Explainable Interface Design for Human–AI Collaboration in Confocal Laser Endomicroscopy for Head and Neck Cancer Surgery: Mixed Methods Study (Preprint)

BACKGROUND Confocal laser endomicroscopy (CLE) enables real-time, in vivo tissue assessment during head and neck cancer surgery, where the balance between oncological safety and functional preservation is particularly delicate. Artificial intelligence (AI) can support CLE interpretation, yet clinical adoption remains limited, a gap attributed less to model performance than to a lack of interfaces that make AI outputs interpretable and trustworthy at the point of care. OBJECTIVE This study aimed to investigate how explanatory, user-centered interfaces can enable reliable human–AI collaboration for AI-assisted CLE across the surgical workflow. METHODS Following a user-centered design approach, a user journey map informed the development of three high-fidelity prototypes addressing intraoperative, pre-/postoperative, and educational use cases. A mixed-methods evaluation with five experienced surgeons from four institutions combined think-aloud exploration, the short User Experience Questionnaire (UEQ-S), and semi-structured interviews; qualitative data were analysed inductively through thematic analysis. RESULTS Explainability, conveyed through influential morphological features, class activation maps, and example-based case comparisons, anchored AI recommendations in surgeons’ established clinical reasoning and supported verification against their own judgement. Explicit communication of model uncertainty emerged as a distinct prerequisite for calibrated trust. Clarity, simplicity, and adaptive information density were identified as central design principles, while diverging preferences pointed to a need for configurable interaction modes. All three prototypes were rated as excellent on the UEQ-S (overall means 2.53–2.68). CONCLUSIONS Trustworthy integration of AI into CLE depends primarily on explanatory interfaces aligned with clinical reasoning and on transparent communication of uncertainty, rather than on model performance alone. Configurable, workflow-adaptive designs and case-based educational applications represent promising directions for clinical translation.