Patient safety and health equity of AI-enabled triage tools in primary care: a systematic review
5 affiliations
- 1Division of Population Health, Health Services Research and Primary Care, School of Health Sciences, Faculty of Biology, Medicine and Health, The University of Manchester, Manchester, UK
- 2Department of Public Health, Health Informatics Program, School of Nursing and Health Sciences, Jazan University, Jazan, Saudi Arabia
- 3Division of Informatics, Imaging and Data Sciences, School of Health Sciences, Faculty of Biology, Medicine and Health, The University of Manchester, Manchester, UK
- 4Division of Family Medicine, Yong Loo Lin School of Medicine, National University of Singapore, Singapore
- 5Centre for Research in Health Systems Performance, National University of Singapore, Singapore
Abstract
Objectives: This review synthesised real-world patient safety evidence for artificial intelligence (AI)-enabled triage tools in primary care and mapped whether safety has been examined across sociodemographic groups.
Methods: PubMed, Embase, Web of Science and Scopus were searched (January 2015–November 2025) with citation searching. Prospective studies evaluating AI triage tools in real-world primary care settings were eligible. Quality was assessed using the Mixed Methods Appraisal Tool; equity dimensions were mapped using PROGRESS-Plus. Random-effects meta-analysis pooled comparable proportion-based outcomes.
Results: Four studies were included; all evaluated rule-based or decision-tree tools as no machine learning or generative AI system met the eligibility criteria. Safe triage rates ranged from 81% to 97.6%, under-triage rates ranged from 2% to 17.2% and over-triage rates ranged from 8% to 29%. An exploratory random-effects meta-analysis of three studies (n=1303) gave a pooled safe triage proportion of 0.908 (95% CI 0.319 to 0.995) with substantial heterogeneity (I²=97.0%); given incompatible outcome definitions across studies, these estimates are hypothesis-generating rather than summary measures. No study analysed differential triage accuracy by any sociodemographic factor despite available demographic data.
Discussion: Safety outcomes were confined to technical performance proxies rather than direct measures of harm, and the absence of equity-stratified analysis may reflect a systemic evidence gap.
Conclusion: To our knowledge, this is the first systematic review to examine both real-world patient safety and equity of AI-enabled triage in primary care; findings may not extend to machine learning or generative AI systems. A Safety-Equity Evidence Mapping (SEEM) matrix and an evidence maturity framing are proposed to characterise this evidence and inform emerging consensus on evidence thresholds.
PROSPERO registration number: CRD420251179046.
Related
- Clinical, safety, and patient-reported outcomes associated with second-line and later therapies in primary biliary cholangitis: A systematic review
- Switching Beyond Rituximab in Primary Glomerular Diseases: A Systematic Review of Within-Patient Efficacy and Safety of Non-Rituximab Anti-CD20 Antibodies.
- Artificial intelligence-enabled mental health applications for adolescents: a scoping review of safety, effectiveness and clinical implementation
- Comparative efficacy and safety of targeted disease-modifying therapies in adults with primary IgA nephropathy: a systematic review and network meta-analysis
- PARP inhibitors and newly second primary malignancies in cancer patients: a systematic review and safety meta-analysis of placebo randomized controlled trials
BibTeX
@article{alamoudi2026patient,
title = {{Patient safety and health equity of AI-enabled triage tools in primary care: a systematic review}},
author = {Aymn Alamoudi and Yanying Gai and Salwa S Zghebi and Evangelos Kontopantelis and Benjamin C Brown},
journal = {BMJ Health \& Care Informatics},
year = {2026},
month = oct,
doi = {10.1136/bmjhci-2026-102427},
url = {https://doi.org/10.1136/bmjhci-2026-102427},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - Patient safety and health equity of AI-enabled triage tools in primary care: a systematic review
AU - Aymn Alamoudi
AU - Yanying Gai
AU - Salwa S Zghebi
AU - Evangelos Kontopantelis
AU - Benjamin C Brown
PY - 2026
DA - 2026/10/28
JO - BMJ Health & Care Informatics
DO - 10.1136/bmjhci-2026-102427
UR - https://doi.org/10.1136/bmjhci-2026-102427
AB - Objectives: This review synthesised real-world patient safety evidence for artificial intelligence (AI)-enabled triage tools in primary care and mapped whether safety has been examined across sociodemographic groups. Methods: PubMed, Embase, Web of Science and Scopus were searched (January 2015–November 2025) with citation searching. Prospective studies evaluating AI triage tools in real-world primary care settings were eligible. Quality was assessed using the Mixed Methods Appraisal Tool; equity dimensions were mapped using PROGRESS-Plus. Random-effects meta-analysis pooled comparable proportion-based outcomes. Results: Four studies were included; all evaluated rule-based or decision-tree tools as no machine learning or generative AI system met the eligibility criteria. Safe triage rates ranged from 81% to 97.6%, under-triage rates ranged from 2% to 17.2% and over-triage rates ranged from 8% to 29%. An exploratory random-effects meta-analysis of three studies (n=1303) gave a pooled safe triage proportion of 0.908 (95% CI 0.319 to 0.995) with substantial heterogeneity (I²=97.0%); given incompatible outcome definitions across studies, these estimates are hypothesis-generating rather than summary measures. No study analysed differential triage accuracy by any sociodemographic factor despite available demographic data. Discussion: Safety outcomes were confined to technical performance proxies rather than direct measures of harm, and the absence of equity-stratified analysis may reflect a systemic evidence gap. Conclusion: To our knowledge, this is the first systematic review to examine both real-world patient safety and equity of AI-enabled triage in primary care; findings may not extend to machine learning or generative AI systems. A Safety-Equity Evidence Mapping (SEEM) matrix and an evidence maturity framing are proposed to characterise this evidence and inform emerging consensus on evidence thresholds. PROSPERO registration number: CRD420251179046.
SN - 2632-1009
ER -