Operations Research 2021
Abstract Submission

536. Classifying ready-for-transfer patients in the intensive care unit based on clinical data

Invited abstract in session TD-16: Logistics in Health Care, stream Health Care Management.

Thursday, 14:20-15:40
Room: Schilthorn

Authors (first author is the speaker)

1. Franz Ehm
Industrial Management, TU Dresden
2. Volkmar Franz
Department of Anesthesiology and Critical Care Medicine, University Hospital Carl Gustav Carus Dresden
3. Maic Regner
Department of Anesthesiology and Critical Care Medicine, University Hospital Carl Gustav Carus Dresden
4. Udo Buscher
Industrial Management, TU Dresden
5. Hanns-Christoph Held
Director of the surgical intensive care unit, University Hospital Carl Gustav Carus Dresden
6. Peter Spieth
Department of Anesthesiology and Critical Care Medicine, Univer-sity Hospital Carl Gustav Carus Dresden


In the intensive care unit (ICU), a limited number of beds and personnel are available for the treatment of the most severely sick patients. In the event of new arrivals physicians often have to decide upon patients who are ready to go to a lower ward and thereby free up capacity. The choice is complicated as it depends on numerous clinical and operational factors and comes at the risk of causing negative patient outcome. Clinicians at Uniklinikum Dresden (UKD) expressed their need for a data-driven decision support when identifying ICU patients for transfer. For this purpose we build a mathematical classification model that is trained on historical clinical data to evaluate actual patient information. Patient outcome is labelled according to readmission to the ICU within 72 hours following discharge. The transfer decision is then modelled as a classification problem which predicts the event of readmission. Our research is based on clinical data from a total of 41108 episodes providing information on transfer history, patient characteristics, vital parameters and treatments. In a first step, relevant features for transfer are specified by expert clinicians at UKD. We use these as a guideline to construct a first classifier which employs a specific threshold value of the number of critical parameters in a patient to predict her chance of readmission. In a next step, logistic regression models are deployed using binary variables from critical parameter counts as well as metric features from measurements, scores and patient data. Performance of the trained classifiers is evaluated on a test set using accuracy metrics as well as the receiver-operating-characteristic.


Status: accepted

Back to the list of papers