Types of AI

There are in general two types of AI. The ability of a machine to represent the human mind and perform any intellectual task that a human can perform is termed artificial general intelligence.6 Artificial general intelligence was the initial focus of early AI research and the predominant representation of AI in popular culture.7 However, given the challenges and complexity associated with developing this kind of AI, many researchers turned their attention to artificial narrow intelligence: the ability of a machine to perform a single task extremely well. Virtually all contemporary AI health applications are considered artificial narrow intelligence .7

No universally accepted classifications of AI subfields relevant to health exist. For that reason, we have chosen to focus on areas that have been repeatedly cited by researchers and are consistent and recurring themes at the Artificial Intelligence in Medicine conference, a prominent biannual meeting of researchers focused on AI applications related to medicine and health.8 While additional themes have also been repeatedly discussed in recent years, some of these areas are more upstream (ie, focused on research into foundational AI principles) and thus not included here.8 Many of these areas are often used together in a single application.

Complementary advances in digital health technologies in resource-poor settings

High-income countries are already benefiting from integrating AI into their healthcare ecosystems. One analysis has found that the use of AI applications could result in approximately $150 billion in saved healthcare costs annually by 2026 in the USA.13 There are many reasons  to be optimistic that AI could also prove transformative for public health in resource-poor countries. The large volume of data being generated has created a plethora of opportunities for applying AI to improve individual and population health. This is particularly true in resource-poor settings where there has been strong mobile phone penetration, developments in cloud computing and substantial investments in digitising health information and introducing mobile health (mHealth) applications. As a result, many places now have the necessary basics in place to initiate meaningful applications of AI. Advances in AI could also help expand and strengthen the impact of these and other digital health technologies in such settings.

AI applications in resource-poor settings

Little has been documented in the academic literature on AI applications for health in resource-poor settings. However, this should not be taken as an indicator of the current activities in this area. There are many examples of promising technologies covered in the media and at conferences. The ‘AI for Development’ community has published several observations of how AI has been implemented in resource-poor settings. A major lesson from the experience of those working on AI in resource-poor settings is that ‘AI should build intelligence into existing systems and institutions rather than starting from scratch or hoping to replace existing systems, however broken’.21 The success of AI applications requires knowledge of local markets, clear usability requirements and access to adequate training data via field testing.21 Examples of how AI is currently being deployed are noted below. While many of these interventions have not yet been robustly evaluated, they do provide insight into how AI subfields are being applied and their potential beyond current pilots.

In resource-poor settings, expert systems can be used to support health programmes in several ways. First, medical expert systems can support physicians in diagnosing patients and choosing treatment plans as is done in high-income countries. For some conditions, they can act in place of a human expert if one is not readily available, which is often the case in poor communities. In this way, expert systems have already been deployed in resource-poor settings. Birth asphyxia is not always predictable or preventable. Cases often require immediate, skilled resuscitation in the delivery room. Researchers from Brazil and USA developed a fuzzy logic-based expert system to predict birth asphyxia for use in developing countries. The system relates maternal medical, obstetric and neonatal characteristics to the clinical conditions of the newborn to predict the need for resuscitation. Researchers found that the application was 77% sensitive and 95% specific for identifying the need for birth asphyxia at tertiary perinatal medical centres.22

AI is already being used to predict, model and slow the spread of disease in epidemic situations around the world, including in resource-poor settings. Dengue fever is a vector-borne disease that has spread rapidly around the globe in recent years. About half of the world’s population is currently at risk.23 Researchers have developed a machine learning tool to identify weather and land-use patterns associated with dengue fever transmission in Manila. The machine learning algorithm has learnt over many iterations how to fine-tune its model to predict dengue occurrence with increasing accuracy.24 Researchers are now working with the Philippine government to expand the programme.

There are currently two ways in which NLP is explicitly being used to address health challenges in resource-poor settings. First, it is being used for surveillance and outbreak predictions using data from electronic health records and online media and social media sources. Global Health Monitor is one example of an online system for detecting and mapping infectious disease outbreaks. The tool works by identifying English-language news stories and classifying them for topical relevance. Relevant articles are then plotted on a map using geo-coding information, which can help epidemiologists and programme managers monitor the spread of diseases.25

NLP is also being used to analyse unstructured text in the medical literature and EMRs to support clinical decision making or track population health behaviour. Using data from the web, for example, NLP has been applied to a wide range of public health challenges, from improving treatment protocols to tracking health disparities.26 27 NLP and machine learning are also being used to guide cancer treatments in low-resource settings including in Thailand, China and India.28 Researchers trained an AI application to provide appropriate cancer treatment recommendations by giving it descriptions of patients and telling the application the best treatment options. The AI application uses NLP to mine the medical literature and patient records—including doctor notes and lab results—to provide treatment advice. When examining different patients, this application agreed with experts in more than 90% of patients in one study and 50% in another.28

AI planning is also already being applied to improving the provision of primary healthcare services in resource-poor settings. These tools—versions of which are under development or are being piloted—are being applied to improve the efficiency of programmes ranging from immunisation, to supply chain, or referral services within a multitiered system. CHWs often visit several households per day—often separated by long distances—where they provide health information, dispense some medicines or perform limited medical procedures. A partnership comprising researchers and a social enterprise have been developing an AI planning application for optimising CHW scheduling in communities in Africa.29 As AI planning, like many AI disciplines, relies on the availability of relevant high-quality data, advancements in mobile connectivity and digitised health system performance and resource allocation data present additional opportunities for AI planning in resource-poor settings.

Signal processing is another related area that could be buoyed by the rapid expansion of mobile devices that can capture and transmit signals and the emergence of cloud computing. Specific opportunities in resource-poor settings currently focus on signals that can be collected with mobile phones or single-function, add-on devices that link with mobile phones or multipurpose instruments that can collect multiple kinds of digital data. Such signal processing opportunities in resource-poor settings are promising, particularly when paired with machine learning principles and cloud computing. For example, signal processing and machine learning have been used in Nigeria to predict birth asphyxia using a mobile phone.30 The application uses the birth cry of children collected using a mobile device to identify children likely to be experiencing birth asphyxia in resource-poor settings. Applications that collect and transmit signals collected by sensors that continuously record whatever is being monitored (eg, activity trackers) are further upstream in such settings.

As resource-poor settings become more connected and the data they produce are of higher quality, the ability of AI to address health challenges will likely expand. Furthermore, as countries develop, several of the emerging digital technologies described here could be employed to monitor patient populations and medical data to recognise patterns that could potentially identify a pending outbreak or other public health emergencies. This would allow public health agencies, such as the WHO, to investigate and monitor in real-time to model cause-and-effect relationships that could mitigate the progression of an epidemic.

Unanswered questions and remaining challenges

Sheikhtaheri et al 31 describe many challenges with designing and implementing expert systems for supporting clinical decision making. In general, they conclude that the successful implementation of any expert system requires a clear definition of the clinical problem to be addressed. Building and updating the knowledge base is challenging in the best of circumstances and would be compounded in resource-poor settings. Many expert systems also lack an accuracy tracking mechanism, which could undermine the trust of clinicians and patients. Sheikhtaheri et al 31 also raise interesting questions of whether all disease domains require expert systems and how various AI systems should be integrated. They suggest that the lack of answers for these questions  has prevented expert systems from advancing from research to implementation.

Supervised machine learning applications require high-quality datasets that can be used to train machine learning algorithms to identify risk factors or make disease diagnoses.31 For many diseases and conditions relevant to resource-poor settings, such datasets can be difficult and time-consuming to collect.31 In addition, better diagnosis does not equate to access to appropriate or quality treatment options. While remote diagnostics and machine learning applications might help to identify diseases, there is no guarantee that the condition identified can be treated in any given setting. As such, the principle of ‘do no harm’ and the ethics around providing treatment following testing for and confirmation of disease (‘test and treat’) when deploying AI applications are relevant as much as in any other context.

Carrel et al 32 describe some of the challenges associated with adapting clinical NLP systems to diverse healthcare settings. Using colonoscopy screening as an example, they highlight the substantial resources necessary to compile the natural language corpora, address different record structures and deal with idiosyncratic linguistic content. These challenges would likely be multiplied when considering healthcare settings in low-income countries. Countries in low-resource settings sometimes maintain hand-written health records in local languages. Therefore, building the natural language corpora could require substantial effort. The WHO has advocated for the adoption of standardised medical terminologies or the development of local data dictionaries to address some of these challenges.33

Substantial data are necessary to build and implement automated planning and scheduling applications.29 Compiling such data in resource-poor settings is often difficult and time-consuming. Brunskill and Lesh describe the extensive data collection involved in laying the groundwork for the development of an improved CHW schedule in sub-Saharan Africa.29 Furthermore, the effectiveness of automated planning and scheduling applications will depend largely on the quality of data used to develop the application. High-quality health systems data are currently difficult to collect in many resource-poor settings. Such challenges could slow the development of automated planning and scheduling in the settings where they would be most needed.

While internet connectivity is improving throughout the world, some resource-poor settings remain without access to the substantial bandwidth necessary to upload very large signal datasets to the cloud. Some applications, including mHealth tools used by CHW, have the ability to work offline and sync with remote databases when the bandwidth is sufficient. Storing such data locally could also require substantial investment in IT infrastructure. While device prices are currently also a barrier, these are predicted to come down over time as companies take advantage of economies of scale.

There are also environmental challenges that need to be considered. The use of AI in resource-poor settings require a strong understanding of local social contexts, infrastructure requirements and additional related infrastructure needs, including IT, communications networks and platforms for delivering primary health services. Many AI applications depend on the availability of strong electronic health record systems, which require substantial investment to put into place. In addition, AI applications will have limited impact if they do not effectively integrate languages and scripts used in the electronic health records of many developing countries.

In high-income countries, discussions around the ethics of electronic health records and AI have focused largely on privacy, confidentiality, data security, informed consent and data ownership. Most of these same considerations apply to resource-poor settings. However, the relevance of these issues varies depending on differences in culture, literacy, patient–provider relationships, available IT infrastructure and regulatory issues in LMICs. One proposed approach for maintaining secure and transparent health records is the use of ‘blockchain’, the distributed ledger system known primarily for its use by crypto-currencies.

In addition, some experts have raised concerns that some AI applications can potentially exacerbate inequities, including those related to ethnic, socioeconomic and gender. They note that cultural prejudices can be reflected in data, algorithms and other aspects of AI design.34 One recent report found that an application that uses arrest records, postal codes and socioeconomic data to assess the risk of recidivism in US courts was biased against black citizens.35 These challenges are compounded by the fact that many AI algorithms are a ‘black box’ and are therefore less likely to be assessed for bias. However, some researchers are working to assess biases by testing how well they predict by randomly changing key variables for individuals for whom AI applications are attempting to make a prediction.36 For AI to benefit all, including those in resource-poor settings, these biases need to be considered in the design of such applications. In health, ensuring that more women and those resource-poor communities are involved in the development of AI applications will also help to reduce such biases.

The generation of large amounts of data naturally raises questions around the ownership of the data and who can access which specific data for research or commercial purposes. While there have been some initial efforts to address this, including the Data Sharing Principles in Developing Countries, which was put forward in Nairobi in 2014, there has yet to be widespread adoption.37 The underlying concept that initiated  the development of these principles is that data generated with public funds should be viewed as a public good. Establishing a repository for large amounts of data for global health would help in ensuring that such data are made available as a global public good. Many LMICs have signed agreements that data generated with use of public funds should be freely available.37 However, both from patient and developer perspectives, privacy laws and data access and ownership agreements are perceived to be potential threats to successful AI applications and they should therefore be monitored closely by groups working to develop such applications for particular contexts. Applications using new technologies like blockchain, may also help resolve some of the current concerns.