Counterfactual decision support for the control of a clinical, multi-organ failure ICU machine

The leading cause of death in the intensive care unit (ICU) is multiorgan failure (MOF). The mortality of MOF patients increases especially when liver, lung, and kidney functions are impaired. Clinical practice connects a separate device for each failing organ, but this can worsen the patient condition, ignores organ interactions, makes access to the patient difficult, and leads to increased risk of complications. The ADVOS multi-organ support system (ADVITOS GmbH, Germany) provides systemic MOF treatment by combining liver, kidney, and lung support as well as extracorporeal blood pH correction. Studies show an improvement in patient survival rates from 10% to 35–50%. The machine is already deployed in many Austrian and German clinics. 

Due to the complexity of the procedure, use of that system requires a high degree of experience on the part of the user or support staff: thus, in collaboration the Fraunhofer Institute for Cognitive Systems (IKS) and Avditos co-developed an AI-based model for decision control support. This addition allows for effective patient trajectory planning and, more importantly, reduces the workload of the often overloaded staff. Our method leverages experimental data and expert knowledge to (a) recommend action times and treatment options and (b) obtain causal insights into the measurements' relationships. For this, we used methods of counterfactual inference an offline reinforcement approach that builds a latent representation factoring out the factual assignment. The patient status can thus be unbiasedly projected and evaluated even for unseen treatment procedures. This enables the refinement of the decision-making process, offering valuable insights into potential alternative courses of action and improving the overall effectiveness of the treatment strategy.