Evaluation of risk stratified breast cancer screening regimens that employ artificial intelligence driven short-term risk assessment

Background

There is increasing interest in the UK and internationally to replace one-size-fits-all mammography screening for breast cancer by risk-stratified screening, where the frequency and modality of screening are chosen based on risk of breast cancer. Recently an artificial intelligence (AI) model has been developed which interprets data automatically generated from mammogram screenings, providing an immediate estimation of an individual's short-term risk of cancer incidence following a negative mammogram (eg. risk over the next 3 years). Typically, classical models for assessing breast cancer risk consider a timeframe of 10 years or more, which that is less relevant time frame for decision makers than three years, given breast screening in England takes place every three year. The AI model is called Mirai, and it could be used to perform automatic and instantaneous risk assessments during screening appointments, and the Mirai software is free of charge. This suggests it has the potential for easy integration into the National Breast Screening Programme when compared to other breast cancer risk assessment approaches. These alternatives may involve procedures such as obtaining a blood sample for a polygenic risk score or completing an additional questionnaire during a screening appointment and verifying the accuracy of the questionnaire answers.

Aim The aim of this research is to assess the cost-effectiveness of using an AI based risk-stratified screening programs into the UK National Breast Cancer Screening program.

Methods We first develop and apply a framework to determine risk thresholds for implementing risk-stratified screening regimens in women aged 50 to 70 years. Each regimen comprises of screening intervals (1 year to 6 years) that correspond to specific risk groups, with risk evaluated using the AI model. A deterministic model is applied to establish risk thresholds to minimise the expected advanced cancer incidence when constraining the expected number of breast screens to equal the number in the current England breast screening programme, assuming perfect compliance. After a range of different screening strategies (and risk thresholds) are established, we use a discrete event simulation economic model to estimate population health (QALYs) and NHS costs over the lifetime of adult women eligible for breast

screening in England. The economic outcomes are established for four different risk-stratified screening strategies and are compared to the current England screening programme.

Results

All the AI-based risk stratified regimens programs reduced NHS costs and increase QALYs compared with the current screening programme. This remained the case under a range of sensitivity analysis conducted on key model parameters. The strategy of conducting screening every six years (low risk), three years (medium risk), and annually (high risk) was estimated to yield the highest additional net monetary gain per woman invited for screening. This was £365 and £514 under the assumption of QALY values being £20,000 and £30,000, respectively. Consequently, this strategy is expected to provide an annual net monetary benefit within the NHS screening program of £12 million, £63.8 million, and £89.6 million for cost per QALY gained values of £1, £20,000, and £30,000, respectively. The three alternative approaches produced comparable figures for the net monetary benefit gained.

Conclusion

We used an economic model to generate practical predictions of the population health gains within the NHS from the adoption of AI-based risk-stratified breast screening in the NHS. We find that AI-based regimens could in theory be used to improve population health while reducing NHS costs and the demand for screening services. The economic benefits of AI-based risk-stratified regimens amounted to tens of millions of pounds annually and was consistent across the four regimens we developed and evaluated. This implies that the specific choice of an AI risk-based screening regimen is less crucial for attaining population health gains than the overarching decision to implement any form of AI-based risk-stratified screening.

Given the considerable benefits indicated by the economic model we recommend conducting prospective implementation studies for AI-based risk-stratified breast screening. These studies should focus on investigating short-term implementation aspects in detail, including the feasibility of introducing the technology, associated implementation costs, and the immediate impact on population health resulting from the adoption of AI-based risk-stratified breast screening within the NHS.