Human and artificial intelligence true drivers of new medicine
What are synthetic data produced by specific generative artificial intelligence algorithms and how are they used? Their study and use is of paramount importance in health care, as they are not subject to privacy restrictions and, most importantly, make it possible to overcome the problem of scarcity of clinical data. But not only that, they foster a synergy between research and clinical practice in order to maximize positive effects in the adoption of innovative tools to improve patient care.
Patrizia Fortunato, consultant content editor of Forum PA Sanità, discussed this with our researcher Giuseppe Jurman, head of the Data Science for Health unit of at Fondazione Bruno Kessler’s Digital Health and Wellbeing Center.
Jurman, interviewed ahead of the upcoming FORUM Sanità 2023, will participate as a speaker in the “Health Data & FSE” scenario on October 26 from 09:30 a.m. to 11:00 a.m. in Rome.
Data created by particular generative artificial intelligence algorithms (an example is ChatGPT, i.e., all those procedures that can generate images, sounds, from even literary descriptions) are called synthetic data. They do not belong to any patient and can relate to any aspect of diagnostics or prognostics: laboratory tests, images provided by CT or PET scans, genetics data, genome sequencing data.
How critical can their use be in medicine?
The importance of synthetic data in healthcare
“There are two aspects to take into account: the first is data scarcity; the second has to do with privacy,” says Giuseppe Jurman.
The fragmentation of data available for many diseases due to the small size of a given hospital, the rarity of the disease, or the high cost in producing such data result in what is called data poverty in healthcare.
“The use of synthetic data,” Giuseppe Jurman says, “can help increase the amount of data available to enable a device to learn and improve its predictive, prognostic or diagnostic capabilities with respect to a specific clinical task under study.”
In the statistical sense, drawing robust conclusions from evaluating too small samples carries risks (the very basis of probability theory is the principle of “sampling variability”). Similarly, data sparsity affects the performance of Machine Learning models.
Synthetic data preserve privacy. “Because they are not linked to any actual patient, they are not subject to restrictions arising from the GDPR, nor do they need ethics committees, permissions or authorization from the patient,” as Jurman puts it. They are data used by pure research. Thus, they do not carry ethical implications; rather, they expose clinical caveats, warnings of danger to prompt the unaware to take into account the relative risk and acknowledge their reliability.
The problems and challenges associated with the use of synthetic data in health care
We are just starting to use them; their potential is still being understood. There are numerous sensitive issues related to synthetic data, in the sense that synthetic data is valid if it is realistic. “The main problem in using synthetic data in health care,” Jurman says, “is ensuring that the algorithm used has produced a data that realistically belongs to the distribution of the real data we are studying. The functionality of the data comes down if it cannot belong to a real patient, simply because biological motivations prevent that data from being generated in that way.”
This is the biggest obstacle facing research, and this is true for any type of data. iuseppe Jurman points out that there are types of data that are easier to create: “the first data that have ever been reproduced, from a synthetic point of view, are bioimaging data, such as MRI data, biopsies in pathological anatomy, and this is because images are somewhat more controllable during generation; it is much more difficult, on the other hand, to generate clinical data, especially longitudinal data, which describe the health trajectory of a patient over the years. If you take a patient and monitor them over time, continuing to have clinical examinations, reliably and realistically replicating that kind of data becomes much more difficult.” This is one of the open challenges that research faces.
Fostering the adoption of innovative solutions in healthcare: the key role of Fondazione Bruno Kessler
The awareness that from research one must move on to the impact phase, to spillover into clinical practice, must always guide any kind of activity in this area. This is one of the principles that inspires Fondazione Bruno Kessler, which has at its core the Digital Health and Wellbeing Center, a center based on two pillars that support the different activities: research excellence on the one hand and impact on the health system on the other hand.
When initiating a research project with the clinician, it is important to consider that “the clinician wants to be able to see the effects at the clinical practice level in a reasonable time and with an appropriate level of reliability,” Jurman says.
So when Fondazione Bruno Kessler initiates projects with health care units, whether local or national, there is always the problem of getting a very good result from the research point of view, but also moving quickly in the direction of certification, validation, and implementation of translational medicine, i.e., taking this product from the laboratory to clinical practice. In this, the Foundation needs the clinician’s guidance and the help of specific guidelines for certification of the research algorithm toward the clinical product.
“It is well known in the literature,” Jurman says, “that the performance results you achieve in research are unlikely to be replicated to the same level in clinical practice because of numerous problems related to data, implementation, and the fact that the data you have trained your machine on does not exactly match what you are confronted with in everyday practice. Data in practice is more impure, less clean.”
“If you’re studying a disease about which you know very little, any tool you can use in research is welcome; only then,” Jurman continues, “do we deal with implementation. We try to understand, first of all, the biological mechanisms we are dealing with, while in so many activities, more properly routine, artificial intelligence has to act as a support to the daily work. In these cases, the impact becomes perhaps more important, and therefore it is necessary to have a good-performance tool that is certified and validated in a reasonable timeframe.”
Clinicians and research: an essential combination
Research has always been collaborative and is even more so when we start introducing innovative tools for clinicians as well, and “clinicians may not always be inclined to accept them,” Jurman notes.
“We come from two different worlds and with different languages. The new revolution imposed by artificial intelligence and Data Science imposes coming closer together,”Jurman says. The medical profession is changing, and elements must be introduced to help physicians understand in the best possible way what is being used and how it can be integrated into their work. From our perspective, as data scientists, it is critical to equip ourselves with the skills to understand as much as possible the clinicians’ point of view on the issues. So there is an attempt being made to converge the two areas, and this is one of the principles that the scientific literature is also following.”
In fact, Eric Topol, one of the leading experts in this field, published an article in Nature Medicine in 2019 entitled “High-performance medicine: the convergence of human and artificial intelligence,” highlighting how the result of the convergence of the two different intelligences, human and artificial, will be the real driving force behind the new medicine. Personalized, precision medicine will necessarily be the result of the convergence of the two strands of human and artificial expertise.
Speaker at the Health FORUM 2023 event: Giuseppe Jurman
Oct. 26 from 09:30 a.m. to 11:00 a.m. – LVG Plaza Room, Rome
Reserve your seat “Health Data & FSE“ (THE HUB LVenture GROUP – Via Marsala 29 H, Rome, on Oct. 26).
Share on Facebook Share on Twitter Share on Pinterest
