Beyond Human Error: Building Intelligent Resilience for Medication Safety in the ICU

Abstract

Highlights: What are the main findings? Traditional voluntary reporting systems in the ICU miss approximately 98% of medication errors compared to active observation, creating a dangerous surveillance gap. We propose a five-layer Intelligent Safety Stack that integrates standardised data, intelligent surveillance, signal optimisation, generative AI, and physical engineering controls to actively detect and intercept errors beyond the limits of human vigilance. What are the implications of the main findings? Safety performance should move beyond raw error counts to include rescue rates—the proportion of risks successfully intercepted before patient harm. Sustainable safety involves a sociotechnical strategy that addresses implementation barriers such as alert fatigue and data fragmentation rather than relying on adding more digital tools to an already complex workflow. Background/Objectives: Medication errors (MEs) in intensive care units (ICUs) remain a persistent threat to patient safety. A significant surveillance gap exists where traditional voluntary reporting detects as few as 0.02 MEs per patient-day, leaving approximately 98% of errors invisible to standard audits. This review critically examines how artificial intelligence (AI) and implementation science can bridge this gap through a proposed five-layer Intelligent Safety Stack. Methods: We conducted a critical narrative review of the peer-reviewed literature published between 2000 and 2025, synthesising evidence across medication safety, predictive analytics, generative AI, engineering controls, and sociotechnical frameworks. Results: Reported ME incidence varies widely (1.32% to 31.7%) due to the profound methodological heterogeneity. To achieve sustainable safety, we propose a five-layer framework: (1) Standardised Ontology (e.g., NCC MERP) to establish ground-truth data; (2) Intelligent Surveillance to identify and monitor high-risk patients; (3) Signal Optimisation to filter noise and reduce alert fatigue; (4) Generative Stewardship to automate reconciliation at transitions of care; and (5) Engineering Controls (smart pump interoperability and NRFit™), which have been shown to reduce administration errors by up to 54.8%. Conclusions: Isolated error counting is insufficient. Sustainable medication safety in the ICU involves a sociotechnical fusion of the Intelligent Safety Stack with success measured by rescue rates rather than error prevalence alone.