AI-Assisted Clinical Decision Support Versus Standard Teleconsultation in Adult Medicine: A Systematic Review of Diagnostic Accuracy, Referral and Testing Decisions, and Clinician Confidence

Abstract

Background/Objectives: Artificial intelligence (AI) is increasingly integrated into teleconsultation, but its effects on clinician performance remain uncertain. We assessed whether clinician-facing AI improves diagnostic accuracy, reduces referrals or testing, and increases clinician confidence compared with standard teleconsultation or conventional resources. Methods: This retrospectively registered systematic review followed the PRISMA 2020 and Synthesis Without Meta-Analysis (SWiM) guidance. PubMed/MEDLINE, publisher-indexed platforms, Google Scholar, ClinicalTrials.gov, and backward and forward citation searches were performed from inception to 1 August 2026. Eligible studies evaluated clinician-facing AI in adult primary, urgent, or general medical teleconsultation or in high-fidelity virtual consultation designs and reported diagnostic, referral/testing, or confidence outcomes. One reviewer screened records, extracted data, and assessed the risk of bias using the design-appropriate QUADAS-2, RoB 2, and ROBINS-I domains, with a second pass for consistency. Owing to clinical and methodological heterogeneity, the results were synthesized narratively. Results: Of 31 records identified, 25 remained after duplicate removal, 18 full texts were assessed, and six studies were included: two randomized virtual experiments, two retrospectives virtual care cohorts, one randomized crossover conversational AI study, and one provider-to-provider telemedicine case series. In a telehealth screening experiment, AI improved diagnostic prediction, but physicians given a large language model showed no significant benefit (adjusted difference, two percentage points; 95% CI, −4 to 8). Across 102,059 virtual encounters, an AI-generated differential included the clinician’s final diagnosis in 84.2%, although the reference was not independent. AI-assisted urgent care recommendations were rated optimal more often than final physician recommendations (77.1% vs. 67.1%). AI did not reduce unnecessary referrals; one experiment found more confirmatory testing requests and lower trust than with a familiar clinical score. Certainty was low or very low for the principal clinical outcomes. Conclusions: AI may support teleconsultation decisions, but evidence that it reduces unnecessary referrals or meaningfully increases clinician confidence is insufficient. Pragmatic trials using independent patient-level reference standards are needed. Registration: This completed review was retrospectively registered on the Open Science Framework (OSF); no protocol was published before study selection.