Safety-First Framework for AI-Enabled Anamnesis in Head and Neck Surgery: Evidence Synthesis from a Narrative Review

Objectives: To synthesize evidence on artificial intelligence (AI)-enabled medical history taking (anamnesis)-beyond large language models (LLMs) alone-and to translate findings into implications and research priorities for head and neck surgery. Methods: We performed a PRISMA-informed narrative review. Searches from database inception to 31 December 2025 (updated 3 January 2026) were conducted in MEDLINE (PubMed), Embase, Scopus, Web of Science Core Collection, IEEE Xplore, and ACM Digital Library, supplemented by medRxiv/arXiv screening and citation chasing. We included studies evaluating or describing AI-supported history capture/summarization, conversational interviewing, symptom checker/digital triage, EHR-integrated intake-to-decision support pipelines, voice interviewing, education/training systems, and governance/ethical considerations related to digital anamnesis. Findings were synthesized by system category and by cross-cutting outcome domains, with a head and neck surgery interpretive lens. Results: Fifty studies (2014-2025) were included. Evidence most consistently suggested feasibility and acceptability of pre-consultation computer-assisted history taking and the potential to reduce documentation burden and improve structured capture. In contrast, symptom checkers and digital triage tools showed highly variable diagnostic/triage performance and prominent safety concerns, highlighting the importance of conservative red-flag escalation strategies, continuous monitoring, and clear accountability. LLM-based diagnostic dialogue demonstrated strong performance in controlled evaluations, but prospective real-world validation, governance, and workflow integration remain limited. Conclusions: AI-enabled anamnesis comprises heterogeneous tools with uneven evidence. For head and neck surgery, potential near-term applications may include structured pre-visit intake, clinician-facing summarization, and training applications, whereas autonomous triage warrants harm-oriented, specialty-calibrated validation and robust governance prior to broader clinical reliance.