An agentic AI workflow for spine history taking and surgeon-facing clinical synthesis

<title>Abstract</title> <p>Adequate history-taking is central to safe spine triage but is often constrained by outpatient time pressure. We developed SpineGPT-2, an agentic history-taking system combining a patient-facing interview agent, a deterministic orchestration layer enforcing safety and history coverage, and a surgeon-facing agent that produces a structured, retrieval-grounded summary. In a prospective single-clinic study, 240 patients were allocated by alternating clinic days to the model arm (n = 120) or standard consultation (n = 120). An independent expert panel rated summary adequacy on a five-item scale. Mean adequacy was 4.23 (95% CI, 4.10–4.36), significantly exceeding the pre-specified actionable threshold of 4.0. The panel's leading diagnosis fell within the model's top-three differentials in 88.3% of cases; transcript-based red-flag sensitivity and specificity were 93.8% and 96.2%. Physician-led history-taking time fell from 10.22 to 6.77 minutes. Hallucination occurred in 8.3% of summaries. Agentic, safety-first history-taking produced clinically adequate surgeon summaries, supporting supervised pre-consultation use.</p>