人工智慧眼底鏡篩檢糖尿病視網膜病變與老年性黃斑部病變之成本效用分析

Abstract

隨著糖尿病視網膜病變（Diabetic Retinopathy, DR）及年齡相關性黃斑部病變（Age-related Macular Degeneration, AMD）盛行率持續上升，相關視力損失已成為重要公共衛生議題。人工智慧（Artificial Intelligence, AI）輔助眼底影像判讀可提升篩檢效率與早期診斷能力，但其於臺灣醫療體系之經濟效益仍缺乏本土實證。本研究以臺灣全民健康保險支付者觀點，比較AI輔助眼底篩檢（VeriSee DR與VeriSee AMD）與非AI篩檢策略之成本效用。

本研究採決策樹結合馬可夫模型，模擬50歲族群終身疾病進程，以品質調整生命年（Quality-Adjusted Life Years, QALYs）作為健康效益指標，支付意願門檻設定為新臺幣3,000,000元/QALY。模型參數主要參考臺大醫院人工智慧眼底篩檢臨床試驗資料、本土健保資料及國內外文獻，並進行單因子及機率敏感度分析。

研究結果顯示，AI輔助眼底篩檢策略之總成本為新臺幣345,295.85元，較非AI篩檢增加1,726.62元；健康效益為15.34 QALYs，較非AI篩檢增加1.59 QALYs，增量成本效用比（Incremental Cost-Utility Ratio, ICUR）為1,085.92元/QALY，遠低於本研究設定之支付意願門檻3,000,000元/QALY，顯示AI輔助篩檢具有成本效用。敏感度分析結果顯示，模型主要受AI對AMD之特異度、AMD治療成本、AI對DR之特異度及DR治療成本等參數影響；機率敏感度分析亦顯示，當支付意願門檻約達6,000元/QALY時，AI輔助篩檢之成本效用機率已接近100%，顯示研究結果具有良好穩健性。

綜合而言，AI輔助眼底篩檢於臺灣醫療體系具有良好成本效用，可提升健康效益並促進醫療資源有效配置，研究結果可作為未來人工智慧醫療器材健康科技評估及健保給付政策之參考。

As the prevalence of diabetic retinopathy (DR) and age-related macular degeneration (AMD) continues to increase, vision impairment caused by these diseases has become a major public health concern. Artificial intelligence (AI)-assisted fundus image interpretation has the potential to improve screening efficiency and facilitate early disease detection; however, evidence regarding its economic value within the Taiwanese healthcare system remains limited. This study evaluated the cost-utility of AI-assisted fundus screening (VeriSee DR and VeriSee AMD) compared with conventional non-AI screening from the perspective of the Taiwan National Health Insurance (NHI).

A decision tree combined with a Markov model was developed to simulate the lifetime health outcomes of a hypothetical cohort starting at 50 years of age. Health outcomes were measured in quality-adjusted life years (QALYs), and the willingness-to-pay (WTP) threshold was set at NT$3,000,000 per QALY. Model parameters were derived from clinical trial data provided by National Taiwan University Hospital, Taiwan National Health Insurance data, and published literature. One-way and probabilistic sensitivity analyses were performed to evaluate model uncertainty. The results showed that the AI-assisted screening strategy had a total cost of NT$345,295.85 per person, representing an incremental cost of NT$1,726.62 compared with the non-AI screening strategy, while generating an additional 1.59 QALYs. The incremental cost-utility ratio (ICUR) was NT$1,085.92 per QALY, which was substantially below the predefined WTP threshold of NT$3,000,000 per QALY, indicating that AI-assisted screening was a cost-effective strategy. Sensitivity analyses demonstrated that the model was most sensitive to the specificity of AI for AMD, followed by AMD treatment costs, the specificity of AI for DR, and DR treatment costs. Probabilistic sensitivity analysis further showed that the probability of AI-assisted screening being cost-effective approached 100% at a WTP threshold of approximately NT$6,000 per QALY, supporting the robustness of the findings.

In conclusion, AI-assisted fundus screening is a cost-effective strategy within the Taiwanese healthcare system. It has the potential to improve health outcomes and optimize healthcare resource allocation. The findings of this study may serve as evidence to support future health technology assessment (HTA) and National Health Insurance reimbursement decisions for AI-based medical devices.