APPLICATION OF ARTIFICIAL INTELLIGENCE IN INTENSIVE CARE: CLINICAL POTENTIAL AND CHALLENGES

Vollständiger Abstract

Worum geht es in dieser Arbeit?

Background: Contemporary intensive care units (ICUs) generate vast amounts of real-time data, leading to staff cognitive overload and the phenomenon of alarm fatigue. Traditional prognostic scoring systems (APACHE II, SOFA) are inherently static and fail to account for dynamic clinical trends. Artificial intelligence (AI) algorithms, including machine learning (ML) and deep learning (DL), offer novel capabilities for continuous monitoring and early prediction of life-threatening conditions. Objective: To analyze the clinical potential and implementation challenges associated with the deployment of AI algorithms in intensive care settings. Methods: A comprehensive literature review was conducted across the PubMed, Embase, and IEEE Xplore databases for the years 2018–2026, focusing on studies utilizing multicenter databases (such as MIMIC-IV and eICU). Results: AI algorithms demonstrate high discriminative performance (AUROC 0.79–0.96) in the early detection of sepsis, prediction of hemodynamic instability with a 5-to-15-minute lead time, and dynamic mortality risk estimation. In the context of mechanical ventilation, deep learning models (specifically convolutional neural networks, CNNs) enable automated detection of patient-ventilator asynchrony with an accuracy exceeding 90%, as well as precise forecasting of extubation success. Conclusions: Artificial intelligence holds immense potential for optimizing ICU patient care. However, primary implementation barriers persist, including the lack of model interpretability (the "black box" problem), domain shift, and ethico-psychological dilemmas, such as clinician moral distress and the risk of care dehumanization. AI systems should function strictly as clinical decision support tools under physician supervision.