Weaning patients from mechanical ventilation is often a challenge in intensive care units (ICUs). Traditional approaches, like spontaneous breathing trials (SBTs) and clinical scoring systems, often lack the sensitivity to capture subtle, patient-specific physiological trends, leading to delayed or failed weaning attempts. This project aims to develop a time-aware, interpretable machine learning framework to predict difficult weaning from mechanical ventilation in ICU patients using vital signs and mechanical ventilation parameters. The proposed model will forecast a patient’s readiness for weaning within the next 6–12 hours, providing proactive, data driven clinical support. The project will leverage real-world ICU datasets (e.g., MIMIC-IV, eICU) to build personalized, context-aware models that capture the dynamic nature of patient physiology. In addition, interpretable AI techniques, including Shapley values and attention mechanisms, will be employed to identify the most influential features affecting weaning outcomes. A novel aspect of this research is the exploration of counterfactual scenarios to simulate alternative clinical decisions (e.g., “What if we had attempted weaning earlier?”), providing critical insights for optimizing patient management. Expected outcomes include validated AI models, scientific publications, and practical tools for ICU decision support, ultimately improving patient safety, reducing days on mechanical ventilation, ICU length of stay, and enhancing resource utilization. The project aligns with the broader goals of advancing personalized medicine and supporting sustainable healthcare systems.
Project funding:
Research Council of Lithuania, Projects carried out by researchers’ teams
Project results:
A validated, time-dependent machine learning model was designed to predict complex weaning from MV, reducing ICU length of stay by 20-30% and reintubation rate by 15-20%.
A methodology for integrating interpretive artificial intelligence and counterfactual argument generation in clinical prediction.
Insights into the temporal patterns of biosignal patterns associated with weaning success or failure.
Assessment metrics, risk factors, and functional importance findings that contribute to the scientific understanding of intensive care unit decision support.
Period of project implementation: 2025-12-01 - 2028-11-30
Project coordinator: Kaunas University of Technology
