Extraction and appraisal workbook for a systematic review of large language models in first-contact primary health care

Study-level extraction, quality appraisal and risk-of-bias data for a systematic review of large language models (LLMs) used for diagnostic, triage and referral decision support in first-contact primary health care. The dataset covers 71 studies published in English between 1 January 2022 and 1 June 2026, identified by searching PubMed/MEDLINE, Embase (Ovid), Scopus and Web of Science Core Collection on 2 June 2026, together with backward citation searching. Nine sheets hold the analytic record. Descriptive extraction of study characteristics, technology, evaluation design and outcomes. A five-domain appraisal developed by the author covering clinical realism, trustworthiness, safety, equity and implementation readiness. A reporting-framework appraisal scored on a common eight-item set across MI-CLAIM, STARD-AI, TRIPOD-LLM, DECIDE-AI and FUTURE-AI. A corroborative risk-of-bias layer using QUADAS-2 and PROBAST+AI. Model-level results for studies evaluating more than one model or metric, 334 rows in total. A summary matrix, a full codebook defining every field and banding rule, and reviewer notes recording each borderline decision and its resolution. Extraction and appraisal were conducted by a single reviewer with a documented validation and correction pass. All ratings are the author's own judgements applied to published study reports and record reporting adequacy rather than the quality of the technologies evaluated. The workbook contains no personal or identifiable data. Files are provided as an Excel workbook and as one CSV per sheet with formulas resolved to values. The dataset accompanies an MSc Public Health project report submitted to the London School of Hygiene and Tropical Medicine in September 2026.

Keywords