Predictive modelling for intensive care
Time-series models that forecast oxygenation deterioration up to 72 hours ahead. Six years of my life, and the thing I still find most worth doing: buying a clinician time they did not have.
What the work looks like
A patient's oxygenation, hour by hour. At the marker the model stops observing and starts forecasting — and the range of what might happen next widens the further ahead it looks.
The useful part is not the single line. It is knowing, early, that enough of those trajectories end up somewhere a clinician would want to act on.
Illustrative — not patient data.
- Observed
- Forecast, up to 72 h ahead
- Level worth acting on
Right now
- Research
- Getting an ICU oxygenation model ready for external validation — the real test of whether it travels.
- Teaching
- Version control and reproducible workflows, for research groups who never got taught them.
- Building
- A home lab for local inference, and a food-sharing site for Aachen on the side.
What I work on
Accessible high-performance computing
Plenty of groups have the data and the question but not the cluster experience. Through the NHR4CES Simulation and Data Lab I help close that gap.
Open tools and materials
The papers are the smallest part of what a project produces. I publish the tools and the teaching materials too — every workshop below is free to take and reuse.
Open to what's next
I'm looking for the next thing — a role, a collaboration, or a project worth building. What I like most is the messy middle: real data, a real decision at the end of it, and a model that has to earn its place.
If that sounds like something you're working on, say hello.
Models on data that fights back
Time-series forecasting on clinical records that are irregular, imbalanced and full of holes — and the validation work that decides whether a model survives its second hospital.
Behind it: Machine Learning: Health 2026, SoftwareX 2023
LLMs, in practice rather than in theory
Choosing them, evaluating them, and getting them into other people's hands — from a workshop for a research centre to local inference on my own hardware.
Behind it: LLMs at CCLS, a home lab that runs them
Research engineering that scales
SLURM and national compute allocations, reproducible pipelines, and Python tooling packaged so the next person can actually run it.
Behind it: HPC for Researchers, NHR4CES Simulation and Data Lab
Getting it adopted
Four workshops, materials anyone can reuse, and years of translating between people who write the models and people who make the decisions.
Behind it: every deck below, free to take
Papers
Four2026
First author
First author
Predicting a sudden decrease in oxygenation in mechanically ventilated adult intensive care patients
Machine Learning: Health · Polzin R, Fritsch S, Sharafutdinov K, Bickenbach J, Marx G, Schuppert A
2024
Computational simulation of virtual patients reduces dataset bias and improves machine learning-based detection of ARDS from noisy heterogeneous ICU datasets
IEEE Open Journal of Engineering in Medicine and Biology · Sharafutdinov K, Fritsch SJ, … Polzin R, … Schuppert A
2023
First author
First author
Diagnostic Expert Advisor: a platform for developing machine learning models on medical time-series data
SoftwareX · Polzin R, Fritsch S, Sharafutdinov K, Marx G, Schuppert A
2021
Algorithmic surveillance of ICU patients with ARDS (ASIC): protocol for a multicentre stepped-wedge cluster randomised quality improvement strategy
BMJ Open · Marx G, Bickenbach J, Fritsch SJ, et al. — one of ~45 co-authors
Where I've been
- 2024 →Postdoc, JRC COMBINE
Continuing the ICU prediction work with Prof. Schuppert's group, focused on clinical applicability.
- 2018 – 24Ph.D., JRC COMBINE
Six years on ARDS prediction for intensive-care patients, working with routine clinical data.
- 2016 – 18M.Sc., Maastricht University
Natural language processing and machine learning — sentiment analysis, topic modelling, deep learning.
- 2013 – 16B.Sc., Aachen
Computer science and mathematics, alongside a first developer role at a CNC simulation company.
Projects I'm part of
- 2025 →Fairteiler Aachen
Is there anything in the Fairteiler right now? A live status board for Aachen's food-sharing points — anyone can report what's on the shelf in about ten seconds, no account, no app store. Built and run in my own time.
- 2023 →EDITH CSA
Charting the road from single-organ models to whole-body digital twins.
- 2023 →SDL Digital Patient
Bringing high-performance computing capability to healthcare research groups.
- 2023 →CCLS
RWTH's push to bring computational methods to the life sciences.
- 2018 – 23SMITH
The clinical data platform behind my PhD work on early warning for ARDS.
Workshops I've run
Free to take and reuse
Nov 2025
Version Control for Researchers
A practical introduction to version control for research code: branching, collaboration, and a history you can actually retrace.
Open the deck → Nov 2024Synthetic Data in Medicine
When real patient data can't be shared — open-source tools for generating realistic medical datasets.
Open the deck → Nov 2024HPC for Researchers
From SLURM fundamentals through to applying for time on national compute resources.
Open the deck → Oct 2024LLMs at CCLS
Using large language models for research on CCLS infrastructure — worked examples and practical guidance.
Open the deck →Outside the lab
Roasting my own coffee
Green beans, a roaster, and an ongoing argument with myself about how dark is too dark. It is the one process in my life I am happy to run without logging the parameters.
A home lab that keeps growing
Self-hosted services, my own network, and local model inference — partly to keep my own data close, mostly because it is a good excuse to keep learning how things actually work.
Photography around Aachen
Walking the city with a camera, which is the best way I know to notice a place you think you already know well.
Co-op games, when there's time
Preferably the kind you play with other people in the same room, or at least the same voice channel.
Get in touch
Email is the surest way to reach me, and I answer it. If you have a problem in reach of a model, a dataset nobody has been able to make sense of, or a research group that needs to get off a laptop and onto a cluster — I'd like to hear about it.