Designing clinically actionable visualisations of the AI Clinician: an exploratory study of ICU clinician preferences in sepsis management

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Introduction Artificial intelligence (AI) offers significant opportunities to support complex, time-critical decisions in intensive care units (ICUs). However, clinical adoption depends not only on predictive accuracy, but also on whether AI recommendations are visualised in formats that are interpretable, actionable, and trusted by clinicians. Poorly designed interfaces risk undermining usability, trust, and integration into workflows. This study investigates ICU clinicians’ preferences for graphical user interface (GUI) design and explainability strategies in the AI Clinician, a reinforcement learning tool for sepsis management. Methods An online survey was conducted between November 2023 and November 2024 with 24 ICU clinicians from UK hospitals. Participants evaluated three GUI prototypes of the AI Clinician’s treatment recommendations; 1) discrete numerical doses, 2) dose ranges, and 3) qualitative increase/decrease guidance. Clinicians rated each prototype’s usefulness and ease of use using Likert scales, completed pairwise comparisons, and provided free-text feedback. They also ranked four types of AI explanations (treatment option gap, projected mortality, feature importance, and influential training examples) according to their perceived benefit. Descriptive statistics summarised quantitative data, while thematic analysis was applied to qualitative responses. Results Clinicians expressed a strong preference for the range-based prototype, which best balanced clarity with flexibility for patient-specific adjustments. Prototype 2 (ranges) achieved the highest ratings for meeting requirements (median=6/7) and ease of use (median=6/7) and was preferred in 77.3% of pairwise comparisons. Discrete numbers were considered precise but overly rigid, while qualitative increase/decrease guidance was simple but too vague for critical decisions. Treatment option gap (mean=4.04) and projected mortality (mean=3.71) were ranked most beneficial for explainability, although clinicians emphasised that multiple, complementary, explanations were necessary to enhance transparency and trust. Discussion The findings suggest that AI recommendations for sepsis management should be presented using range-based visualisations that reflect real-world clinical practice by supporting flexibility without sacrificing clarity. Multiple explanation types, particularly those tied directly to treatment rationale and patient outcomes, are essential for building trust. These insights highlight the importance of human-centred design, clinician involvement in development, and layered transparency to optimise AI clinical decision support systems in critical care.