ICU Time series & Early Warning Systems
In intensive care units (ICUs), physicians are facing large quantities of data from many patients, such as physiological signals, lab measurements, observation records, etc. It becomes increasingly challenging to identify the most important information for care decisions using only manpower. Our group develops Early Warning Systems (EWS) for different types of organ failure based on machine learning methods to forecast impending organ failure using both time series as well as other meta-data from the ICU of Inselspital (University Hospital Bern). The circulatory Early Warning System (circEWS) our group developed was able to detect 81.8% of the circulatory failure event 2 hours before the onsets, and this model can be transferred well to the ICU data from another hospital in another country. We actively work on early warning systems for other types of organ failure, such as renal failure and respiratory failure. These early warning systems could provide a useful aid to ICU physicians in making clinical decisions.
Benchmark For the Early Prediction of Rare Events in Healthcare
Healthcare is an application field with a high societal impact. If, for other medical sub-fields, datasets are provided with standardized tasks and pre-processing, this is not the case for the major databases from patients' EHR in the ICU. However, comparing works across multiple benchmarks is a key step toward improving deep learning models, including unsupervised approaches. Here we aim to address this issue by providing benchmark tasks based on the HiRID database  that are reproducible, diverse, and clinically relevant. We developed HiB, a benchmark containing six different clinically relevant tasks for patient monitoring in the ICU, achieving the goals of this work package. This work provides a modular library to pre-process the raw HiRID data and extract labels. This work will supplement MIMIC-III Benchmark, which is the only well-established benchmark for ICU data. We aim to build on this benchmark to evaluate the methods described in the following work packages.
Building prognostic and diagnostic models of health outcomes
We are developing prognostic and diagnostic models of cardiovascular and other disease outcomes by integrating the full spectrum of data, including genetic, imaging, and novel biomarkers using electronic health records datasets (EHR) and other data. Several manuscripts are under development. In one such project, we evaluated ways to improve the calibration of established ASCVD risk prediction models using machine learning tools without sacrificing model interpretability [3,4]. In another project, we created SARS2 simplified risk scores of hospitalizations and death among patients with COVID‑19 and made it available as an online calculator .
 Stephanie L Hyland, Martin Faltys, Matthias Hüser, Xinrui Lyu, Thomas Gumbsch, Cristóbal Esteban, Christian Bock, Max Horn, Michael Moor, Bastian Rieck, et al. Early prediction of circulatory failure in the intensive care unit using machine learning. Nature medicine, 26(3):364–373, 2020.
 Hugo Yèche, Rita Kuznetsova, Marc Zimmermann, Matthias Hüser, Xinrui Lyu, Martin Faltys, and Gunnar Ratsch. Hirid-icu-benchmark—a comprehensive machine learning benchmark on high-resolution icu data. 2021.
 Zinzuwadia, A. N., C. Li, H. Dashti, L. Chen, B. Cade, E. Karlson, S. Mora, N. Paynter and O. Demler (2022). "Combining Multiple Cardiovascular Risk Prediction Models Improves Model Performance: A Novel Flexible Framework Approach Applied To A Contemporary And Diverse Ehr Cohort." J Am Coll Cardiol 79(9_Supplement): 1510-1510.
 Zinzuwadia, A. N., C. Li, H. Dashti, L. Chen, B. Cade, E. Karlson, N. Paynter, S. Mora and O. Demler (2022). "Abstract P021: Performance Of Pooled Cohort Equations And MESA Risk Score Across Race/Ethnicity And Socioeconomic Status To Estimate 10-year Cardiovascular Risk In Diverse New England Cohort." Circulation 145(Suppl_1): AP021-AP021.
 Dashti H, Roche EC, Bates DW, Mora S, Demler OV. SARS2 simplified scores to estimate risk of hospitalization and death among patients with COVID-19. Sci Rep. 2021 Mar 2;11(1):4945. doi: 10.1038/s41598-021-84603-0. PubMed PMID: 33654180; PubMed Central PMCID: PMC7925678.