Original Title: Inteligencia artificial en los programa de cribado de cáncer de mama

Authors' objectives: To assess the clinical efficacy and efficiency of adopting AI systems in Breast Cancer Early Detection Programmes (BCEDPs) through a systematic review of the scientific evidence.

Authors' results and conclusions: Overall, 1 systematic review and 11 individual studies on clinical efficacy and 1 cost-analysis study were included. Nine studies analysed AI-based automated reading systems. The discriminatory power of these systems was considered to be good in five of the studies and acceptable in three. Four studies reported that the AI systems were more sensitive than a single reading by a radiologist for certain cut-off points. Eight studies analysed AI systems as tools to support reading of radiological images. The discriminatory power was considered good in one case, acceptable in five and unacceptable in two. In all eight studies, readings by radiologists were more sensitive when made with than without AI assistance. Seven studies assessed the sensitivity and specificity of AI systems as a classification tool before screening. Sensitivity was high for lower thresholds for the risk of cancer detection using mammography in five studies, while it was variable for higher risk thresholds in two studies. Nine studies analysed the impact of AI systems on specialists’ workload. Six studies (four considering AI as a support tool for the radiologist and two as a classification tool) reported a reduction in radiologists’ workload (reading time). The cost-effectiveness study included indicated that the most costeffective strategy among those analysed was the one that used AI for the initial prediction of breast cancer for the 40- to 49-year-old age group, followed by annual screening only for the women in this group classified as high risk, and then screening from 50 up to 74 years of age, in line with the United States Preventive Services Task Force (USPSTF) guidelines. On the other hand, for a population cohort of 50,000 screened women, it was calculated that the costs per mammography exam carried out using an AI-based system and by a radiologist were €0.82 and €6.39, respectively, with a 44.3 % reduction in radiologists’ workload (mammography screen reading), and overall, the incremental cost for a screening strategy using an AI system was –253,384.62. Conclusion The evidence reviewed suggests that AI systems are most accurate when used in the screening process as a tool to support single reading of mammograms and pre-screening triage. The economic analysis indicates that the following is a cost-effective strategy: obtaining an index mammography in all women aged 40 years old that is then interpreted by AI to predict breast cancer risk, plus annual screening from 40 to 49 years of age in women predicted to be at high risk of breast cancer (relative risk ≥ 1.1), followed by breast cancer screening from 50 to 74 years old in accordance with the USPSTF guidelines. For the base case scenario considered, the costs of breast cancer population screening with a strategy based on an AI system to assist detection, using it to classify exams before single or double reading by specialists, are lower than the standard double reading procedure. The costs associated with the AI system (€41,140) were compensated for by the lower costs associated with mammography reading by radiologists in the intervention group (which required 46,095 fewer readings than the control group). Finally, we should highlight that the use of AI in breast cancer screening is a challenge with ethical, legal and social as well as technical implications, and that should be considered carefully to avoid harmful effects for individuals and groups, especially the most disadvantaged. Indeed, as commercial systems have been developed based on research in specific cohorts and contexts, there is no guarantee that they are applicable to other populations.

Authors' methods: To identify studies on efficacy and cost-effectiveness, a systematic review of the scientific literature was performed using the following databases: Cochrane Library, International HTA Database, Medline (PubMed), Embase (Ovid Web), Web of Science and Scopus. In addition, economic analysis was conducted to estimate the incremental costs associated with screening with an AI support tool for detection, using this tool to assign screening exams to single or double reading compared to standard double reading. The analysis was carried out from the perspective of the funding body of the Spanish National Health System and with a short time horizon, and costs were estimated using direct healthcare costs and the workload of radiologists associated with the strategies analysed. Given potential uncertainties in the data, univariate sensitivity analysis was also performed.

Project Status: Completed

Year Published: 2024

URL for published report: https://www.euskadi.eus/contenidos/informacion/2024_osteba_publicacion/eu_def/adjuntos/E-24-09-Inteligencia-Artificial-cancer-mama.pdf

English language abstract: An English language summary is available

Publication Type: Full HTA

Country: Spain

Organisation Name: Basque Office for Health Technology Assessment

Contact Address: C/ Donostia – San Sebastián, 1 (Edificio Lakua II, 4ª planta) 01010 Vitoria - Gasteiz

Contact Name: Lorea Galnares-Cordero

Contact Email: lgalnares@bioef.eus

Copyright: <p>Osteba (Basque Office for Health Technology Assessment) Health Department of the Basque Government</p>