The application of artificial intelligence (AI) to improve overall health and wellbeing is growing steadily in the medical and healthcare sectors. Cost savings from AI applications in US healthcare may have reached $52 billion in 20211 and could exceed $150 billion by 2026.2
Tasks enhanced by AI range from medical imaging for diagnosis to genetic analysis to aid in treatment. Yet barriers remain, including challenges with data privacy, AI model bias, and a lack of clinical studies. Regardless, AI in medicine offers transformative potential both for the healthcare industry and for life and health insurers.
Modeled on the human brain’s network of neurons, a convolutional neural network (CNN) is a type of deep learning AI algorithm ideally suited for image recognition and task processing. CNNs receive raw image data, usually as a grid of pixel values with three color channels (red, green, blue). After the input layer, these networks employ convolutional layers to detect patterns and features in the data. When connected in multiple layers, these neurons are referred to as a deep learning (DL) network. CNNs have been used in many models for cancer detection because of their ability to extract special and contextual information from images.3
Radiomics, a type of DL, is a new field of research based on radiological imaging analysis and extraction of many quantitative features from medical images. AI algorithms leverage radiomic features noted on mammography, ultrasound, MRI, and positron emission tomography (PET) to enhance accuracy in detecting and classifying malignant and benign lesions. This process captures tissue and lesion characteristics, such as intensity, size, shape, volume, texture, and dimensions often not noticed by the human eye.4
AI can assess medical images faster than humans, evaluate patients at risk of early mortality, reduce the workload of physicians, decrease medical errors, and improve the accuracy of disease diagnostics and efficacy of treatments, directly impacting morbidity and mortality rates.
Using AI algorithms to detect disease and personalize treatments can improve survival rates and life expectancy. For example, compared to human intelligence, AI can more easily prioritize screening for asymptomatic patients at risk of cancer and better identify patients at risk of cancer recurrence.
The process of developing new pharmaceuticals can take more than a decade and cost billions of dollars. Even then, only about 10% of drugs in Phase II trials receive approval for further development. AI models can now predict outcomes from clinical drug trials, significantly expediting the process and lowering the enormous cost of drug development. AI algorithms look for similarities in drug compounds and can project levels of toxicity in patients and any probable drug-target interactions. AI software developed to improve patient participation in drug trials, where drop-out rates lead to the failure of 30% of trials, has increased adherence rates by up to 25%.5
AI can identify patients at minimal risk of disease. With cancer, for example, this means minimal-risk patients covered by recommended cancer screening guidelines may no longer require screening, helping to improve turnaround times and reduce costs. Furthermore, personalized risk predictions could enable those most at risk who do not fall under current screening guidelines to be screened early, facilitating earlier diagnosis and treatment of diseases that would otherwise remain undetected. This would increase survival rates for a myriad of diseases, particularly those conditions with poor prognoses.
With cardiac screening, the AI-read ECG is proving comparable, if not superior, in many cases to human interpretation of ECGs and has the potential to revolutionize clinical treatment. For assessing cardiac function, cardiac magnetic resonance (CMR) imaging is considered the gold standard. Shortages of qualified CMR-trained doctors can result in delays in screening for, and diagnosis of, cardiovascular disease (CVD). In a recent study of 11 types of CVD in 9,719 patients, an AI deep learning model achieved performance comparable to that of physicians with more than 10 years of experience in CMR reading (F1* score of AUC** 0.931 versus 0.927), with faster speed of interpretation (1.94 min versus 418 min for interpreting 500 subjects).6
* F1 = harmonic mean of the predictive positive value and sensitivity. ** Area Under the Curve (AUC) is a measure of model performance accuracy, where 1 is a perfect classification, and 0.5 is random chance.
AI can play a key role in disease diagnosis, assisting doctors with medical interpretation of complex data and scans, allowing them to review more patient data in less time. AI aids in the identification of lesions on scans, classifying individual pixels and recognizing patterns undetectable to the human eye.7 For example, AI algorithms can help radiologists read mammograms and computed tomography (CT) scans and provide a second opinion on their findings. Agreement can range from 75% to 88%, particularly for more complex diseases.8 AI can also detect abnormalities with comparable or improved accuracy vs. experienced radiologists. In a study using CNNs to predict malignancy on CT scans, the AI model was more accurate than the average radiologist reading, achieving an AUC of 95.5%.3
AI can identify gene combinations in patients, helping to predict treatment response to specific drugs and allow for doses to be adjusted or modified.7 In addition, AI can assist with surgical planning and intervention by creating three-dimensional (3D) models detailing a patient’s specific anatomy. This allows surgeons to complete operations with a high degree of accuracy, improving medical outcomes and reducing patient risk. 3D models can also be printed as physical objects, allowing medical specialists to replicate and examine a specific part of a patient’s anatomy.9
Significant barriers with AI in medicine have yet to be overcome. These include data protection concerns about how personal medical information is used, stored, and distributed within AI models. Any data used in training models must be properly de-identified and standardized. Also of great concern is malicious misuse of AI, including fabricated medical research or the use of inauthentic data. In an analysis of references used in a medical article created by a ChatGPT platform, only 7% were authentic and accurately analyzed the information.10
AI modelling is currently limited by a certain amount of bias, swayed by gender, ethnicity, and the use of datasets from high-income countries. This could lead to inappropriate decisions for patients in low-income countries. Many AI models are limited in demographic diversity, making it hard to translate models into clinical practice.
The lack of large clinical studies to validate AI modelling outcomes is also an issue. A review of 172 AI solutions showed that 93% of them fell below clinical readiness stage 4 in real-world applications. Only 2% had undergone prospective validation. Stage 4 should demonstrate the potential for prediction or decision support or optimizing model and validation on characterized data.11 Models will need to show consistency in their results across multiple datasets before they can be integrated in real-world settings.
Another significant concern involves blame when things go wrong. Should AI algorithms provide doctors with an incorrect diagnosis or treatment plan, it will be difficult to hold any person accountable.
AI-based prediction tools can make erroneous decisions, as can humans, but there are currently no regulations or guidelines regarding who is legally responsible when AI makes an incorrect decision that harms a patient.
Further limitations and challenges include the substantial initial costs of the AI applications and their associated maintenance that could ultimately prove prohibitive, as well as the impact of resource-intensive AI on the environment.
AI is already being used in diverse ways within the insurance industry, from fraud detection and client verification to chatbots in customer support applications. AI can identify risks in underwriting applications and pricing models and can aid decision-making in claims submissions. It can help streamline processes, analyze data, and improve efficiency across the business. It is particularly useful in processing tasks and interpreting complex medical data, both of which are a major part of the life insurance process.
AI could speed up underwriting by limiting the need for additional tests based on risk model outcomes. It could identify likely comorbidities and risk of disease development based on an applicant’s responses on the application form. Insurance companies could see significant cost reductions, from reduced claims payouts for death and critical illness claims to more efficient administration tasks. According to Boston Consulting Group, AI has the potential to reduce claims payouts by 3% to 4%.12 A separate report on AI trends in the insurance market stated AI could lower claims processing costs by 50% to 65% and processing times by 50% to 90%.13
AI has the potential to change the landscape of modern medicine, improving disease detection and treatment efficacy. AI is likely to benefit many patients and improve overall health outcomes, particularly for individuals diagnosed with cancer. However, due to current barriers and limitations, it may be some time before AI predictive models replace healthcare experts and are routinely integrated into preventive screening, patient diagnostics, and personalized treatments.
RGA experts are eager to engage with clients to better understand and tackle the industry’s most pressing challenges together. Contact us to discuss and to learn more about RGA's capabilities, resources, and solutions.
