Abstract
<sec> <title>BACKGROUND</title> <p>Early detection of patient deterioration is critical to patient safety in acute care hospitals. While electronic health records (EHRs) and early warning systems such as the Queensland Adult Deterioration Detection System (QADDS) support routine monitoring, these systems are reactive by design, prompting a response only once physiological thresholds have been breached. Artificial intelligence (AI) driven clinical decision support (CDS) tools offer potential to anticipate these triggers and extend the window for intervention yet evidence to guide real-world implementation remains limited.</p> </sec> <sec> <title>OBJECTIVE</title> <p>This study aimed to understand existing workflows for detecting and managing patient deterioration in a tertiary hospital setting to inform the pilot implementation of a machine learning-based CDS tool, Predicting Risk of Deterioration (PRoD). Specifically, the study sought to identify optimal integration points and intended end users, evaluate clinician feedback on the prototype’s usability and value, and explore contextual factors influencing implementation and adoption.</p> </sec> <sec> <title>METHODS</title> <p>A contextual inquiry study was conducted within the medical emergency team (MET) and critical care outreach service of a tertiary hospital in Queensland, Australia. Data collection comprised desk-based policy analysis, in-situ observation of clinical practice, usability testing of the prototype using simulated deterioration scenarios and semi-structured interviews. Workflows were mapped using the Actor, Action, Context, Target, Time (AACTT) framework. Qualitative data were analysed using narrative synthesis, grounded theory-informed usability analysis and AI-assisted inductive thematic analysis. Themes related to implementation were mapped to the Consolidated Framework for Implementation Research (CFIR) and the identified constructs were aligned with Expert Recommendations for Implementing Change (ERIC) strategies to derive actionable recommendations to support implementation.</p> </sec> <sec> <title>RESULTS</title> <p>Ten clinicians (5 doctors, 5 nurses) participated. A total of 44 workflow steps across seven phases were identified. Observations demonstrated that clinical decision-making relied heavily on senior clinician expertise, contextual patient knowledge and synthesis of fragmented information across multiple systems. Participants identified the greatest potential value of PRoD during routine patient review, handover and after-hours care rather than during active emergencies. Usability testing generated four themes emphasising seamless workflow integration, role-specific dashboard views, transparent and clinically actionable risk explanations and minimisation of alert fatigue. Inductive analysis produced three themes and nine subthemes which mapped to 17 CFIR constructs spanning staffing pressures, communication gaps, documentation quality, governance and information technology infrastructure. Alignment with ERIC strategies informed a staged, workflow-sensitive pilot implementation plan.</p> </sec> <sec> <title>CONCLUSIONS</title> <p>This formative study demonstrated that the effective use of AI-enabled deterioration prediction depends as much on socio-technical alignment as on predictive performance. PRoD was perceived as offering greatest value when positioned as a non-mandatory, prognostic decision-support tool integrated into ongoing clinical review process and handover workflows. By grounding design and implementation planning in observed practice, this study provided practical guidance for responsible adoption of predictive tools in acute care.</p> </sec>