key: cord-0029241-54fqg1pu authors: Gomez Rossi, Jesus; Rojas-Perilla, Natalia; Krois, Joachim; Schwendicke, Falk title: Cost-effectiveness of Artificial Intelligence as a Decision-Support System Applied to the Detection and Grading of Melanoma, Dental Caries, and Diabetic Retinopathy date: 2022-03-15 journal: JAMA Netw Open DOI: 10.1001/jamanetworkopen.2022.0269 sha: 9fe54dc33e5d310956b48e59e7b4a51306aacedd doc_id: 29241 cord_uid: 54fqg1pu OBJECTIVE: To assess the cost-effectiveness of artificial intelligence (AI) for supporting clinicians in detecting and grading diseases in dermatology, dentistry, and ophthalmology. IMPORTANCE: AI has been referred to as a facilitator for more precise, personalized, and safer health care, and AI algorithms have been reported to have diagnostic accuracies at or above the average physician in dermatology, dentistry, and ophthalmology. DESIGN, SETTING, AND PARTICIPANTS: This economic evaluation analyzed data from 3 Markov models used in previous cost-effectiveness studies that were adapted to compare AI vs standard of care to detect melanoma on skin photographs, dental caries on radiographs, and diabetic retinopathy on retina fundus imaging. The general US and German population aged 50 and 12 years, respectively, as well as individuals with diabetes in Brazil aged 40 years were modeled over their lifetime. Monte Carlo microsimulations and sensitivity analyses were used to capture lifetime efficacy and costs. An annual cycle length was chosen. Data were analyzed between February 2021 and August 2021. EXPOSURE: AI vs standard of care. MAIN OUTCOMES AND MEASURES: Association of AI with tooth retention–years for dentistry and quality-adjusted life-years (QALYs) for individuals in dermatology and ophthalmology; diagnostic costs. RESULTS: In 1000 microsimulations with 1000 random samples, AI as a diagnostic-support system showed limited cost-savings and gains in tooth retention–years and QALYs. In dermatology, AI showed mean costs of $750 (95% CI, $608-$970) and was associated with 86.5 QALYs (95% CI, 84.9-87.9 QALYs), while the control showed higher costs $759 (95% CI, $618-$970) with similar QALY outcome. In dentistry, AI accumulated costs of €320 (95% CI, €299-€341) (purchasing power parity [PPP] conversion, $429 [95% CI, $400-$458]) with 62.4 years per tooth retention (95% CI, 60.7-65.1 years). The control was associated with higher cost, €342 (95% CI, €318-€368) (PPP, $458; 95% CI, $426-$493) and fewer tooth retention–years (60.9 years; 95% CI, 60.5-63.1 years). In ophthalmology, AI accrued costs of R $1321 (95% CI, R $1283-R $1364) (PPP, $559; 95% CI, $543-$577) at 8.4 QALYs (95% CI, 8.0-8.7 QALYs), while the control was less expensive (R $1260; 95% CI, R $1222-R $1303) (PPP, $533; 95% CI, $517-$551) and associated with similar QALYs. Dominance in favor of AI was dependent on small differences in the fee paid for the service and the treatment assumed after diagnosis. The fee paid for AI was a factor in patient preferences in cost-effectiveness between strategies. CONCLUSIONS AND RELEVANCE: The findings of this study suggest that marginal improvements in diagnostic accuracy when using AI may translate into a marginal improvement in outcomes. The current evidence supporting AI as decision support from a cost-effectiveness perspective is limited; AI should be evaluated on a case-specific basis to capture not only differences in costs and payment mechanisms but also treatment after diagnosis. Sensitivities and specificities of disease detection with and without AI assistance were derived from the economic analysis and compared with the existing meta-analysis. If possible, we calculated mean values and 95% confidence intervals or ranges to estimate normal distributions (in parentheses) for random sampling during microsimulation. Sensitivities and specificities of disease detection with and without AI assistance were derived from the economic analysis and compared with the existing meta-analysis. If possible, we calculated mean values and 95% confidence intervals or ranges for random sampling during microsimulation. *The population modelled consists of already diabetic patients, reason why utility even without DR is never 1. Transition probability to other stages can be sudden, however, patients are assumed to already accrue costs for their standard treatment. Adherence to the American Diabetes Association retinal screening guidelines for population with diabetes in the United States Health-related quality of life associated with diabetic retinopathy in patients at a public primary care service in southern Brazil Stage-specific survival and recurrence in patients with cutaneous malignant melanoma in Europe-a systematic review of the literature An estimate of the annual direct cost of treating cutaneous melanoma device US Food and Drug Administration