Advanced Machine Learning for Intelligent Decision Support in the Intensive Care Unit

Short Summary
Treatment of critically ill patients requires sequential decisions based on the patient’s current state and medical knowledge. The enormous quantity of integrated data would cognitively overload humans. In this project, which is a collaboration between the Biomedical Informatics Group (ETH Zürich) and Inselspital (Universitätsspital Bern), we aim to integrate artificial intelligence (AI)-based decision support tools into the intensive care unit (ICU) to assist healthcare professionals in detecting critical patient changes while reducing information overload and alarm fatigue.
Goals
The project is structured along three main aims: The primary aim is to implement ready-to-use organ failure alarm systems and evaluate them in a prospective observational clinical study. The researchers will develop the ICU patient data visualizer software, presenting ICU data in a similar format to the bedside data. They will also create approaches to self-monitor the alarm systems to detect unreliable predictions. The performance of the alarm system will be compared with human assessors, and treatment recommendations will be evaluated. The second aim is to conduct a randomized, controlled clinical pilot trial to assess a live alarm system. The trial will also evaluate clinical stakeholders’ acceptance and adoption of the technology. The researchers estimate a sample size of 1,000 patients for the pilot trial to detect a meaningful reduction in circulatory and respiratory failure events. The final aim is to advance the methodology for patient state modeling and deterioration prediction. The researchers plan to develop machine learning (ML) models using a comprehensive ICU dataset from a hospital in Bern. These models will characterize a patient’s real-time health state based on various medical time series data. The goal is to create representations that summarize the patient’s medical history, predict near-future health events, and differentiate physiological subsystems. By leveraging AI techniques and mining large ICU datasets, the researchers aim to develop precise patient state representations that can predict health-state changes such as circulatory or respiratory failure.
Significance
We aim to model how treatment decisions in the ICU impact a patient’s condition and understand how the change in patient states can be predicted. Machine Learning (ML) methods enable combining the clinical experience of multiple physicians, represented in the data by treatment decisions following specific patient states. Overall, this project aims to develop AI-based support tools for ICUs to assist healthcare professionals in detecting patient deterioration, improving treatment decisions, and ultimately enhancing patient outcomes.
Background
Critically ill patients require treatment in an ICU where continuous monitoring of organ function parameters enables early recognition of physiological deterioration and the onset of appropriate interventions. Healthcare professionals must react to frequent alarms when monitored organ function parameters cross preset thresholds, leading to information overload and alarm fatigue.

iDoc

csm_Gunnar_Raetsch_0498_b04d3a8306

Prof. Dr. Gunnar Rätsch

ETH Zurich
Co-Investigators
  • Prof. Dr. med. Jörg Schefold, Inselspital
  • Prof. Dr. med. Stephan Jakob, Inselspital
  • Prof. Dr. med. David Berger, Inselspital
  • Dr. Tobias Merz, Auckland City Hospital
Consortium
  • Auckland City Hospital

Related Posts