VIRTUAL HEALTHCARE ASSISTANTS IN TELEMEDICINE: A STRUCTURED EVIDENCE REVIEW OF EFFICACY CLAIMS AND ASSURANCE REQUIREMENTS

Abstract

Virtual healthcare assistants (VHAs) built on large language models (LLMs) are widely promoted as a remedy for administrative burden and access constraints in telemedicine, and the effect sizes attached to them are striking. This paper asks what the peer-reviewed evidence supports. Following PRISMA 2020 reporting guidance, 117 records were identified through five structured Crossref queries and targeted searching of regulatory, standards and journal registries; 8 duplicates were removed, leaving 109 unique records screened on title and abstract. Seventy-three were excluded at screening, 36 were assessed for eligibility, 12 were excluded with reasons, and 24 were retained. Because retrieval relied on Crossref rather than on MEDLINE, Embase, Scopus or Web of Science, this work is reported as a structured evidence review rather than as an exhaustive systematic review. Retained studies were graded by evidence tier and appraised for risk of bias using RoB 2 and QUADAS-2. The principal finding concerns provenance. Five widely circulated figures — 45 percent administrative burden reduction, 40 percent diagnostic accuracy uplift, 30 percent hospitalisation reduction, 35 percent satisfaction improvement and US$150 billion in annual savings — were traced to source. Two were located in vendor material but could not be reconciled with any primary evaluation, two could not be traced to any source at all, and the fifth is a 2017 consultancy projection for the United States rather than a measured global outcome. Controlled measurement reports self-triage accuracy of 57.8 to 76.0 percent for LLMs and 47.3 to 62.4 percent for laypeople as ranges across included studies, and symptom checkers listing the correct diagnosis first in 34 percent of vignettes. Randomised evidence on remote patient management is mixed rather than uniformly negative. The paper proposes a conceptual five-layer reference architecture whose principal element is an explicit clinical assurance layer.