access is available.

The project involves patient data, so read the Data privacy and Collaboration policies before you design the protocol, and plan the timeline with us early. A medical student on the team helps, and so does German for talking to patients.

Heart rate from a PPG sensor, streamed to the actibelt

Photoplethysmography (PPG) detects the pulse with light, as smartwatches and pulse oximeters do. Measure heart rate with a PPG sensor worn on the finger, on the wrist, in a silicone patch or elsewhere. The sensor must have a live Bluetooth Low Energy (BLE) link to the actibelt. Validate its heart rate against an ECG chest strap, for example at rest and during movement, where PPG tends to struggle. An external supervisor is involved.

For a custom topic in this area, you could for example:

  • extract and compare spectral or temporal features of heart or lung sounds;
  • evaluate algorithms that detect murmurs or breathing irregularities in controlled or open datasets;
  • work on signal quality measures, denoising or automated screening;
  • study how breathing depth or posture affects the recordings.

Acoustic projects need clean, well-labeled recordings. If you collect your own, describe the recording environment, the equipment and how you calibrate it. Useful skills: audio and signal analysis, feature extraction, machine learning and validation methods.

Flexeal mask patch​

Flexeal is a silicone patch that seals FFP2 masks at the nose bridge so that air doesn't leak toward the eyes. Its soft, adhesive surface could also carry small sensors, for example for pressure, temperature or gases.

Sensors for the Flexeal mask patch

A partner group will use Flexeal masks with pregnant women. Which ultra-thin PCBs, batteries and sensors could go into the patch to gather data in many different situations? Brainstorm the options and design the patch electronics as a Bluetooth Low Energy (BLE) peripheral of the actibelt. This semester you carry out a design study and build a bench prototype of one sensor.

For a custom topic in this area, you could for example:

  • measure how well the patch seals across face shapes, movement and humidity;
  • embed small environmental sensors (CO₂, humidity, temperature gradient) to study breathing or comfort;
  • model airflow and leaks from measurements or simulations;
  • build a pipeline for sensor calibration and quality control.

Useful skills: sensor integration, measurement, signal processing, mechanical modeling and low-power electronics.

Surveys, reviews and clinical practice​

Not every CACoM project is technical. Some study how clinicians use, interpret or trust data-driven tools, through structured literature reviews, surveys, interviews or meta-analyses. These must be rigorous and quantitative.

Which tools matter to midwives

A delivery ward in Munich is run by midwives alone. Which tools are important to them? Set up a structured dialogue: draw up a semi-structured list of tools, put it on a web platform and interview the midwives so that the answers are complete. You visit the ward and see its work first-hand. The early feedback from users is also useful to us. If the approach works, it can be extended to other clinics in Munich. German is required.

For a custom topic in this area, you could for example:

  • run a systematic review of machine learning methods in obstetric monitoring or wearable sensing;
  • design and analyze a survey on how clinical researchers view the interpretability of algorithms or the transparency of data;
  • review computational medicine papers quantitatively for reproducibility statements and dataset availability;
  • analyze how fetal monitoring or wearable studies report uncertainty, validation metrics and dataset limitations.

Survey and interview studies need a plan for recruitment, dissemination and the ethical handling of responses, and studies with human participants may need approval. Useful skills: survey design, statistics, data visualization and systematic review methods.

Custom topics​

You can also propose a topic of your own, inside or outside the areas above. Combining areas is welcome, for example IMU data with physiological modeling. So are studies that integrate several data sources, and theoretical or simulation work on biomedical mechanisms or model validation. A custom topic needs at least one mini-pitch in class before the proposal (see the Proposal guide).

Be persuasive

Custom topics face a higher bar. Every custom topic must show exceptional initiative, a deep understanding of the problem and a realistic plan that fits CACoM's scientific and educational goals, and you must show that you can get the data or devices it needs. Unless an approved external collaborator (a clinician, research group or industry partner) proposed the topic, you must also convincingly justify its clinical relevance, its computational challenge and how you will evaluate the results.