TOP
The topic of artificial intelligence (AI) has gained increasing attention in the medical scientific world over the last years and there is a lot of excitement about the potential future role of AI in this field. The number of scientific papers on this topic has tripled within the last years and typing “AI in medicine” into the search field of medical publications platform nowadays brings more than 23,000 results. On top of that, increasing numbers of AI‑based devices are currently U.S. Food and Drug Administration-approved, mostly in the field of radiology which currently stands out as the most AI‑invested medical specialty, followed by cardiology.1
AI encompasses a wide set of methods for replicating human-like abilities – such as perception, reasoning, and decision-making – using rules, statistics, or learning-based approaches. Machine learning arose as an AI technique focused on learning patterns from data. Deep learning is a specialized form of machine learning using multilayer neural networks, and generative AI is an advanced offshoot that creates new content using these models.
Machine learning is an AI‑based technique that is often applied in the medical field. Machine learning starts with a large amount of data on which data machine-learning algorithms are employed. Algorithms use statistical techniques to detect patterns and make predictions or decisions based on historical data that they are trained on. The information is used to generate a useful output to solve a well-defined problem in the medical system followed by an action or a categorization.
Machine learning is very promising in medical fields where making of diagnoses is based on images, e.g., histopathology. AI can be used to assist pathologists in making more accurate, quicker and higher number of diagnoses. It includes the goal of reducing errors in making diagnoses, e.g., cancer diagnoses. Systematic reviews and meta-analyses including diagnostic accuracy studies from around the world using AI applied to whole slide images for any disease, report a mean sensitivity of 96.3% and a mean specificity of 93.3% for making any type of diagnosis. With the assistance of deep-learning systems, the accuracy of pathologists’ diagnoses improved significantly, representing an 85% reduction of human error rate.2
AI‑enabled devices are also used in the field of radiology, which itself is a very digital, computer image-based science. AI in medical imaging uses algorithms to improve the speed, accuracy, and efficiency of analyzing images such as X‑rays, MRIs, and CT scans, assisting in disease diagnosis, treatment planning, and monitoring. Benefits include faster analysis and the detection of subtle anomalies that might be missed by the human eye, leading to more precise diagnoses and potentially better patient outcomes.3
But AI in medicine it is not only about better image recognition. AI has been shown to be efficient at solving problems and predicting medical conditions. This ability is welcomed in clinical medicine because the efficacy of treatment and diagnosis increases with the prediction of specific situations and conditions as it gives doctors the opportunity to treat patients in a timely manner.
For example, an important prediction in the clinical field is the early prediction of an underlying sepsis. Sepsis is a leading cause of death, accounting for half of all hospital deaths. It occurs when an infection triggers a chain reaction throughout the body. Inflammation can lead to blood clots and leaking blood vessels and can ultimately cause organ damage or organ failure. About 3.4 million people develop sepsis every year in Europe and nearly 700,000 of them die. About one-third of sepsis survivors endure lifelong physical, mental, or cognitive problems.4
The challenge is that sepsis is easy to miss because symptoms such as fever and confusion are unspecific and common in other conditions. On top comes that if the patient develops symptoms, clinicians are already lagging behind treatment and racing against time. One of the most effective ways of improving sepsis outcomes is early detection, and applying efficient treatments in a timely way, but historically this has been a difficult challenge due to lack of systems for accurate early identification of sepsis in patients.
To address this challenging problem doctors and researchers have developed AI‑based targeted real-time early warning systems to identify sepsis patients earlier. The systems catch symptoms hours earlier than traditional methods by scoring structured data and unstructured data in the medical records and clinical notes to identify patients at risk for life-threatening complications. By combining a patient’s medical history with symptoms, examination results such as lab results, measurements of vital signs, etc., the machine-learning system shows clinicians when someone is at risk for sepsis.5
The potential of AI creates many opportunities in clinical medicine with great visions for the future in terms of getting a clearer overview by combining all data and information, putting the puzzle pieces together to obtain “the whole picture”.
AI may use data coming from ECGs, lab values, imaging procedures and many more medical examinations, all included in the clinical records, which are increasingly available in digital form. Data from different wearable sensors such as smart watches may be added. The patient-specific information may be combined with available medical evidence on large biological datasets, such as genomics, proteomics or metabolomics. Information on a patient’s lifestyle, such as nutrition or physical activity, and many more data could be included on top. AI‑based computational models create the potential to integrate all information and to make a decision that takes all the data into account.
AI could also be used to predict the outcome of a treatment applied to a specific diagnosis and to find the best treatment for the individual patient. With AI it might become possible to predict drug responses and to optimize dosing. It may help clinicians in chronic diseases management with the opportunity to reduce hospital readmissions and time of inpatient treatment.
AI could support researchers in the development of new drugs or in repurposing existing drugs by finding out about the full potential of already existing medication. As an example, the group of glucagon-like peptide‑1 receptor (GLP‑1 R) analogues are generally perceived as a new substance group and are mostly known as “the new weight loss medication”. But the first GLP1 analogue was released in 2004, for the treatment of type 2 diabetes. It took 10 years in the medical field to approve it as a drug for long-term weight control in non-diabetic, obese people and another eight years to approve it as a cardioprotective drug. AI may accelerate this process of faster identification of drug potential by faster molecule screening.
Another potential benefit may be the prediction of treatment efficacy. Studies are currently looking at some existing medications, such as immune checkpoint inhibitors (ICIs). ICI are a type of immunotherapy and are used in treatment of different types of cancer, showing very positive effects on treatment outcomes. However, only some patients respond to ICIs, while others don’t. It currently remains unclear what causes this difference in responsiveness or how it may be predicted which patient is going to respond positively to treatment.
Non-efficient treatment is expensive and, above all, a very frustrating experience for the patient who might suffer from severe side-effects on top. A Korean study revealed that a machine learning-based tool was able to identify key communication pathways with crucial players responsible for the patient’s immune response or resistance to ICIs.6
AI may also bring doctors forward in performing precision medicine in invasive treatment. AI‑based devices such as in robot-assisted surgery are already broadly applied in the clinical field. It is likely that this technique will spread more and become broadly available for multiple surgeries and invasive types of treatment. The currently available techniques will also continue improving.
Robot-assisted surgery means less-invasive procedures with fewer complications, less blood loss and faster recovery, leading to a reduction of inpatient treatment and intensive care treatment.
AI has the potential to revolutionize medicine, but there are many challenges to overcome. The quality of the data used to train the AI is a major concern. AI depends on reliable, transparent and peer-reviewed data to produce useful information that can be transferred to a general population. One has also to deal with variability of machines used in the clinical field to collect all the medical data. This problem of non-comparability is a major concern that needs to be addressed in the development of reliable AI‑based models.
Access to large datasets for training of models delivering meaningful results, remains essential, posing a huge data protection issue. There are problems concerning data ownership, lawful collection and the information of probands, and there is risk of system hacking, abuse of heath information or criminal or political misuse. On top of that comes the question of legal responsibilities if a treatment fails or a diagnosis was wrong.
In summary, the future of AI in healthcare looks promising. Much remains to be done, but medical research has never gone backwards in human history, and it is very unlikely that this is about to change.
How will this affect the Life insurance industry? It is possible that formerly devasting diseases may turn into more chronic conditions. Chronic and acute diseases are likely to be detected earlier, monitored more closely and treated more efficiently. Life insurance can expect better diagnosed applicants and insureds with clearer long-term prognoses, being healthier due to earlier diagnosis, better treatment and earlier risk stratification in the clinical field.
