Artificial intelligence and public health: prospects, hype and challenges

Introduction

We live in interesting times. Each week sees a clutch of new publications identifying potential applications of artificial intelligence (AI) in healthcare. These promise a digital revolution in healthcare through improvements across a range of clinical practices and tasks as well as streamlining many aspects of clinical and health services operations. Because of the rate of growth of information and its promotion, some health organisations have developed ‘living evidence’ databases to make sense of this rapidly evolving landscape.¹

AI is a widely but often inaccurately used term. Although various definitions exist, AI refers to the ability of algorithms encoded in technology to learn from data so that they can perform automated tasks without every step in the process having to be programmed explicitly by a human.² AI algorithms are ‘trained’ using large datasets to identify patterns, make predictions, recommend actions and progressively adapt to respond to unfamiliar situations – learning from new data and thus improving over time. The ability of an AI system to improve automatically through experience is known as machine learning (ML).³

Although most attention has been given to the possibilities for applications of AI in healthcare, practical applications in public health are now being considered more closely. This paper provides a brief summary of the progress in the use of AI in healthcare to consider potential crossover in technology use between healthcare and public health before considering the broader potential for the use of AI technologies in public health interventions, analysis and research. We consider the applications of AI that have more immediate, practical implementation at scale in public health; what needs closer scrutiny, development and evaluation; and where we need to exercise caution going forward.

Artificial intelligence applications in the health system

There has been understandable excitement at potential AI applications in healthcare that make use of accessible data to provide clinical decision support for diagnosis, prognosis and disease management. For example, AI algorithms can be applied to the analysis of medical images, such as X-rays, magnetic resonance imaging and computed tomography scans, to speed up diagnoses and have the potential to match or improve comparable human accuracy.^4,5 AI algorithms can also be applied to the analysis of patient data to predict the likelihood of outcomes, such as readmission, complications, or progression of diseases, to assist with patient management and resource allocation.⁵ AI platforms can also be used to support continuous monitoring of patients’ health data from wearable devices or sensors, detecting changes in health status in real time and alerting healthcare providers to intervene when necessary.⁵ This may be of particular use in managing chronic conditions and preventing hospital readmissions, but also has obvious potential for application in public health, which is discussed further below.

AI also has the potential to support more effective operational management of health services, for example through the use of algorithms to predict patient admission rates, staffing needs, streamlining workflows and resource allocation.^6,7 Natural language processing can be used to interrogate unstructured clinical notes, with the potential to convert them into structured data that can be more easily analysed for quality improvement, decision support and research purposes.⁶

AI also has potential applications in the analysis of high-volume biological and chemical data to assist with drug discovery and development, as well as in the personalisation of treatment to support optimally targeted interventions.⁶ Virtual health assistants and chatbots are already being deployed to provide patients with personalised health advice, assistance in scheduling appointments, medication reminders and answering health-related questions as shown in Parmar et al. 2022.⁸ Again, several of these technologies have potential for application in public health.⁹

All of these potential applications are being developed, tested and exploited with a mix of serious scientific study and (sometimes) over-hyped commercial interests. To date, most reports of early experiences and (limited) research generally offer a positive view of potential, but they are often reporting on implementation in optimal circumstances and with motivated providers and consumers. Most potential applications are yet to be closely scrutinised as to their feasibility for scale-up in large organisations and health systems.

The adoption and sustained use of these technologies at scale in complex healthcare systems are rapidly evolving but are not currently established. Adoption is most observable in the world’s wealthier countries, and the World Health Organization is already warning that the existing digital divides between and within nations could exacerbate long-standing inequitable access to healthcare technologies.² Concerns about inequitable access, uptake and impact of AI-assisted technologies are of great significance in the consideration of their impact on public health.

Artificial intelligence in public health

In our public health communities, progressive growth in the volume and accessibility of health-related data combined with improvements in computational power and related speed have, for some time, created the potential for more complex and ambitious analysis of population health aetiology, health trends and future predictions. The use of ‘big data’ – volumes of large, complex, linkable information – for these public health purposes has been embraced by some public health scholars and practitioners as a tool for supporting better targeting of potential interventions to individuals and communities and has enabled the evolution of the concept of ‘personalised public health’.^10,11

These same enthusiasms have also led to advocacy for the use of AI in public health over the past decade. Several excellent reviews have examined both the potential and practical application of AI in public health in epidemic and pandemic early warning systems (EWS),¹² in synthetic data generation,^13,14 in sequential decision-making in uncertain conditions (reinforcement learning; RL)¹⁵ and in disease risk prediction.¹⁶ These are all considered below.

Although most of these applications require specialist skills to develop, use and interpret outputs, the launch of more publicly accessible AI platforms, such as OpenAI’s ChatGPT, has prompted broader awareness of and access to the technologies for public health researchers and practitioners. There are some immediate practical public health uses for these new tools in health education and primary prevention.¹⁷ Most written health information (in print and online) uses technical language and requires a reading level that is not aligned with the health literacy of most consumers and patients.¹⁸ This is especially the case for those who are older, with lower education and those who do not speak a community’s dominant language. This disconnect between the use of technical language in health information and its utility to consumers can be found in information developed by government, health services and non-government organisations.¹⁹

Rapidly evolving AI platforms, such as ChatGPT, have the potential to significantly improve accessibility and understanding of all forms of written health information. ChatGPT is one of a number of large language models (LLMs) that have been trained on an extensive database of text data to produce human-like responses to prompts, such as questions or requests, to produce written information within defined limits (for example, Supplementary file S2 provides a description of current applications of AI in public health, disease prevention and health promotion generated by ChatGPT). These LLMs do not synthesise or evaluate evidence, but rather they predict what should come next in written text through learning from large volumes of training data.³ When asked to simplify existing health information, ChatGPT can, on average, improve the grade reading score of texts, use less complex language and provide information in an active voice. Work by Ayre and colleagues demonstrated that the use of simplified language could be achieved whilst retaining 80% of the key ‘technical’ messages.²⁰ These improvements were particularly notable for texts that were more complex to begin with. The authors concluded that ChatGPT can provide a useful ‘first draft’ of plain language health information, highlighting that further human intervention is required to check accuracy and fine-tune messaging.²⁰

These same technologies can also provide the same information in multiple languages for those working in diverse communities.²¹ Current technology is sufficiently advanced to provide good working translations that make the most efficient and value-adding use of a human interpreter’s time to ensure cultural sensitivity and relevance. Over time, as AI platforms evolve through training, platforms like ChatGPT will become more capable of communications that reflect language and cultural nuances.²¹

AI platforms, such as ChatGPT, are already being used in communities for simplifying language and for translation; but, with experience, a number of emerging risks are becoming better understood. The information provided by platforms, such as ChatGPT and others, comes in accessible, persuasive conversational language but may be inaccurate. Responses may be different depending on the form of the prompt question.²² Concerns have also been consistently raised that the advice will perpetuate existing inaccuracies and biases in society in the source material from which machines are ‘learning’, leading to bias by gender, community of origin and social group, as examples.²³

In response to such concerns, LLMs can be improved by careful parameter tuning – restricting the materials from which they learn to credible, ‘pre-approved’ sources while preserving similar capabilities. Restricting source materials and specifying internal dialogues will not only lead to more evidence-based advice but also allow for easier tracking of reasoning to reach the output. Work on fine-tuning LLMs in this way is currently underway but at a very early stage as discussed in Christophe et al. 2024.²⁴

In a related application, AI can also be used to provide targeted, personalised advice on how to address health risks by matching individual characteristics, preferences and needs. The practical application of this can be seen in the increasing use of chatbots and virtual health assistants as conversational agents that mimic human interaction through written, oral and visual forms of communication with a user.²⁵ Chatbots offer the flexibility of on-demand, personalised support and content and consistent connectivity.²⁶ They can adapt their responses based on the user’s input and preferences and are already being used as behaviour change tools in areas including promoting healthy lifestyles, mental health, smoking cessation and reduction in substance misuse.²⁶ Importantly, they are an application of AI that can be brought to scale.

Although great potential exists, reviews of the use of AI chatbots in health behaviour change have concluded that despite promising results, studies had low internal validity, lacked sufficient description of AI techniques used and lacked generalisability.²⁶ These findings highlight the importance of further research, with robust methodologies to move us past the advocacy of programmers to draw more definitive conclusions about the optimal uses of AI in public health education. Part of this should, again, involve the use of active parameter tuning of the materials from which machines learn.

Beyond applications of generic, publicly accessible tools in health education and communication, AI algorithms are also being used to augment existing epidemic and pandemic EWS. By processing and organising vast amounts of data from diverse sources, such as medical records, social media and environmental monitoring devices, AI platforms and customised algorithms offer the potential to detect patterns and anomalies that may indicate the onset of a disease, an infection outbreak, an epidemic or a pandemic.¹² An excellent review of AI applications for epidemic and pandemic EWS by El Morr and colleagues¹² indicates that a wide variety of AI-based EWS has been effectively implemented in different countries, using a variety of AI techniques with a diversity of approaches. The authors concluded that effectiveness of these AI techniques varied across studies but demonstrated ‘promising results’ in predicting disease outbreaks, with advanced ML techniques consistently achieving higher accuracy rates than traditional statistical methods. There is an increasing volume of studies reporting on this potential application of AI to public health surveillance, and the review by El Morr identifies some of the areas of consistency, highlighting the benefits of ‘random forest’ ML algorithms that combine the output of multiple decision trees to reach a single result. However, the diversity of reported experimental applications makes it hard to provide clear guidance at this stage in the evolution of the methods, and the authors conclude that challenges remain with data quality, bias and model transparency.¹²

A further potential application of AI in public health is in synthetic data generation. Synthetic data is created by statistically modelling original data and then using those models to generate new data values that reproduce the original data’s statistical properties. In essence, the use of synthetic data allows researchers to conduct a wide range of experiments and simulations without the risk of exposing the patients’ identity.¹⁴ It also improves data utility while preserving the privacy and confidentiality of information.¹³ It is a promising solution to overcome the challenges posed by data scarcity and privacy concerns in public health, for example when studying communicable diseases and stigmatised populations where there are several barriers to data sharing, including people diagnosed with HIV and illicit drug users. Synthetic data also help overcome the challenge of training AI algorithms on unbiased data with sufficient sample size and statistical power.¹³ Through this approach, the use of synthetic data can facilitate the development of more diverse and accessible data to improve the generalisability of AI models across diverse populations. However, it is critical that potential biases in synthetic data generation are addressed to ensure fairness and equity in the AI models, particularly in data with underrepresented populations.¹⁴

The use of RL in public health has also been advocated. RL is the subfield of ML focused on optimal sequential decision-making under uncertainty.^15,27 RL represents a holistic approach to decision-making that evaluates the impact of every action (i.e. data collection, allocation of resources and treatment assignment) in terms of short-term and long-term utility to stakeholders. In an excellent review paper on potential applications in public health, Weltz and colleagues promote RL as an ideal model for a number of complex decision problems that arise in public health – as a means of efficiently using data to inform optimal evidence-based decision-making. The authors point to the use of RL in several domains, including precision medicine, which uses ML to identify how to tailor interventions to the uniquely evolving health status of each patient.¹⁵ Weltz and colleagues¹⁵ also point to potential public health applications, specifically highlighting the potential to support resource allocation ‘if, when, and where they can be most impactful’. The authors acknowledge current limitations that have stifled uptake in public health, including the volume of high-quality data needed for analysis, learning and adaptation; divergent objectives among stakeholders; and the same challenges in bias in data sources that are faced in all current AI applications.¹⁵

AI algorithms are also being used to analyse large datasets to identify patterns and predict health risks or outcomes for individuals.¹⁷ By assessing factors such as demographics, medical history, health-related behaviours and environmental risks, generative AI can be used to predict who is at higher risk for future diseases. Despite the promise, the evidence suggests that this application is emerging and significantly untested rather than ready for practical use. A recent review of the use of AI in risk prediction models of cardiovascular disease identified 486 different models, of which the majority were in development (n = 380), and none had undergone independent external validation.¹⁶

Finally, although this paper has focussed on the positive potential of AI technologies and platforms to improve public health, they can also be maliciously used to amplify the problem of health-related misinformation and disinformation.²⁸ The infodemic challenge has emerged in the past decade and came to much closer community attention during the COVID-19 pandemic. In the worst-case scenario, despite efforts to limit such misuse, deliberate manipulation of chatbots by economic and political interest groups, cyber criminals, or ‘disinformation farms’ can already be observed.²⁸

Making progress with caution

Despite the evident potential of AI platforms and technologies to assist in health promotion, disease surveillance and public health research, full realisation of this potential requires continued experimentation and fine-tuning, especially to minimise bias in source data and establish clearer guidance on optimal applications.

The use of AI to provide more personalised information, real-time feedback and support for individuals in their health behaviour and decision-making has the most immediate and practical public health applications. But the populations we serve have very different capabilities to respond to the emerging opportunities that AI presents, and for this reason, we need to be wary of unintended consequences of AI use, including further exacerbation of existing health inequalities.²⁹ Generally, those with greater education and greater access to resources are most likely to secure the benefits of these emerging technologies.³⁰ Community members vary considerably in their access to digital technology, in their skills to discriminate the accuracy and reliability of information, and in their trust and responsiveness to what digital technologies have to offer.³¹ Hand-in-hand with advances in technology, public health practitioners will need to work with communities to help build digital health literacies that enable all community members to find, understand and have confidence in reliable sources of information about their health.¹⁸

The potential for AI to offer highly personalised behavioural advice and continuous follow-up at scale makes it a valuable public health tool. But care needs to be taken to ensure that the form of advice it provides doesn’t inadvertently lead to new types of victim-blaming in public health. Although an attractive concept, ‘personalised public health’ also runs the risk of directing attention away from critical social and economic determinants of health that fundamentally limit the ability of individuals to respond to personalised exhortations to behave differently.

At this stage in its evolution, the potential applications of AI in public health education and communication appear to favour those who are already most advantaged and may have the unintended effect of drawing our attention to individual responses to complex public health problems at the expense of examining the social and economic conditions that drive existing inequities.

Beyond the more immediate ‘direct to public applications’, there are several promising, emerging applications that, with refinement, have great potential to support epidemic and pandemic predictions and responses and resource allocation decisions, as well as provide techniques for improved population health surveillance, research quality and representativeness. These all need further development, testing and training of the public health workforce to convert the potential into material improvements in public health research and practice.

At this point in the advancement of AI and its use to support public health, it is essential that there is a strong public health voice in current debates about the evolution of the technology and its regulation. Integrating a public health perspective into public and political discourse will be essential to ensuring that the great potential uses of AI in improving public health are actually realised. This involves a more systemic, critical evaluation of AI applications and their practical uses, transparency in their impact on equity, and public accountability for ethical use of these technologies across different populations and settings. By advocating for policies and practices that prioritise public health considerations in the evolution and application of AI, we have a better chance of optimising the potential of AI to improve health outcomes and wellbeing. By establishing clear ethical, legal and regulatory frameworks, we can harness the power of AI to advance public health goals in a responsible and equitable manner.

Conclusion

The application of AI represents a genuine inflection point for many aspects of our lives. It has the potential to revolutionise healthcare delivery, improve population health outcomes and advance broader global health goals. However, realising the full potential of AI requires careful consideration of its ethical implications, thoughtful regulation and collaborative efforts to ensure that these technologies serve the best interests of individuals and communities worldwide and don’t exacerbate health inequalities. The scientific process should be used to develop the evidence base that informs regulatory and policy frameworks that govern the use of AI in healthcare and public health. By informing policy decisions pertaining to the use of AI in healthcare and public health with scientific research and evidence, we can ensure that AI meets rigorous standards for safety, efficacy and ethical conduct.

Supplementary material

We asked ChatGPT to summarise current applications of AI in public health, disease prevention and health promotion. Supplementary File S2 contains the response we received. Supplementary material is available online.