Spine patient history-taking: a two phase study of a prompt-engineered large language model assistant

Large language models (LLMs) are transforming medicine, yet their use has been largely confined to back-end administrative tasks, leaving their potential for real-time patient interaction critically unexplored. This study aimed to address this gap by developing and evaluating Spine-GPT, a novel AI assistant built on multi-layered prompt engineering to conduct pre-consultation interviews in the field of spine surgery and enhance clinical decision-making.

A two-phase study was conducted. Phase 0 involved simulation-based evaluation using 24 fictional clinical scenarios assessed by three board-certified spine surgeons across seven domains (adaptive questioning, complaint comprehension, diagnostic accuracy, justification clarity, investigation appropriateness, red-flag recognition, and professional tone) on 5-point Likert scales. Phase 1 was a prospective clinical pilot study involving 60 patients, allocated to a control group (n = 30) receiving conventional consultations and an intervention group (n = 30) using Spine-GPT before face-to-face appointment. The primary outcome was active surgeon history-taking time. Secondary outcomes included information completeness, red-flag detection, surgeon and patient satisfaction.

In simulations, Spine-GPT demonstrated high performance, with mean scores exceeding 4.0 out of 5 across all evaluation domains. In the clinical study, the AI-assisted group achieved a 31.3% reduction in active surgeon history-taking time (11.47 min vs. 7.88 min; p < 0.001). This efficiency gain was achieved alongside a significant increase in information completeness (11.7%-point increase; p < 0.001). All three predefined red-flag cases were correctly identified. Surgeon and patient satisfaction were significantly higher in the intervention cohort (p < 0.001), with 70% of the AI-generated summaries requiring no edits.

This study provides proof-of-concept that a prompt-engineered LLM can safely increase clinical efficiency while simultaneously improving the quality of data collection in a specialized surgical practice. By structuring its reasoning process to be transparent, the system functions as an explainable "probabilistic partner" rather than an uninterpretable "black box," helping to bridge the gap between automated inference and clinician trust. This work ultimately positions the clinician not as a passive bystander to automation, but as a critical editor and beneficiary of machine-reasoning, supporting the new paradigm for human-AI collaboration in clinical care.