Real-time Acute Kidney Injury (AKI) Prediction

A deterioration in renal function and acute kidney injury (AKI) is one of the most common complications patients in the intensive care unit (ICU) experience. AKI is now recognized as a heterogeneous condition that extends beyond immediate morbidity and mortality. It affects a patient’s long-term prognosis, emphasizing the need for comprehensive understanding and management. While recent studies have shown promising potential for interventions to mitigate and reduce the risk of AKI or its progression, secure methods to predict a deterioration in renal function are needed to apply such measures before AKI occurs and apply them to the right patients as they are not without risk themselves. Artificial Intelligence’s (AI) potential to transform AKI diagnosis and treatment is immense.

However, ICU data poses a unique challenge to AI research as it is often sparse and highly inhomogeneous. Robust and well tested pipelines for working with such complex data are needed in order to facilitate research for smaller teams, make it more resource efficient and ensure the validity of results in this high-stakes environment.

OpenICU

In recent years, the field of intensive care has seen more and more applications in the field of machine learning. Being one of the most data dense areas of medicine, several programs have been established in order to share data across sites and across the community. Yet currently, there are no open source standard tools for working with these large sets of ICU. Each study site uses their own methods for extracting data, preprocessing and delivery. The lack of standardization might introduce bias into the data at a very early level, negatively impacting model development. It is also highly resource intensive to reinvent the wheel at each single study site. Unifying research on a common platform facilitates research, provides access to a wider audience of scientists and ensures common methodologies for research questions. Which is why, in cooperation with Christian Porschen, we kicked off the “OpenICU” project. OpenICU represents an open source platform of microservices and tools that facilitate working with ICU data in the future, by providing well tested pipelines from data extraction over preprocessing, up until the model implementation.

Deep Kidney

The goal of the “DeepKidney” project is to develop a “Digital Twin” of the kidneys for patients in critical condition. Instead of focusing on thresholds, defined in order to algorithmically approach renal dysfunction, the goal of this strategy is to directly simulate organ function over time with predictions into the future. In order to do this, we utilize modern time series-ready algorithms with the ability to leverage this data dense information space. By directly stimulating organ (dys-)function and including domain knowledge into our models, we also intend to open up possibilities for simulation of interventions for patients. The vision behind it is to provide physicians with a clinical decision support system where they can safely test out interventions on a digital twin of their patient’s organ to further personalize critical care.