The role of artificial intelligence in global health

Deemed to drive the upcoming “fourth industrial revolution”, artificial intelligence has ramped up its reach dramatically in the past decade due to advances in computing infrastructure, overall scientific knowledge of AI and related algorithms, and the amassment of large data sets across many industries.

Medicine is no exception, and numerous AI applications have been created and even begun to be deployed successfully in clinical practice. Automated interpretation of medical imaging (e.g. in radiology, histopathology, ophthalmology) is an area that shows significant promise for instance, with some AI algorithms performing on par and in some cases even better than entire panels of trained medical professionals. Other areas of development are clinical decision support, prediction of clinical outcomes, or even medical education in simulated VR environments.

AI has potential to automate many medical processes, and thus reduce variations in medical care, which is a well known source of higher cost and poorer health outcomes. In some instances, AI applications could even raise the quality threshold directly, through simply better performance compared to their human counterparts.

AI can also help allocate limited resources more wisely, free up funds through automation, and lower medical costs for clinical facilities and patients alike.

On the other hand, global health brings significant challenges, in addition to just more limited resources and poorer access to care. Patient education and their desire to seek preventive care as well as acute care, when needed, are oftentimes in need of improvement, even more so than in developed countries. Beyond just overall reduced material resources, the quality of care is also generally significantly lower in less developed countries. This translates oftentimes directly into increased patient mortality.

For example, one study has shown more than a 5 fold higher number of years of life lost to poor quality (per 1000 population) in Eastern Europe compared to Central Europe, solely because of the lower quality of medical care. Figure 1, displaying low and mid income countries with GDP per capita lower than 33,000$Int, shows an inverse relationship between GDP per capita and poor quality related deaths.

Figure 1. GDP data from World Bank. Quality data from Kruk et al, Lancet 2018;392(10160):2203-2212.

Based on all of the above, we, as a society, have a tremendous opportunity to employ technology and artificial intelligence to address some of the biggest challenges of global health. AI can help provide care in areas that are simply too remote or where the healthcare system is less developed.

There are however significant global disparities in the development of and access to AI applications. The distribution of AI-patents by country looks like a power-law where very few countries are making a disproportionately large impact. Only 5 countries have produced more than 86% of all AI patents since 1997. These top contributors are China, US, Japan, South Korea and Germany (Figure 2). Surprisingly, many remaining higher GDP countries are making apparently limited contributions at this time. These differences are even more pronounced in poorer countries.

Figure 2. Data from Statista.com

With overall societal good will, AI scientific knowledge can potentially be ported directly from the developed world, especially in an area such as medicine, where economic or political competition should be less of a deciding factor. The lack of large medical datasets needed to train AI models in developing countries is however more difficult to address.

The limited availability of large medical data sets in poorer countries is due to the lack of informatization of medical records and oftentimes the absence of nation-wide efforts to centralize existent data sources. This can be a significant barrier to creating AI applications that are specific to those particular patient populations. Different local characteristics of the patient populations and their living environment oftentimes require retraining of AI models to ensure proper performance.

There are some solutions to tackle this challenge.

Using data sets that may be more widely available than electronic medical records, such as medical images or billing claims, could be a good starting point of focus. Another possibility is to train AI models on data sets from countries where such large clinical data exist and only customize them further using smaller, local data sets, where available. Creating synthetic data sets is another way to circumvent some of these barriers.

Making patient data and AI applications more available creates some potential issues on its own.

While making clinical data more available may sometimes raise the challenge of ensuring patient confidentiality, this can be addressed with proper data security protocols as well as alignment of the specific national legislation. Development of ethical AI policies as well as wide dispersion of these technologies should be ensured for true democratization of patients´ access. Furthermore, in time, data from all types of patient groups should be collected in each country. Subsequent refinement of the AI models will subsequently help decrease the risk of AI bias that has been brought to light in recent years in the developed world.

Furthermore, AI should be seen as a tool that supports and augments the work of human medical professionals and not necessarily as a replacement for them. Although some clinical processes can be completely automated, this only increases efficiency and frees up time for the currently strained medical professionals. They will thus be more available to address the more complex, more procedural or more creative parts of medicine, validate AI findings, integrate clinical results, and ensure medical quality.