Cost-effectiveness of artificial intelligence-assisted liquid-based cytology testing for cervical cancer screening in China

Summary

Background

The 2021 World Health Organization (WHO) guidelines for cervical cancer screening recommend human papillomavirus (HPV) DNA or mRNA testing. Artificial intelligence (AI)-assisted liquid-based cytology (LBC) systems also have the potential to facilitate rapid scale-up of cervical cancer screening. We aimed to evaluate the cost-effectiveness of AI-assisted LBC testing, compared with the manual LBC and HPV-DNA testing, for primary cervical cancer screening in China.

Methods

We developed a Markov model for a cohort of 100,000 women aged 30 years over a lifetime to simulate the natural history of cervical cancer progression. We evaluated the incremental cost-effectiveness ratios (ICER) of 18 screening strategies (a combination of the three screening methods with six screening frequencies) from a healthcare provider's perspective. The willingness-to-pay threshold (US$30,828) was chosen as three times the Chinese per-capita gross domestic product in 2019. Univariate and probabilistic sensitivity analyses were performed to examine the robustness of the results.

Findings

Compared with no screening, all 18 screening strategies were cost-effective, with an ICER of $622–24,482 per quality-adjusted life-year (QALY) gained. If HPV testing after scaling up to population level screening costs $10.80 or more, screening once every 5 years using AI-assisted LBC would be the most cost-effective strategy with an ICER of $8790/QALY gained compared with the lower-cost non-dominated strategy on the cost-effectiveness frontier. Its probability of being cost-effective was higher (55.4%) than other strategies. Sensitivity analyses showed that the most cost-effective strategy would become AI-assisted LBC testing once every 3 years if the sensitivity (74.1%) and specificity (95.6%) of this method were both reduced by ≥10%. The most cost-effective strategy would become HPV-DNA testing once every 5 years if the cost of AI-assisted LBC was more expensive than manual LBC or if the HPV-DNA test cost is slightly reduced (from $10.8 to <$9.4).

Interpretation

AI-assisted LBC screening once every 5 years could be more cost-effective than manually-read LBC. Using AI-assisted LBC could have comparable cost-effectiveness to HPV DNA screening, but the relative pricing of HPV DNA testing is critical in this result.

Funding

National Natural Science Foundation of China, National Key R&D Program of China

Research in context.

Evidence before this study

2021 WHO guidelines for cervical cancer screening recommend human papillomavirus (HPV) DNA or mRNA testing. Artificial intelligence (AI)-assisted liquid-based cytology (LBC) systems also have the potential to facilitate rapid scale-up of cervical cancer screening. We searched PubMed, Embase, and Web of Science between January 1, 2000, and July 30, 2022, with no language restrictions, using the terms “China” or “Chinese”, “cervical cancer”, “artificial intelligence” or “AI” or “deep learning” or “automated” or “computer-assisted” or “image-read”, and “cost-effectiveness”, to identify published economic evaluations on artificial intelligence-assisted (AI-assisted) models for primary cervical cancer screening in China. We found no previous studies describing the cost-effectiveness of AI-assisted screening in China. We also searched for studies on the cost-effectiveness of AI-assisted screening in other countries using the same search terms, without “China” or” Chinese”. Only two previous studies evaluated the cost-effectiveness of AI-assisted screening based on conventional neural network algorithms in the UK and Australia.

Added value of this study

We evaluated the cost-effectiveness of 18 screening strategies, i.e., a combination of the three screening methods (AI-assisted liquid-based cytology (LBC) testing, manual LBC and HPV-DNA testing) with six screening frequencies (once per lifetime, twice per lifetime, once every 10 years, once every 5 years, once every 3 years and once every year). We found that screening once every 5 years using AI-assisted LBC would be the most cost-effective strategy with an ICER of $8790/QALY gained. If the sensitivity (74.1%) and specificity (95.6%) of AI-assisted LBC testing were both reduced by ≥10%, the most cost-effective strategy would become AI-assisted LBC testing once every 3 years. The most cost-effective strategy would become HPV-DNA testing once every 5 years if the cost of AI-assisted LBC was more expensive than manual LBC or if the HPV-DNA test cost is slightly reduced (from $10.8 to <$9.4).

Implications of all the available evidence

Our findings suggest that AI-assisted LBC screening once every 5 years could be more cost-effective than manually-read LBC. Using AI-assisted LBC could have comparable cost-effectiveness to HPV DNA screening, but the relative pricing of HPV DNA testing is critical in this result.

Introduction

Cervical cancer is preventable if the precursor lesions are detected and treated early or if human papillomavirus (HPV) vaccines are widely used. However, it remains the fourth leading cause of cancer-related death in women worldwide.¹ In 2020, an estimated 604,127 cases of cervical cancer and 341,831 related deaths occurred worldwide.¹ On 17th November 2020, the World Health Organization (WHO) launched a strategy for the global elimination of cervical cancer through national cervical cancer control plans involving vaccination, screening, and treatment.² The declaration set three ambitious targets by 2030: 90% of girls receive the human papillomavirus (HPV) vaccine by age 15, 70% of women receive cervical cancer screening by the age of 35 and again by the age of 45, and 90% of diagnosed women with pre-cancer or cancer diagnoses receive treatment. On 6th July 2021, WHO updated the guidelines for cervical screening that now recommend HPV DNA testing for cervical cancer screening.³

Cervical cancer has substantially increased in China, but screening coverage remains low. In contrast to the declining trends in incidence and mortality in many developed countries, China has witnessed a substantial increase in cervical cancer incidence (9.1/100,000–16.3/100,000) and mortality rate (2.4/100,000–5.1/100,000) during 2005–2015.4, 5, 6, 7, 8 China has launched a free cervical cancer screening programme for women aged 35–64 years living in rural areas since 2009,^9,10 and the cumulative number of women who received cervical cancer screening reached 120 million by 2019.¹¹ Despite this, only one-fifth of eligible Chinese women received cervical cancer screening over a 3-year period,¹² which was far less than the 70% target. One underlying cause that limits the expansion of the cervical cancer screening program, especially with cytology, is the shortage of trained pathologists, especially in less developed regions of China. Notably, China currently has only 9000 licensed pathologists nationwide and requires at least a further 90,000 to meet its routine medical needs entirely.¹³ The pathologist shortage, together with participation and accessibility issues related to women, limited the scale-up of cervical cancer screening in China. Training a sufficient number of pathologists requires significant time and monetary investment. Therefore, it is a top priority to develop novel strategies to improve the cervical cancer screening capacity to meet the current demand and WHO's goals.

Artificial intelligence (AI)-assisted screening may provide a practical alternative to overcome the shortage of pathologists and improve cervical cancer screening efficiency in China. Previous automated screening systems with liquid-based cytology (LBC) slides are mainly based on the ThinPrep Imaging System and the Becton Dickinson Focal Point GS Imaging System using conventional neural network algorithms. The sensitivity of these systems for detecting cervical intraepithelial neoplasia or invasive cancer (CIN2+) varies from 8% lower in the UK^14,15 to 5.5% higher in Australia, the United States, and Scotland,16, 17, 18, 19, 20 compared to manual LBC testing. These two systems are also used in hospital settings and small-scale screening programs in China but are not currently suitable for large-scale screening due to their high cost.^21,22 With the fast development of AI technology using deep learning algorithms, the AI cytology system (Landing CytoScanner) developed by the Landing Medical Laboratory (Wuhan, China) demonstrated that AI-assisted LBC testing outperformed manual LBC testing by a 5.8% higher screening sensitivity for detecting CIN2+ with affordable cost for screening.23, 24, 25 Kitchener et al.¹⁴ evaluated the cost-effectiveness of AI-assisted LBC versus manual LBC testing as the primary screening while using HPV-DNA testing as the triage to select women with low-grade cytology for colposcopy referral in the UK. They reported that AI-assisted LBC was more cost-effective than manual LBC testing. An evaluation of a range of screening methodologies, including AutoLBC and primary HPV screening found that primary HPV screening would save costs and save more years of life compared to other methods.²⁰ However, comparing the cost-effectiveness of AI-assisted LBC testing based on a newer deep-learning approach with HPV-DNA testing for women has not been conducted.

We developed a Markov model to evaluate the cost-effectiveness of AI-assisted LBC testing developed based on a domestic AI-assisted cytology system.^23,24 AI-assisted LBC testing was then compared to manual LBC and HPV-DNA testing, which is currently the primary screening strategy for China's national cervical cancer screening programme. We calculated the incremental cost-effectiveness ratios (ICER) of 18 screening strategies to identify the most cost-effective strategy. We also evaluated the impact of variation in costs, sensitivity, and specificity of AI-assisted LBC testing on cost-effectiveness.

Methods

Study design

From a healthcare provider perspective, we conducted a model-based economic evaluation to assess the cost-effectiveness of AI-assisted LBC testing based on the domestically manufactured Landing CytoScanner Medical System (Landing CytoBrain, Wuhan, China)^23,24 compared with manual LBC and HPV DNA testing in China's cervical cancer screening programme. The model was constructed using TreeAge Pro 2021, and the analysis was reported according to the Consolidated Health Economic Evaluation Reporting Standards 2022 (CHEERS 2022) statement²⁶ and the HPV-FRAME consensus statement.²⁷

Modelling

Mathematical models had been widely used to evaluate prevention and intervention programs for sexually transmitted infections.28, 29, 30, 31 We used a Markov model (Appendix Fig. S1a) that was extended from our published model.³² It simulated the disease progression of high-risk HPV infection to cervical cancer or regression to the susceptible state in an initial cohort of 100,000 unvaccinated and unscreened women aged 30 years (based on WHO's updated cervical cancer screening guidelines³) for a lifetime (average life expectancy of 80 years). The model was calibrated by the type-specific and age-specific incidence of high-risk HPV infection and validated by the morbidity and mortality of cervical cancer, together with the 5-year and 10-year cumulative risk of CIN2+ for HPV infection among Chinese women in our previous model.³² Since the cost of AI-assisted LBC is dependent on the number of screened individuals (Table 1), we performed the simulation for 100,000 women in our primary cost-effectiveness analysis and conducted a sensitivity analysis with a smaller population size of 50,000, 20,000 and 10,000 women, respectively (Appendix).

Table 1.

Parameters used in the model.

Parameter	Base-case	Range	Distribution	Reference
Manual LBC performancea				33,34
Sensitivity	0.70	0.60–0.79	Beta (14.92, 6.40)
Specificity	0.96	0.95–0.97	Beta (353.37, 14.72)
AI-assisted LBC performancea				23,24
Sensitivity	0.741	0.635–0.836	Beta (12.77, 4.46)
Specificity	0.956	0.946–0.966	Beta (385.56, 17.74)
HPV DNA testing performancea				33
Sensitivity	0.93	0.89–0.95	Beta (62.07, 4.67)
Specificity	0.89	0.87–0.91	Beta (208.48, 25.77)
Colposcopy performancea				32,35
Sensitivity	0.95	0.86–0.98	Beta (47.24, 2.49)
Specificity	0.42	0.26–0.61	Beta (12.41, 17.14)
Annual self-initiated examination				32,36
CIN2/3	0.01	0.005–0.02	Triangular (0.005, 0.01, 0.02)
Local cancer	0.1899	0.15–0.2	Triangular (0.15, 0.1899, 0.2)
Reginal cancer	0.5999	0.4–0.65	Triangular (0.4, 0.5999, 0.65)
Distant cancer	0.9	0.85–0.95	Triangular (0.85, 0.9, 0.95)
Efficacy of LEEP	0.92	0.85–0.99	Beta (50.49, 4.45)	32,37
Proportion of hysterectomy in CIN3	0.2	0.1–0.5	Triangular (0.1, 0.2, 0.5)	32,38
Compliance for re-visit	0.85	0.5–1	Triangular (0.5, 0.85, 1)	32,39
Costs (US$)b
Manual LBC	9.5	6.7–15.6	Triangular (6.7, 9.5, 15.6)	Survey (Appendix Table S1)
HPV DNA testing	10.8	5.4–16.2	Triangular (5.4, 10.8, 16.2)	Survey (Appendix Table S2)
AI-assisted LBC (100,000 cases)	5.1	3.8–6.3	Triangular (3.8, 5.1, 6.3)	Survey (Appendix Table S3)
AI-assisted LBC (50,000 cases)	6.0	4.5–7.5	Triangular (4.5, 6.0, 7.5)	Survey (Appendix Table S3)
AI-assisted LBC (20,000 cases)	7.9	6.0–9.9	Triangular (6.0, 7.9, 9.9)	Survey (Appendix Table S3)
AI-assisted LBC (10,000 cases)	10.7	8.0–13.4	Triangular (8.0, 10.7, 13.4)	Survey (Appendix Table S3)
Colposcopy	5.9	1.7–10.1	Triangular (1.7, 5.9, 10.1)	32
Biopsy	16.4	11–21.8	Triangular (11, 16.4, 21.8)	32
LEEP	123.5	92.4–181.6	Triangular (92.4, 123.5, 181.6)	32
Hysterectomy	315.6	151.3–479.8	Triangular (151.3, 315.6, 479.8)	32
Per-capita screening administration	2.8	±50%	Triangular (1.4, 2.8, 4,2)	32,40
Cervical cancer treatment				32,41
Local cancer	4299	2047–6550	Gamma (14.01, 4.90)
Regional cancer	5420	1766–9073	Gamma (8.45, 2.35)
Distant cancer	7176	4670–9682	Gamma (31.50, 6.61)
Annual health care				32,42
Local cancer	633	316–949	Triangular (316, 633, 949)
Reginal cancer	798	399–1197	Triangular (399, 798, 1197)
Distant cancer	1056	528–1585	Triangular (528, 1,056, 1585)
Utility score				32,43
LEEPc	0.98	0.9–1.0	Beta (183.64, 3.75)
Post-hysterectomy	0.85	0.82–0.88	Triangular (0.82, 0.85, 0.88)
Local cancer	0.83	0.79–0.87	Beta (280.36, 57.42)
Regional cancer	0.72	0.65–0.78	Beta (131.26, 51.05)
Distant cancer	0.6	0.43–0.77	Beta (18.54, 12.36)
Cured cancer	0.87	0.70–0.99	Triangular (0.70, 0.87, 0.99)
Discount rate	0.03	0.0–0.08	–	32,44

Abbreviations: AI, artificial intelligence; CIN, cervical intraepithelial neoplasia; LBC, liquid-based cytology; LEEP, loop electrosurgical excision procedure; NDRC, National Development and Reform Commission.

^aSensitivity and specificity of each screening method were estimated based on the detection of CIN2+ on histology.

^bAdjusted according to Chinese consumer price indexes in the category of health, 2019 as the base year.

^cThe disutility to quality of life was considered only in the year of treatment.

We defined no screening as the status quo and investigated 18 screening strategies that included a combination of three primary screening methods (AI-assisted LBC, manual LBC and HPV DNA testing) at six screening frequencies^3,32 (WHO recommended five frequencies, i.e., once per lifetime, twice per lifetime, once every 10 years, once every 5 years, once every 3 years, plus annual screening we proposed in this paper). Women were aged 30−64 years were targeted for cervical cancer screening.^3,5 We assumed screening once per lifetime was conducted at the age of 35 years,³² and screening twice per lifetime was conducted at 35 and 45 years of age.⁴⁵ As the primary screening test, cytology with the threshold of atypical squamous cells of undetermined significance (ASCUS+) was followed by immediate colposcopy for women with abnormal cytology.³ HPV DNA testing as the primary screening test was followed by HPV partial genotyping and cytology triage, and then immediate colposcopy was used for those who were positive for HPV-16/18 or non-HPV-16/18 genotypes combined with abnormal cytology (ASCUS+).⁴⁶ After the colposcopy, the colposcopy-directed biopsy and histopathological examination were used to confirm the diagnosis of the lesions. Women who were positive for non-HPV-16/18 genotypes with normal cytology were recommended to undergo repeat testing after 12 months based on expert consensus.⁴⁶ However, we assumed the infection in these women progressed naturally with no 12-month follow-up in the model, as the national screening programme did not cover the costs of follow-up, and the follow-up rate is extremely low. Women who were diagnosed with CIN2 or worse were referred to treatment immediately. The detailed screening, diagnosis, and treatment procedures were shown in Appendix Fig. S1b, and the diagnosis and treatment compliance and effectiveness of treatment were included in the model (Table 1). We assumed all women in the cohort would receive screening once they reached the above age interval.

Data analysis

Our methodology for estimating the type- and age-specific incidence rates of high-risk HPV infection and epidemic parameters (the probabilities of annual transition, the cancer mortality rates, the background age-specific mortality rates) followed our previous study³² (Table 1). Based on a Cochrane meta-analysis³³ and a large-scale population-based study of a cervical cancer screening programme in China,³⁴ the sensitivity and specificity of manual LBC were 0.70 (0.60–0.79) and 0.96 (0.95–0.97), and the sensitivity and specificity of HPV DNA testing are 0.93 (0.89–0.95) and 0.89 (0.87–0.91), respectively. There is no direct comparison between HPV DNA and AI-assisted LBC testing in a primary study, but there is a primary study based on a cervical cancer screening programme in 700,000 women in Hubei Province, China,²³ which showed that the relative sensitivity and relative specificity of AI-assisted LBC testing were 1.058 and 0.996, respectively, compared to manual LBC testing. We multiplied the relative sensitivity and specificity factors with the sensitivity and specificity values of manual LBC to obtain the corresponding sensitivity and specificity of AI-assisted LBC testing. These values were estimated as 0.741 (0.635–0.836) and 0.956 (0.946–0.966) in our base-case analysis.

The screening programme costs consisted of the costs of screening, diagnosis, treatment, and administration (Table 1). Manual LBC and HPV DNA test screening costs were collected from surveyed hospitals (Appendix Tables S1 and S2). The costs of screening by AI-assisted LBC was obtained from the Landing Medical Laboratory (Wuhan, China^23,24) for varied screening population size (10,000, 20,000, 50,000 and 100,000 persons, see Appendix for details) based on the cost of equipment, maintenance and staff time associated with the AI technologies (Appendix Table S3). The costs of cervical cancer treatment consisted of the costs of initial hospitalisation and subsequent annual healthcare costs (Table 1).^32,41,42 All costs were converted from Chinese yuan (CNY) to US dollars ($1 = 6.8968 CNY, in 2019).

We used published utility scores for HPV-related states from the quality-of-life assessments in Chinese patients with cervical lesions^32,43 (Table 1). We discounted cost and quality-adjusted life-year (QALY) at 3% (0−8%) annually.47, 48, 49 We calculated the incremental costs and incremental QALYs gained for each screening strategy compared with no screening (Appendix Table S4) and the next best strategy (lower-cost non-dominated strategy, Table 2). The incremental cost-effectiveness ratio (ICER) was defined as the incremental cost/QALY gained. We identified the cost-effectiveness frontier to obtain the most cost-effective strategy (Fig. 1). Using WHO standards,²⁶ we denoted strategies with an ICER <1, 1−3, and >3-times the per capita gross domestic product (GDP; $10,276 for China in 2019) as very cost-effective, cost-effective and not cost-effective, respectively.

Table 2.

Benefits and costs of screening strategies over the lifetime.

Strategy	Total costsa (US$)	QALYsa	Incremental costs (US$)	Incremental QALYs	ICER (US$/QALY)
No screening	2,275,412	2568620.31	–	–	–
Manual LBC (once per lifetime)	2,929,628	2569014.45	–	–	Dominatedb
HPV DNA testing (once per lifetime)	3,005,208	2569082.45	–	–	Dominated
AI-assisted LBC (once per lifetime)	2,534,890	2569037.53	259,479	417.22	622
Manual LBC (twice once per lifetime)	3,441,044	2569264.54	–	–	Dominated
HPV DNA testing (twice per lifetime)	3,632,359	2569333.42	–	–	Dominated
AI-assisted LBC (twice per lifetime)	2,766,351	2569293.76	–	–	Dominated
Manual LBC (once every 10 years)	3,612,554	2569484.53	–	–	Dominated
HPV DNA testing (once every 10 years)	3,910,523	2569546.87	–	–	Dominated
AI-assisted LBC (once every 10 years)	2,876,080	2569514.20	341,190	476.66	716
Manual LBC (once every 5 years)	5,878,463	2569660.00	–	–	Dominated
HPV DNA testing (once every 5 years)	6,567,703	2569691.20	–	–	Dominated
AI-assisted LBC (once every 5 years)	4,313,429	2569677.73	1,437,349	163.53	8790
Manual LBC (once every 3 years)	9,497,686	2569743.07	–	–	Dominated
AI-assisted LBC (once every 3 years)	6,663,760	2569750.51	2,350,331	72.79	32,290
HPV DNA testing (once every 3 years)	10,735,231	2569759.44	4,071,472	8.93	456,157
Manual LBC (once every year)	26,818,228	2569767.48	–	–	Dominated
AI-assisted LBC (once every year)	17,863,654	2569767.32	7,128,422	7.88	905,044
HPV DNA testing (once every year)	30,569,375	2569776.04	12,705,721	8.72	1,456,585

Abbreviations: AI, artificial intelligence; LBC, liquid-based cytology; QALYs, quality-adjusted-life-years.

^aQALYs and costs are expressed as the present value in 2019 (3% annual discount rate).

^bThe strategy that is dominated yields fewer QALYs at higher costs than the comparator.

Fig. 1.

Cost-effectiveness frontier for all 18 screening strategies (100,000 cohort members). Incremental QALYs and incremental costs of intervention strategies are obtained by comparing with the no screening scenario. Strategies on the cost-effectiveness frontier and their incremental cost-effectiveness ratios (ICER) compared with the lower-cost non-dominated strategy are shown. The strategies on the upper left of the frontier are dominated by the strategies on the lower right. Incremental QALYs and incremental costs are expressed as the value on Dec 31, 2019. Abbreviations: AI, Artificial intelligence; GDP, gross domestic product; LBC, liquid-based cytology; QALY, quality-adjusted-life-year.

Sensitivity analysis

We conducted a probabilistic sensitivity analysis based on 10,000 simulations to determine the probability of being cost-effective for each screening strategy (cost-effectiveness acceptability curves, Fig. 2). The distribution of all model parameters was based on the properties of the parameters and data informing the parameters.

Fig. 2.

Cost-effectiveness acceptability curves for all strategies. Within the willingness-to-pay threshold of three times the Chinese per-capita GDP, screening strategies with the probability of being cost-effective consistently being zero are not shown. Abbreviations: AI, Artificial intelligence; GDP, gross domestic product; QALY, quality-adjusted-life-year.

We examined the impact on ICER of potential reduction of sensitivity and specificity of AI-assisted LBC testing (5 scenarios in Appendix Table S5). Despite a reported high sensitivity of AI-assisted LBC testing in our study, previous studies^14,15 reported a lower sensitivity (8% lower) of automated image reading than manual image reading. To account for this potential reduction in sensitivity, we repeated our simulations by reducing the sensitivity and specificity of AI-assisted LBC testing by 10% (scenario 1, Appendix Fig. S2) and 20% (scenario 2, Appendix Fig. S3) of the current baseline level. In these scenarios, sensitivity and specificity for manual LBC testing and HPV DNA testing remained unchanged. However, unsatisfactory sampling, quality of slide reading by cytotechnicians, and the sensitivity of HPV reagents in the large-scale screening programme may affect the performance of these screening methods. We hence created three additional scenarios as follows: (1) the sensitivity and specificity of all three screening methods were reduced by 10% (scenario 3, Appendix Fig. S4); (2) the sensitivity and specificity of AI-assisted LBC testing and manual LBC testing were reduced by 20%, whereas HPV DNA testing was reduced by 10% (scenario 4, Appendix Fig. S5); and (3) the sensitivity and specificity of all three screening strategies were reduced by 20% (scenario 5, Appendix Fig. S6).

We examined the potential impact on ICERs in a smaller population (50,000, 20,000 and 10,000 persons) as a smaller population size may result in a higher price for AI-assisted LBC testing. We presented the cost-effectiveness frontier (Appendix Fig. S7) and the cost-effectiveness acceptability curves (Appendix Fig. S8).

We performed univariate and two-way sensitivity analysis (Appendix Figs. S9 and S10) to examine the impact of model parameters within their respective ranges on the ICER when comparing three primary screening methods with the same frequency as in the most cost-effective strategy and identify the most sensitive parameters.

Role of the funding source

The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.

Results

Cost-effectiveness of AI-assisted LBC screening

Compared with the ‘no screening’ scenario, all 18 screening strategies were cost-effective, with an ICER of $622–$24,482/QALY gained (Appendix Table S4).

The cost-effectiveness frontier in Fig. 1 showed that the most cost-effective strategy is AI-assisted LBC testing once every 5 years with an ICER of $8790/QALY gained (Table 2), according to a 3-time GDP willingness-to-pay (WTP) threshold ($30,828). As WTP decreases, the optimal strategy becomes AI-assisted LBC testing once every 10 years (ICER = $716/QALY gained), then once per lifetime (ICER = $622/QALY gained). However, the next four strategies on the cost-effectiveness frontier, i.e., AI-assisted LBC testing once every 3 years (ICER = $32,290/QALY gained), HPV DNA testing once every 3 years (ICER = $456,157/QALY gained), AI-assisted LBC testing once every year (ICER = $904,044/QALY gained), and HPV DNA testing once every year (ICER = $1,456,585/QALY gained), were not cost-effective compared with the lower-cost non-dominated strategy.

The probabilistic sensitivity analysis in Fig. 2 showed the cost-effectiveness acceptability curves for all strategies using a range of WTP between 0 and 3 times the per-capita GDP. At a WTP threshold of three times the per-capita GDP, AI-assisted LBC testing once every 5 years showed a 55.4% probability of being cost-effective and outperforming other strategies. For a WTP between $863 and $10,730, AI-assisted LBC testing once every 10 years had the highest probability of being cost-effective. The ‘No screening’ strategy was cost-effective when WTP was reduced below $863.

Impact of varying screening sensitivity and specificity

Varying the sensitivity and specificity of the screening methods resulted in different strategies being the most cost-effective. For all scenarios with reduced sensitivity and specificity in three screening methods (scenarios 1–5), the most cost-effective screening strategy became AI-assisted LBC testing once every 3 years with an ICER of $20,239–25,287/QALY gained (Appendix Figs. S2–S6).

Impact of varying target population size

Varying population size may result in different strategies being the most cost-effective due to the change in the cost of AI-assisted LBC testing (Appendix Figs. S7 and S8). When the target population size for screening was reduced to 50,000 or 20,000, AI-assisted LBC testing once every 5 years remained the most cost-effective strategy, with an ICER of $9873 or $12,182/QALY gained, respectively. However, when the population size for screening was reduced to 10,000, the most cost-effective strategy became HPV DNA testing once every 5 years with an ICER of $22,094/QALY gained.

Pairwise comparison between AI-assisted LBC testing and other screening methods

Varying model parameters (Appendix Figs. S9 and S10) did not vary the conclusion that AI-assisted LBC testing once every 5 years was more cost-effective than manual LBC testing once every 5 years. However, if the cost of HPV DNA testing was reduced by 13% from $10.8 to $9.4, HPV DNA testing once every 5 years was more cost-effective than AI-assisted LBC testing once every 5 years, irrespective of the cost of AI-assisted LBC testing in its range.

Discussion

This study evaluated the cost-effectiveness of 18 screening strategies based on three screening methods, each with six screening frequencies. We found that all 18 screening strategies were cost-effective compared to the status quo (no screening). Given the base case assumptions on cost (i.e., that HPV screening, even at high screening volumes, would be $10.80), AI-assisted LBC testing once every 5 years was the most cost-effective strategy for cervical cancer screening in China, but this finding was contingent on many factors, especially primary screening test price. If the sensitivity and specificity of the three screening methods were decreased by 10–20%, AI-assisted LBC testing once every 3 years may become a more cost-effective option. If the HPV test cost was slightly reduced to $9.4, the most cost-effective strategy became HPV DNA testing once every 5 years. To the best of our knowledge, this is the first study to evaluate the cost-effectiveness of AI-assisted cervical cancer screening in a Chinese setting.

Our result of good cost-effectiveness of AI-assisted LBC testing, compared with manual LBC testing, should be interpreted with caution. The cost-effectiveness of the AI strategy is largely determined by its sensitivity and cost. Our baseline results demonstrated that AI-assisted LBC testing, with higher sensitivity and lower cost than manual LBC, would be more cost-effective than manual LBC testing. Our sensitivity analysis (Appendix Figs. S2 and S3) by reducing the testing sensitivity and cost of AI-assisted LBC testing demonstrated consistent findings with a similar study in the UK¹⁴ where AI-assisted LBC testing has lower sensitivity and lower cost than manual LBC. Conversely, our sensitivity analysis (Appendix Fig. S10b) by increasing the testing sensitivity and cost of AI-assisted LBC testing resulted in better cost-effectiveness of manual LBC testing than AI-assisted LBC testing, consistent with a previous study in Australia.²⁰

Our research presents a potentially more cost-effective strategy for using AI-assisted LBC testing to scale-up cervical cancer screening in China. In a Chinese setting, AI-assisted LBC testing can substantially improve cervical screening accuracy and efficiency without substantial additional investment in medical personnel training and qualification nor the delay associated with the personnel training. It is also potentially beneficial for remote and rural areas of China, where qualified pathologists are scarce. With advances in technology and accumulating clinical data for HPV and cervical cancer diagnoses, an AI-assisted diagnostic system could become increasingly accurate, efficient, and economical for population-wide screening. AI-assisted LBC testing can substantially increase China's cervical cancer screening rate in its effort to achieve the ambitious WHO targets of cervical cancer elimination. Further, a feasible AI-assisted LBC testing strategy may also be transferrable to other developing country settings to aid their screening scale-up and the achievement of cervical cancer elimination targets worldwide.

Our finding demonstrates that the sensitivity and specificity of AI-assisted LBC testing can significantly influence its cost-effectiveness. This may have practical implications. The high sensitivity and specificity of AI-assisted testing are associated with the quality of the slide production, the data quality and the efficiency of the deep learning algorithms. Particularly, it hinges on sophisticated sample collection by the trained medical staff because a poor sample collection may lead to inadequate slides and low performance of the AI-assisted diagnostic system. As a result, AI-assisted LBC testing still requires an initial establishment of the diagnostic system and sufficient training of clinical and technological personnel in primary care. Adequate training for medical staff in sampling ensures the subsequent performance of the AI-assisted approach.50, 51, 52 Nevertheless, the required skill set and training turnover time for primary medical staff would be much less than specialised training for pathologists. In this vein, however, it is important to note that HPV testing also offers the potential for highly automated, quality-controlled testing.

Our finding of better cost-effectiveness of AI-assisted LBC than HPV DNA testing once every 5 years is an important indicative result for China but may not be a replacement for HPV DNA testing. The worldwide push for cervical cancer elimination driven by WHO may further reduce the cost of HPV DNA testing. Our sensitivity analysis indicates that a 13% reduction in the current cost of HPV DNA testing would render it more cost-effective than AI-assisted LBC testing, irrespective of the cost of AI-assisted LBC testing in its range in China. Also, to date, direct comparison data for this version of AI-assisted LBC versus HPV DNA screening is not available from other studies and settings, and more primary data on the longitudinal performance of AI will be required over the long term before this technology could be generalised to broader scenarios outside China. For under-developed regions with limited primary care access to perform any form of cytology (including AI-assisted LBC), self-collection of HPV samples and posting the sample for HPV DNA testing to a qualified laboratory may be one available option with ease and privacy.⁵³ However, for settings with an established cytology-based screening system, AI-assisted LBC testing could provide an additional and feasible option for rapid cervical screening scale-up in the population.

Our study has several limitations. First, the sensitivity, specificity and costs associated with AI-assisted LBC testing were collected from socioeconomically deprived areas in a Chinese province, which may vary in other settings. The primary study directly compared AI-assisted and manual LBC testing in a Chinese setting²³ but did not directly compare AI-assisted LBC with HPV DNA primary screening. More evidence about the cost-effectiveness of the AI system would be needed for the population in other provinces in China or other countries in the future. Second, the actual sensitivity and specificity would likely be lower than what is measured in a controlled environment, as in our study. We are yet to quantify how much the sensitivity and specificity of AI-assisted LBC testing may reduce in clinical practices and how this may affect its cost-effectiveness. Third, the management strategy on the threshold for referral and management of triage negatives may affect the cost-effectiveness. We assumed that women who were positive for non-HPV-16/18 genotypes but with normal cytology would not be followed up at the 12th month based on the national screening programme in China. This, which is different to 2021 WHO recommendations, is likely to underestimate both the costs and benefits of HPV DNA testing. Fourth, we assumed the current coverage of the primary care network in China is sufficient to provide cytology to the targeted women, which is essential for implementing AI-assisted LBC testing. Fifth, the use of AI technology is common in China. Whether the study finding is generalisable to other LMIC settings requires further investigations due to the lack of primary data on the longitudinal performance of AI in the longer term. As HPV-based screening algorithms may vary across LMIC countries,^3,46 country-specific cost-effectiveness analysis needs to be conducted to accommodate the differences. Sixth, we consider the utility value change with the cancer stage at a wide range while without considering the decreased utility value with age directly⁵⁴ due to limited data on utility value stratified by both age and cancer stage. Our sensitivity analysis results show that this would not change the conclusion of cost-effectiveness. Finally, while our results were robust to parameter uncertainty, future screening costs or efficacy changes, and the potential impact of HPV vaccination on the disease-related parameters may affect our estimates.

In conclusion, our study indicates that screening once every 5 years using AI-assisted LBC has comparable cost-effectiveness for cervical cancer screening to HPV DNA screening, and we note that the relative pricing of HPV DNA testing is critical in this result. These findings are of importance for China as the evidence for the technical performance in long-term studies continues to build.

Contributors

M.S., Z.Z., H.B., L.W., and L.Z. conceived and designed the study. M.S., Z.Z. and H.B. collected the data. M.S. and Z.Z. analysed the data, carried out the analysis and performed numerical simulations. M.S. wrote the first draft of the manuscript. L.Z., L.W., and K.C. critically revised the manuscript. All the authors contributed to writing the paper and agreed with the manuscript results and conclusions.

Data sharing statement

The data used in this study are referenced in the article and included in the supplementary materials files.

Ethics approval and consent to participate

Not appliable.

Declaration of interests

All authors declare that they have no competing interests. CKF holds shares in CSL Behring.