Projects 2024

Contact

Optimizing Resource Planning in Surgical Wards Through Data-Driven Resilience

Data science PI: Emilia Grass

Life science PI: Sabine Jung-Klawitter

Daily resource management in surgical wards is complex due to the uncertain nature of surgeries, patient complications, high associated costs, and the need for specialized supplies and workforce. The complex resource planning and management task can be further exacerbated by unforeseeable events such as adverse events.  Adverse events can be natural hazards or man-made disasters.

Mismanaging a surgical ward's resources during adverse events can lead not only to financial losses but also to non-financial losses such as higher casualty rates and personnel burnout. To minimize such losses, a hospital needs to improve its operational resilience before, during, and after adverse events.

In this project, we propose a data-driven resilient optimization framework to assist hospital decision-makers in managing and planning the surgical ward’s resources to withstand such adverse events. By optimizing resource use in advance or immediately after the event, resilient surgical wards could improve the quality and speed of urgent patient care, limit direct and indirect casualties, and decrease mortality.

Contacts

Using machine learning for citizen science surveillance of ticks and prediction of tick-borne diseases with climate change

Data Science PI: Rainer Stiefelhagen

Life Science PI: Joacim Rocklöv

TAC Member: Ulrich Köthe

Vector-borne diseases, transmitted by species such as ticks and mosquitoes, represent a critical global public health threat exacerbated by climate change and shifting land-use patterns. Current surveillance efforts, particularly those leveraging citizen science and machine learning, face significant technical hurdles, including high battery and memory demands on mobile devices, limited internet connectivity in remote areas, and the difficulty of processing low-quality images with complex backgrounds. Furthermore, existing machine learning approaches often lack interpretability and fail to integrate socio-economic drivers or immunity dynamics.

This research addresses these gaps by optimizing data collection and disease modelling. Key objectives include:

• Efficient Data Collection: Investigating model compression techniques (pruning and quantization) to improve the performance and efficiency of object detection models on mobile devices.
• Explainable Modeling: Developing interpretable machine learning models that incorporate known biological mechanisms and utilize importance and sensitivity analyses to identify key factors in disease transmission.
• Risk Prediction: Utilizing citizen science surveillance data to build spatial risk models and predict tick-borne encephalitis (TBE) risk maps to inform public health policy.

Contact

Model-informed machine learning for the prediction of postoperative atrial fibrillation and its targeted prevention

Data science PI: Axel Loewe

Life science PI: Christian Niklas & Rosa Klotz

Postoperative atrial fibrillation (POAF) complicates roughly 20 % of surgical procedures and can trigger hemodynamic instability, an elevated stroke risk, longer hospital, intensive-care stays, and higher costs (Dobrev et al., 2021). Onset typically occurs 2-4 days after surgery, providing a promising window for preventive interventions, that high-risk patients can be identified during or shortly after the operation.The University Hospital Heidelberg (UKHD) has assembled a rich dataset of biosignals and clinical measurements taken before, during, and after surgery. Although the cohort size is substantial, it remains underpowered for conventional deep-learning approaches. To overcome this limitation, we will augment the real-world data with synthetic recordings generated by validated multiscale simulations of cardiac electrophysiology. As a result, creating a precision-medicine tool for individualized POAF prevention. The project will first encode the most relevant POAF risk factors into a mechanistic simulation model, producing large volumes of synthetic data suitable for deep-learning training. Parallelly, we will explore feature-based machine-learning methods on the clinical dataset to establish a performance baseline. Building on these results, we will develop and evaluate explainable deep-learning models trained exclusively on in-silico data, reserving the clinical dataset solely for testing. Ultimately, the project aims to deliver translatable insights into preventable causes and effective, targeted interventions for POAF onset in surgical patients, grounded in a mechanistic understanding of its multifactorial etiology.

Contact

Mathematical modelling and probabilistic forecasting of infectious disease spread

Data science PI: Melanie Schienle

Life science PI: Johannes Bracher

The project addresses statistical challenges in epidemic modelling and forecasting. In infectious disease modelling, we often encounter the issue of overdispersion, where observed data variability considerably exceeds the expected value. One of the modelling stages, where overdispersion occurs is nowcasting. Infectious disease data typically suffers from delayed reporting, artificially biasing the data downwards, painting an inaccurate picture of an outbreak development. Nowcasting is the process of estimating the current state of an epidemic by correcting for delays in disease reporting. While various heuristic approaches exist to account for overdispersion in nowcasting e.g., via negative binomial distributions, these remain poorly motivated and often show unsatisfactory performance in applied analyses. In this project, we will compare and evaluate nowcasting approaches accounting for overdispersion.

Similarly, overdispersion affects the estimation of the effective reproductive number, a key characteristic of epidemic spread. Widely used tools like the EpiEstim R-package rely on models using the Poisson distribution, which assume equidispersion. In this project we demonstrate, that the Poisson assumption often leads to overconfident estimates and underestimated uncertainty.

The final stage of the project addresses the evaluation of forecasts under conditions of partial missingness, which is a common issue in collaborative forecasting hubs, where not all teams manage to submit their forecast on time. Existing methods often introduce bias or incentivize strategic omissions. We aim to introduce a comparison framework that accounts for missing forecasts and encourages honest forecasting, without just creating incentives to “game the system”.

Contacts

Rooting Medical Decisions in Multi-modal Guidelines

Data science PI: Rainer Stiefelhagen

Life science PI: Jens Kleesiek

Funding phase: 02/2025 - 01/2028

Clinical decision-making unfolds across repeated encounters and diverse data modalities, yet evidence-based guidelines are still largely published as static documents that are hard to operationalize and to match to patient-specific evidence. This project roots AI decision support in explicitly structured guidelines by converting guideline flowcharts into machine-readable graphs, aligning them with longitudinal patient trajectories, and learning to recommend plausible next clinical steps with calibrated uncertainty.

Using prostate cancer radioligand therapy in nuclear medicine as a case study, we develop an automated pipeline that extracts decision steps, allowed actions, and transitions from guideline documents through document layout analysis and large language models. We then align real-world patient histories with guideline stages and integrate multimodal evidence from imaging, laboratory measurements, pathology, and clinical text using specialized neural models. Building on these representations, we will train a Clinical Decision Assistance Model to predict and justify likely next actions while quantifying uncertainty, including situations in which clinical practice deviates from guideline recommendations.