Introduction

Generative artificial intelligence (AI) has been a hot topic in recent months. Early adopters have begun to apply generative AI to various fields such as healthcare, where it has widespread uses, including but not limited to analysing medical images more rapidly and diagnosing the presence and severity of diseases, identifying early prevention opportunities and educating patients at risk of developing a disease, predicting patient outcomes and personalising care and treatment to individual patients. Generative AI can also speed up the discovery of new drugs by proposing new chemical compounds and synthesis methods to researchers.

In this paper, we will describe several use cases in healthcare for generative AI and explore the potential implications of this growing technology for healthcare payers.

Case study 1: Diagnostics

The analysis of medical images, such as CT, MRI and PET scans, using AI technology has the potential to significantly improve the speed and accuracy of diagnosis, potentially enabling early detection of disease and allowing treatment to be initiated promptly. This not only may improve patient outcomes but could also help optimise healthcare resources by reducing the need for unnecessary tests and interventions.

Pulmonary nodule detection AI is one of the most important diagnostic imaging AIs, often positioned as a medical device that reduces the risk of false negative diagnoses. In addition, by providing quantitative information on nodules, it enables the following evaluation methods for lung cancer, which have been difficult to perform in the past:

• Assessment of malignancy using radiomics (e.g., internal heterogeneity).
• Evaluation of malignancy using volume doubling time.
• Evaluation of treatment efficacy using tumour volume / mass.

The information added by diagnostic imaging AI is simple statistics on regions of interest, which is not necessarily a new concept in medicine. However, generative AI may be useful in integrating the findings of imaging AI to provide diagnostic support to radiologists. For example, it may be possible to comprehensively assess the malignancy of lung cancer by integrating image findings such as the primary area of the lung section, internal density and edge heterogeneity, along with the growth rate in the same way as human radiology.¹ While this function alone might be possible for a single diagnostic imaging AI, generative AI would be more technically compatible with the development of an AI that uses all the information available to the doctor, including the integration of information about the patient's clinical status and complications. This is because most medical information other than images is documented, so it is relatively easy to feed this broader contextual information into the generative AI engine.

It may also be possible to integrate the image findings (e.g., a list of “There is xxx shadow in lung section yyy”) from an imaging AI to list differential diagnoses and ultimately generate diagnostic imaging reports. In doing so, the generative AI can also refer to the patient's chief complaint, the imaging request form and information in the patient’s medical record. This is expected to enable comprehensive thinking and accurate diagnosis, which has been difficult with imaging diagnosis AI alone.

To achieve these things, generative AI needs to have medical knowledge or a system that can refer to medical best practice guidelines etc. Several studies have investigated the level of medical knowledge of generative AI, and it has been shown to provide information that exceeds the average for physicians.^2,3,4

Case study 2: Diabetes disease management

Diabetes is a chronic disease affecting more than 350 million people worldwide, and its prevalence is expected to increase in the coming years.⁵ The American Diabetes Association recommends that patients with diabetes use pharmacological therapy, dietary therapy, weight loss and exercise programmes to manage symptoms and prevent complications of diabetes by keeping blood glucose levels within a target range.⁶ In addition, among US adults aged 18 years and older, 2017-2020 crude data revealed that 38.0% had prediabetes, based on their fasting glucose or A1C level, while only 19.0% of adults with prediabetes were diagnosed as such⁷ (i.e., had been told by a health professional that they had the condition).

There have been recent advances in blood glucose monitoring that can provide data to generative AI for analysis and patient-specific information. Continuous glucose monitoring (CGM) devices provide instant, real-time measurements of blood glucose levels and rates of change, alerting the user to impending glycaemic events and motivating the patient to take remedial action.⁸ Technology company Apple is rumoured to be developing a wearable CGM device that uses optical absorption spectroscopy to determine blood glucose levels without the need for a skin prick or subcutaneous device.⁹

Generative AI can help people with diabetes manage their condition in several ways:

• Predicting blood glucose levels: A generative adversarial network was used to predict 30-minute blood glucose levels with an average root mean square error of 18.34 mg/dL. Predicting blood glucose levels may help patients avoid adverse glycaemic events, including hypoglycaemia and hyperglycaemia, by administering appropriate doses of medication.¹⁰
• Patient education: Generative AI-powered chatbots can provide patients with quick advice on how to manage their chronic conditions. Chatbots can answer patients’ questions, provide advice on how to live a healthy lifestyle or suggest individualised meal plans for the diabetic patient, especially those who lack easy access to a face-to-face doctor.¹¹
• Calorie calculation based on images: Recent research has shown increasing accuracy in estimating the nutritional value of meals based on images. Researchers have developed an algorithm using generative AI that can estimate the calorie content of food with an average error of 209 kcal per meal.¹²

As diabetes is typically one of the top five areas of healthcare spend, targeting prediabetes diagnosis and subsequent preventive care interventions to narrow the gap between those who are predisposed to diabetes and those who follow a preventive care programme may be an important future role for generative AI.

Case study 3: Drug discovery

Traditionally, there have been four phases in the drug development process, and, according to Víctor Gallego et al.,¹³ all of them can add up to many years, if not decades, before a new treatment or medicine is brought to market. The process starts with many competing molecules only a small fraction of which reaches the final stage. According to Kit-Kay, M. & Mallikarjuna, R.P.,¹⁴ only about 38% of new chemical entities in Phase IIb and Phase III clinical trials reached the market, with safety and efficacy being the main reasons for failure. According to Chen et al.,¹⁵ drug development can be processed via retrosynthesis, which means that scientists start with the desired molecules that might be suitable for a specific disease and work backwards, converting the target molecule into chemical reactants and planning a multi-synthesis pathway. This process can be time-consuming, as testing multiple reactions can be challenging and costly. For example, the Tufts Centre for the Study of Drug Development (CSDD) found that drugs and biologics approved by the US Food and Drug Administration (FDA) spent an average of 89.8 months in clinical trials between 2014 and 2018.

Generative AI uses neural networks to identify patterns and structures in existing training data. It can create new data samples based on these patterns, and it is powerful because it can leverage different learning approaches, including unsupervised learning. In drug development, generative AI can help discover new molecular structures for drug candidates, can automate chemical reactions for given molecules, and can improve the selection of patients or biomarkers for clinical trials. As a result, generative AI has the potential to significantly accelerate the traditional drug development process, particularly the initial discovery of drugs, to a significant extent. At the same time, it is important to note that, while generative AI can speed up the process, the time to market for drugs developed with generative AI is still dependent on clinical trials, which can take several years to complete. However, generative AI could also optimise the clinical trial process by identifying the most suitable patients for the trial and ensuring that the trial is conducted efficiently.

An example of how generative AI can help in drug discovery is the use of the G2Retro framework developed by researchers at the Ohio State University (Chen et al.). This AI framework can generate multiple potential chemical reactions for any given molecule and can quickly determine which of these reactions is best suited to produce the target drug molecule. To test the system, the Ohio State team trained G2Retro on a dataset of 40,000 chemical reactions collected between 1976 and 2016. The AI framework used deep neural networks and unsupervised learning to learn from the representations of a given molecule and generate potential reactants, or input chemical compounds, that could be used to synthesise the desired target molecule via multiple synthetic pathways, while ranking the different options. In addition, the team conducted a validation case study using four drugs already on the market and concluded that the generative AI system could accurately generate the same patented route for these drugs but was also able to provide alternative feasible synthetic routes.

There are several systems, such as those described above, that have the potential to improve the speed at which drugs are brought to market. However, researchers still need to test these molecules in clinical trials to see whether the process can be validated. At the same time, it will be important to carry out further case studies of generative AI processes to see their “thought pattern” for the many already existing and patented drugs. This may provide additional reassurance that newly created drugs are clinically safe and effective.

Insurance implications

The technologies described above have many potential implications for health payers, such as insurers and governments. In the example of using generative AI to improve disease management care for chronically ill patients, payers will need to make benefit coverage decisions based on the expected costs of management versus the likely financial benefit over an agreed timeframe, as well as customer satisfaction considerations. Payers will also need to develop reimbursement mechanisms for technology that may need to look very different from traditional fee-for-service (FFS) environments.

Subscription or capitation payments to third parties for technology are likely to become more common and may make up an increasingly large part of the insurance premium. While a predictable capitation-based payment for third-party monitoring and disease management technology may be relatively easy to forecast compared to fee-for-service claims, third-party pricing is likely to become more sophisticated over time, differentiating by age and chronic disease profile, depending on usage. This may result in changes to the typical age curve for medical insurance and different types of trends to consider for forecasting purposes.

Actuaries will need to consider the medium- to long-term implications of the various generative AI applications and the likely impact on medical inflation, considering the types of claims that may increase or decrease and changes in the relative proportion of these claims over time, when forecasting medical inflation. Far more granular analysis of claim cost trends is likely to be required, which means detailed information on different kinds of services and the cost drivers for those services will be needed. For example, data linkages of usage of generative AI monitoring devices and subsequent claims at a patient level will be necessary. This has proved problematic to date, as many of the third-party services used by insurers do not provide patient-level usage data to enable linkages to secondary care claims.

Historically, a key driver of medical inflation has been the introduction of new drugs and technologies, often to treat previously untreatable conditions. If the routes to market for new medicines are made faster and easier using generative AI, then insurers will need to increase their horizon-scanning activities and be prepared to make rapid changes to benefits, medical policy, claim processing capabilities and policy terms and conditions where necessary. Agility in governance and systems will become even more important, given the potential of new and possibly expensive personalised treatments to drive up medical inflation.

Higher medical inflation typically leads to a higher degree of uncertainty in the forecast of medical inflation. This potential additional price volatility may affect solvency requirements of health insurers, as a large part of the solvency capital requirement under risk-based capital regimes is typically related to mispricing risk.

Other implications for insurers include emerging risk management around data privacy and customer fairness and treatment. Much has been written about potential model bias and the perpetuation of historical discrimination, which may be implicit in the data used to train generative AI. Board members will want assurances that any use of generative AI does not inadvertently discriminate against certain types of customers.

Thoughts for the future

Whether it is a potential reduction in the need for unnecessary tests and interventions, faster drug development or more efficient disease management through potentially earlier identification of the underlying pathology or determination of the most appropriate interventions, the use of AI in personalised medicine appears to be getting a boost with generative AI. This power can also be harnessed for preventive medicine. With 2.8% of healthcare costs in the EU spent on prevention, according to Eurostat, the use of generative AI in prevention could also be transformative. By analysing the huge amount of patient data already available, generative algorithms can predict the risk of certain individuals developing certain diseases or complications of diseases. Generative AI algorithms can potentially be used to assess whether a particular individual should be targeted for intervention because they are at increased risk of developing a disease or experiencing a disease complication.

As in many other fields, generative AI has the potential to be transformative in healthcare. For insurers, there is a risk that their systems and capabilities will struggle to anticipate the high levels of short-term uncertainty around the operational implications and impact on medical inflation. Rapid changes in medical practice patterns, and therefore health insurance claims, mean that insurers need to be more agile and understand emerging supply chain risks in order to shape operational responses in claim management, product design and pricing capability. Those insurers with the foresight to understand the need for granular patient-level claim data with rich clinical insights to understand and predict trends will have a competitive advantage. Insurers that have focussed on offering more third-party services as a customer benefit but have not integrated the data streams from these services, will find it difficult to understand and react to fast-moving trends.

Another important consideration for generative AI in the healthcare field is the expected regulation in the EU, specifically the AI Act.¹⁶ AI systems used to diagnose and initiate treatment, due to the potential for harm to patients if recommendations turn out to be incorrect; systems used to monitor patients’ vital signs, due to the potential for missing needed treatment; and drug discovery processes, due to the potential for developing ineffective or harmful drugs, or robotic surgical assistants, due to the potential for errors during surgery that can harm patients, are all examples in healthcare and life sciences where the use of generative AI can pose many risks and is likely to be subject to significant regulation in the EU and other jurisdictions, which may slow adoption.

The AI Act includes regulatory compliance, risk assessment and the use of transparent and accountable systems, as well as more stringent requirements for usage of generative AI in business services. It also expands “high-risk” scenarios included in a more comprehensive testing for users of generative AI systems. As such, the AI Act will have a significant impact on the use of generative AI in healthcare and life sciences, and users of such systems will need to comply to ensure the safety and effectiveness of these processes.

¹ Fleischner Society pulmonary nodule recommendation.

² Daiju Ueda, Shannon L Walston, Toshimasa Matsumoto, Ryo Deguchi, Hiroyuki Tatekawa, Yukio Miki (5 May 2023). Evaluating GPT-4-based ChatGPT’s Clinical Potential on the NEJM Quiz. MedRxiv Retrieved 4 February 2024 from https://doi.org/10.1101/2023.05.04.23289493.

³ Peter Lee, Sebastien Bubeck, Joseph Petro (2023). N Engl J Med; 388:1233-1239 DOI: 10.1056/NEJMsr2214184.

⁴ David A. Asch (4 April 2023). NEJM Catalyst.

⁵ Pouya Saeedi, Inga Petersohn, Paraskevi Salpea, Belma Malanda, Suvi Karuranga, Nigel Unwin, Stephen Colagiuri, Leonor Guariguata, Ayesha A Motala, Katherine Ogurtsova et al. (2019). Global and regional diabetes prevalence estimates for 2019 and projections for 2030 and 2045: Results from the international diabetes federation diabetes atlas, Diabetes Research and Clinical Practice, 157, 107843.

⁶ American Diabetes Association (2014). Standards of medical care in diabetes—2014. Diabetes care, 37 (Supplement_1), S14-S80.

⁷ Centers for Disease Control and Prevention. National Diabetes Statistics Report. Retrieved 5 February 2024 from https://www.cdc.gov/diabetes/data/statistics-report/index.html.

⁸ David Rodbard (19 January 2016). Continuous Glucose Monitoring: A Review of Successes, Challenges, and Opportunities. Diabetes Technology & Therapeutics. S2-3-S2-13. Retrieved 5 February 2024 from http://doi.org/10.1089/dia.2015.0417.

⁹ Clover, Juli (22 February 2023). Apple’s Noninvasive Blood Glucose Technology for Future Apple Watch Reaches ”Proof-of Concept” Stage.” MacRumors. Retrieved 5 February 2024 from https://www.macrumors.com/2023/02/22/apple-blood-glucose-monitoring-technology/.

¹⁰ Zhu, Taiyu et al. (2020). Blood glucose prediction for type 1 diabetes using generative adversarial networks. CEUR Workshop Proceedings. Vol. 2675.

¹¹ Javaid, Mohd, Abid Haleem & Ravi Pratap Singh (2023). ChatGPT for healthcare services: An emerging stage for an innovative perspective." BenchCouncil Transactions on Benchmarks, Standards and Evaluations 3.1: 100105.

¹² Fang, S., Shao, Z., Kerr, D.A., Boushey, C.J. & Zhu, F. (18 April 2019). An End-to-End Image-Based Automatic Food Energy Estimation Technique Based on Learned Energy Distribution Images: Protocol and Methodology. Nutrients, 11, 877. Retrieved 5 February 2024 from https://doi.org/10.3390/nu11040877.

¹³ Gallego, V., Naveiro, R., Roca, C., Ríos Insua, D. & Campillo, N.E. (25 August 2021). AI in drug development: A multidisciplinary perspective. Mol Divers. (3):1461-1479.

¹⁴ Kit-Kay, M. & Mallikarjuna, R.P. (2019). Artificial intelligence in drug development: Present status and future prospects. Drug Discovery Today;24(3):773–780.

¹⁵ Chen, Z., Ayinde, O.R., Fuchs, J.R. et al. (2023). G2Retro as a two-step graph generative models for retrosynthesis prediction. Communications Chemistry 6, 102.

¹⁶ European Parliament (9 December 2023). Artificial Intelligence Act: Deal on comprehensive rules for trustworthy AI. Press release. Retrieved 5 February 2024 from https://www.europarl.europa.eu/news/en/press-room/20231206IPR15699/artificial-intelligence-act-deal-on-comprehensive-rules-for-trustworthy-ai.