Cost-Effectiveness of an Artificial Intelligence as a Medical Device (AIaMD) for Triaging Patients Presenting to Primary Care with Concerns that They Have a Skin Cancer: A Modelling Study.

INTRODUCTION: Health services are struggling to cope with the growing numbers of people coming with skin lesions they are worried could be cancer. Using artificial intelligence (AI) to assist in the triage process is one potential approach. Deep Ensemble for the Recognition of Melanoma (DERM) is one such AI device which has achieved Conformité Européene (CE) Class III regulatory approval in the UK and Europe. The effectiveness and cost-effectiveness of DERM have already been carefully examined post-referral. This study examines the cost-effectiveness of DERM pre-referral in the community setting. METHODS: We modified a decision-analytic model developed for post-referral evaluation of DERM to the community setting. The model compared standard care (SC) with two DERM-based strategies: with second reading (2R) when discharge was recommended (DERM_2R) and without it (DERM_autonomous). The outputs were numbers of diagnostic outcomes, numbers of key events (GP appointments, DERM assessments, hospital referrals), costs and quality-adjusted life-years (QALY). The cost-effectiveness metrics were incremental cost per QALY gained. Analysis was conducted in the UK from an NHS and personal social services perspective, with a life-time horizon up to a maximum of 100 years, with uncertainty assessed mainly using deterministic sensitivity analyses. RESULTS: We report three base-case results on the basis of different assumptions regarding GP diagnostic sensitivity and specificity. When GPs operate with maximised sensitivity as suggested by the literature, SC is the most cost-effective option. SC dominates DERM_autonomous and DERM_2R has an incremental cost-effectiveness ratio (ICER) relative to SC of £30,370. If GPs operate with maximised specificity, DERM_autonomous dominates SC and DERM_2R has an ICER of £397. If GPs diagnose cancer with a performance back calculated from NHS routine data, the cost-effectiveness of both DERM options dominate SC. CONCLUSIONS: The priority should be for better estimates of GP accuracy to be obtained to reduce uncertainty in the estimates of cost-effectiveness. The results presented do, however, show that it is plausible that DERM could be cost-effective justifying use in practice with further data collection.