Project: Cost-effective screening for breast cancer and RHD

This project focuses on two diseases that affect a large number of people in the world: breast cancer and rheumatic heart disease (RHD), using artificial intelligence (AI) models to analyse ultrasound and thermal images and to determine the cost-effectiveness of different imaging techniques. This is particularly useful for low- and middle-income countries (LMICs), where there is a shortage of medical doctors. The project involves researchers from UNED and HM Hospitales. It is funded by the Spanish Ministry of Science, Innovation and Universities, with support from the European Union.

Breast cancer

The burden of breast cancer

Breast cancer is the most common type of tumour in women. According to the World Health Organization (WHO), “in 2020, there were 2.3 million women diagnosed with breast cancer and 685,000 deaths globally”. In high-income countries 12.5% of women (one in 8) develops breast cancer in their life and more than 4% (about one in 24) die of it. Around 40,000 new cases were diagnosed in Spain in 2022, with increasing incidence: from 106.5 cases per 100,000 women in 2002 to 126.0 in 2020, i.e., 19% increase in less than two decades.

In LMICSs, cancer incidence is lower but rising much faster due to changes in life habits and longer life expectancy. In sub-Saharan Africa, it is projected to increase by more than 92% between 2020 and 2040.

Despite lower incidence, mortality is much higher because screening is not widely available and effective treatments are inaccessible to most people. In addition, in Africa and Latin America, breast cancers occur at a much younger age and are more aggressive due to ethnic factors. In 2020, 70% of cancer-related deaths occurred in LMICs, and mortality rates, unlike the case of high-income countries, are increasing alarmingly; for example, in Mexico they have doubled over the past 10 years—see the references in this document.

Breast cancer screening in LMICs

Breast cancer survival rate is 99% when the disease is detected early, but drops to 22% when it is detected in stage 4. Early detection also saves economic resources: treating breast cancer in this late stage is three times more expensive than treating it in stage 0. For this reason, all countries are interested in implementing screening programmes, aimed at detecting the disease in asymptomatic women.

Mammography is the standard screening modality in high-income countries, but it is not feasible in LMICs due to the cost of equipment and the difficulty of implanting it in rural areas. Ultrasound [Omidiji et al., 2017; Sood et al., 2019; Martei et al., 2022] and thermography [Davalagi et al., 2022] have been proposed as cost-effective alternatives, since they are portable and much less expensive. They also have the advantages of being safe (because they do not use ionising radiation) and painless (because they do not require compression of the breast).

Another problem, even more serious than the cost of the equipment, is the lack of qualified personnel to interpret the images, as the few radiologists trained each year in these countries often move to other places that offer much higher salaries and better living conditions. For this reason, AI systems for automatic analysis of ultrasound and thermography images and for decision support could save the lives of many women.

Open research questions

There is currently considerable controversy about the optimal screening pattern for breast cancer: at what age to start, when to stop, which technique or combination of techniques to use, and how often. Even within the USA, several scientific societies (USPSTF, ACS, ACOG, ACR, ACP...) have proposed discrepant guidelines. Many experts argue that the screening plan should be tailored to each woman, depending on her age, personal and familiar antecents, breast density, etc., rather than applying a one-size-fits-all programme.

Another controversial issue is whether breast thermography can play a role in breast cancer screening and diagnosis. This technique originated in the 1950s but, after some years of enthusiasm, it was abandoned by radiologists, to the point that now most scientific societies advise against its use. However, improvements in the thermal sensitivity and spatial resolution of modern cameras, which are more and more affordable, and the analysis of images with AI, which has increased the sensitivity and specificity of this technique, have led to a renewed interest among experts—see the systematic review cited below.

Aims of our project

One goal of this project is to develop AI applications for the automatic analysis of breast ultrasound and breast thermography images. For this purpose, we will mainly use deep convolutional neural networks, with special emphasis in generating visual and natural language explanations of their conclusions.

Another goal is to determine the optimal screening pattern for each woman with cost-effectiveness criteria, considering all the imaging techniques available, not only mammography, ultrasound and MRI, as in other studies published in the literature, but also breast thermography, whose cost-effectiveness has never been analysed. For this purpose we will use two new types of probabilistic graphical models (PGMs), namely Markov influence diagrams and DESnets, created by our group and implemented in OpenMarkov, an open-source tool developed at UNED.

Our publications about breast ultrasound

Our publications about breast thermography

Other links

Rheumatic heart disease (RHD)

The burden of RHD

RHD has been virtually eliminated in high-income countries (with the exception of some indigenous populations in Oceania), but it is still the leading cause of heart failure in LMICs. It affects about 40 million people, and causes 300,000 deaths and many cases of severe disability each year, most of them in children and young adults, with a higher incidence in girls and women [Rwebembera et al., 2022]. It is also a cause of maternal, foetal and neonatal mortality [Liaw et al., 2021; Prasad et al., 2023]. However, it can be considered a neglected disease: “the Big 3”—namely HIV, tuberculosis and malaria—cause 3 to 5 times more deaths, but RHD only receives around 0,01% the funding of those diseases, when it should receive at least 20% [Macleod et al., 2019; Belay and Aliju, 2021].

RHD screening

The most effective prevention consists in screening children and adolescents in schools and applying antibiotic prophylaxis with penicillin when the disease is still latent. The main barrier is the lack of cardiologists in LMICs; for example, Sierra Leone, with a population of almost 9 million people, has only one cardiologist.

For this reason, several countries have adopted a task-shifting approach in which the first phase of screening is carried out by briefly trained nurses or technicians, sometimes with the support of an AI, and in a second phase those suspected of having RHD are examined by expert cardiologists using standard or portable sonographs [Rwebembera et al., 2022].

Open research questions

As a consequence of this new approach, there is a growing interest in developing applications to assist the paramedical personnel in the first phase of RHD screening.

Another issue is to refine the criteria for diagnosing the disease in screening programmes, as it is sometimes difficult to distinguish pathological signs from trivial findings and, in the case of valvulopathy, to determine whether it is of rheumatic origin and whether it is serious enough to start secondary prophylaxis. Several authors have developed different sets of criteria, and in November 2023 the World Heart Federation (WHF) updated the guidelines it had proposed a decade earlier [Reményi et al., 2012; Rwebembera et al., 2023], but there are still some cardiologists who disagree with them.

Aims of our project

Our main goal is to develop AI applications for the detection of RHD by echocardiography using deep convolutional neural networks. In order to obtain images to train these networks, our group conducted a study in Makeni, Sierra Leone, in July 2022, in collaboration with the NGO Viva Makeni, in which a Spanish paediatric cardiologist screened 168 children. In February 2024, two cardiologists examined 604 secondary school students in Lunsar, Sierra Leone, in collaboration with the Mile 91 Foundation.

The 13 cases of suspected RHD found in this study were examined with the seven sets of criteria proposed in the literature, which has brought to light important discrepancies between them. We believe that a probabilistic causal model (more specifically, a Bayesian network) combining expert knowledge on valvulopathies with clinical data could be more accurate than consensus-based sets of criteria.

Finally, we intend to conduct a cost-effectiveness analysis to determine whether RHD is worth its cost in Sierra Leone, as well as a budget impact analysis to estimate the cost of implanting secondary prophylaxis in this country. In collaboration with researchers abroad, we will also try to carry out similar analyses for other countries, using the same types of probabilistic graphical models mentioned above for breast cancer.

Our publications about echocardiography

Other links