Background
Glaucoma is the second leading cause of blindness after cataracts, affecting 13.12 million people in China [
1,
2]. Among them, primary angle-closure glaucoma (PACG) is the predominant type of glaucoma, accounting for 7.14 million patients with a prevalence of 1.40% among China populations aged over 45 [
2]. The number of PACG patients in China is projected to reach 7.5 million in 2020, surpassing that in India and ranking first in the world [
3]. Considering the irreversibility of vision and visual field loss in glaucoma, early detection and intervention are essential for glaucoma management, which can effectively delay optic nerve damage [
4,
5]. However, as the initial phase of glaucoma is usually asymptomatic, over 90% of patients in China remain unaware of their conditions until in late stage [
6,
7]. Once the disease progress to late stage, the effects of medical interventions are unsatisfactory, underpinning the importance of population-based screening to identify asymptomatic glaucoma patients in communities and subsequent timely referral and treatment.
Glaucoma screening requires manually assessing optic nerve structure from digital fundus image. The process is labor-intensive and time-consuming, and the accuracy of diagnosis heavily relies on the skill and experience of ophthalmologists [
8]. Currently, the problem with population-based screening for glaucoma is lack of ophthalmologists and poor ophthalmology capability at grassroot hospitals, making community screening difficult to implement [
9]. The past several years has witnessed significant technology advancement of artificial intelligence (AI) in glaucoma detection [
8,
10‐
12]. The idea is that a large amount of glaucoma specialist-labelled fundus images is used to train deep learning system (DLS) so that the algorithms can establish the association of abnormality patterns of the cup-to-disk ratio and optic disc hemorrhage specifically with glaucomatous optic neuropathy [
13]. The advantages of AI automated glaucoma diagnosis are not only simple and fast, but also improved accuracy without relying on the subject judgement of experts [
13,
14]. Therefore, researchers have advocated that integrating AI into community screening can overcome resource and capability deficiencies of primary care centers by providing diagnostic support to ophthalmologists [
15].
However, the application and performance of AI automated diagnosis system in real-world clinical settings has been rarely reported, leaving the validity and suitability of such technology in population-based screening to be determined [
16]. Moreover, when consider adoption of a new health intervention in real-world settings, policymakers not only evaluate its safety and effectiveness, but also the potential cost impact on the health system. For example, the local government of Changjiang county, Hainan province, China has decided to fund the Changjiang Medical Group of People’s Hospital of Wuhan University to design and conduct a community screening in its jurisdiction. Despite the expected increase of public health benefits from the screening program, the local authority, facing with the pressure of already skyrocketing health expenditures, was deeply concerned of the budget impact of such intervention. Whether the AI-assisted, population-based screening for glaucoma is worth financing from a budgetary perspective has not been explored to date.
This present study was aimed to address this issue by modelling the health outcomes and costs incurred by the screening program for glaucoma and comparing them with the health outcomes and costs that would have incurred if no screening had been implemented in the context of Changjiang county. Our hypothesis was that the screening program would inevitably increase health costs in the short-term as glaucoma patients of any stages would be identified and treated earlier. However, incremental costs would be offset by the long-term health benefits since early intervention would decrease risks of progression, which in turn would lead to cost-savings from less medical resource utilization. This study took solely PACG as illustrative example because: (1) PACG is the most prevalent type of glaucoma in China; (2) for simplicity; (3) most cost-effectiveness/cost-utility studies has focused primary open angle glaucoma, while PACG has rarely been investigated [
17].
Discussion
To our best knowledge, this was the first study evaluating community glaucoma screening from a budgetary standpoint. This study set out to address decision-makers’ budget concerns when adopting a glaucoma screening, by developing a Markov prediction model with PACG as illustrative example in which medical care costs under the screening were projected and compared against what they would have been without screening. Our study found that among 19,395 residents aged 65 and above in Changjiang county, the AI-assisted community screening for PACG was able to reduce patients with PACS by 163 (7.7%), PAC by 79 (8.8%), PACG by 45 (16.7%), and prevent 10 patients (33.3%) from any visual blindness. However, additional healthcare costs resulted from the community screening could not be offset by decreased disease progression. The 5-, 10-, and 15-year accumulated incremental costs of screening, compared with the no screening scenario, were estimated to be $396,362.8, $424,907.9, and $434,903.2, respectively. Combing the disease case numbers and incremental costs, our results indicated that the incremental cost per disease case of any stages prevented was $1464.3 over a 15-year horizon.
We also found that although the programme could not lead to cost-saving, there was a consistent downward trend in medical care costs with screening. This related to the fact that with screening, patients at earlier stages (PACS or PAC) would be identifier and treated in a timely manner so that disease progression would be prevented or delayed. Considering the fact that patients seeking medical treatment outside of the Changjiang county were not reflected in our model, we believed that the declining rate of cost might be more pronounced in real-world setting. This was because many patients screened positive in community would be directly referred through green channels to county hospitals for diagnosis and treatment. As a result, the number of patients seeking treatment in tertiary hospitals outside of the county which were generally associated with expensive healthcare services would drop.
The economical profile of glaucoma screening was still controversial in the academic community. Most studies focused on the cost-effectiveness of glaucoma screening. Previous studies by British researchers found that the incremental cost-effectiveness ratio (ICER) of population screening exceeded £30,000 per additional quality-adjusted life year (QALY) gained, making it less cost-effective than no screening [
20,
34,
35]. A recent study in the United States favored economically glaucoma screening based on routine physical checkup, with an ICER of $46,000 per QALY gained [
36]. In developing countries, an India study suggested that population-based glaucoma screening might be a cost-effective alternative to no screening [
17]. Similar results were also observed in China [
19]. On top of the cost-effectiveness aspect, our study added to evidence by demonstrating that an glaucoma screening with AI automated system could not reduce health care expenses induced by glaucoma disability. The 15-year accumulated incremental cost amounted to $434,903.2, which translated into $1464.3 per PACG of any stages prevented. Whether such excess costs were bearable depended on the local government’s financial resources. One possible explanation for the excess costs was due to the low rates of natural progression of PACS, PAC, and PACG, making the benefits of early diagnosis and treatment less pronounced compared with the standard of care. Therefore, it might require a longer time horizon for the cost-savings of screening to be realized. Another explanation might be the substantial upfront capital investment. As our results pointed out, capital costs for equipment represented the most significant driver for cumulative healthcare costs, indicating necessity for optimizing capital equipment purchasing.
China is currently undergoing population aging and transition from infectious diseases to non-infectious chronic diseases. The health system is facing a dilemma of limited health resources and increasing demand for high quality health services. Based on this background, there is an urgent need to create an efficient disease management model that shifts the focus from hospital care to public health services e.g. community screening for chronic diseases. Unlike clinical interventions such as drugs and medical devices that directly affect the pathogenies, the effect of public health is indirect since it works by regulating risk factors to control onset and progression of diseases. Therefore, benefits of public health are not as immediate as clinical treatments. Given this long-term nature, evaluation of public health interventions requires longer observation periods and significant investment in personnel and financial resources. This may explain why there is lack of evidence on the impact of population screening [
37]. In this case, Markov prediction models represent a good alternative, which combine mathematical algorithms with clinical trial and epidemiological evidence to project the impact of health interventions on the health system thereby supporting health decision-making. To the best of our knowledge, our study was the first attempt to utilize modelling technique to predict the budgetary impact of AI-assisted glaucoma screening on healthcare costs. Nevertheless, there were some inevitable limitations in this study. Firstly, as a modelling study, the predictive accuracy relied on parameters used in the Markov model. There were potential risks that parameters obtained from existing studies conducted in other regions could not be extrapolated to Changjiang county. To test the robustness of our model, we employed one-way sensitivity analysis to quantify the impact of all parameters on prediction results. Secondly, there was a lack of concrete evidence on the efficacy of treatment (e.g. LPI) for PACG due to absence of high quality randomized clinical trials. We recognized that this might be a challenge and thus we strived to base our model on the best available evidence. Additionally, the fact that all evidence sources for the treatment effectiveness included in this study have also been cited in other cost-effectiveness analysis has augmented their validity [
19]. Thirdly, we did not consider the impact of newly emerging health technologies and drugs. Fourthly, although various image-based AI automated diagnosis systems for ophthalmic disease were greatly advanced in recent years, their diagnostic performance have not been examined with a rigorous experimental study design.
Open AccessThis article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit
http://creativecommons.org/licenses/by/4.0/. The Creative Commons Public Domain Dedication waiver (
http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated in a credit line to the data.
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.