Artificial intelligence (AI) is revolutionising our world, including the medical sector. It offers innovative tools for data analysis, disease diagnosis, patient management, and drug discovery. One of the first successes of AI has been an incredible improvement in diagnostic accuracy through analysis of medical images using deep learning networks. Their accuracy in the dermatological and radiological fields is comparable to that of the best medical experts (Esteva et al., 2017; McKinney et al., 2020).
The triumphs of AI are not limited to diagnostic imaging. It is also having an impact on research. For example, over 50 years ago, Anfinsen won the Nobel Prize for demonstrating that protein sequences determine their three-dimensional structure. This problem had remained unsolved for over half a century despite the combined efforts of thousands of scientists worldwide. DeepMind’s AlphaFold AI system solved this problem with unprecedented accuracy (Jumper et al., 2021) using an unique tool designed to advance the study of genetic mutations and their disease implications, a tool that is also being used to evaluate new drugs.
AI is accelerating drug discovery through virtual screening techniques that simulate and model molecular interactions to predict the efficacy of new compounds. For example, deep neural networks can be used to predict the affinity between ligands and protein targets. One emblematic case is Insilico Medicine’s use of AI to identify a new compound for pulmonary fibrosis. Thanks to AI, a process that normally takes many years was completed in just a few days (Zhavoronkov et al., 2019).
During the COVID-19 pandemic, AI platforms were used to analyse large volumes of genomic data pertaining to the virus and to design effective vaccine candidates in record time. For example, Moderna used AI to optimise the design of its mRNA-1273 vaccine, which contributed to the speed of its development and approval (Sharma et al., 2022).
New foundational tools for language analysis (like ChatGPT) enable analysis of clinical texts in order to extract useful information from clinical reports and scientific publications. These models allow an expert user to analyse medical records and suggest personalised treatment plans. Medical chatbots can provide patient support. Apps for analysing medical images can provide preliminary diagnoses by also assessing patients’ symptoms, and they can produce recommendations based on medical guidelines.
Despite these advances, there are serious obstacles and problems involved with applying AI in general and particularly for medical purposes. One of the main issues is the quantity and quality of the data. To train algorithms effectively, it is necessary to have high-quality, well-annotated data. Very often, medical data are kept (or scattered) in various departments of hospital or research institutes which are not connected to one another and so there is no possibility of retrieving the relevant data. Think about the volume of medical images (X-rays, MRI, ultrasound, dermatoscope, etc.) and clinical data (blood tests, urine tests, therapies, etc.) we all produce in our lifetimes that cannot be collected systematically. What’s more, everything is complicated by regulations and privacy issues that are making data exchange and integration more rigid.
andSharing information between multiple centres is essential for effective medical research, but there are significant challenges and difficulties involved. Paradoxically, private entities with clear internal process management for data processing can use AI more easily and directly than public research institutions. In this context, federated learning and the development of synthetic data appear to provide promising solutions (Sharma Guleria, 2023). Federated learning involves the use of algorithms that “move” between various centres without transferring data. Privacy is maintained as they collect only the information required for building prognosis and prediction models. Synthetic data are “digital twins” of patients that are not real but act as surrogates for real patients, allowing AI to learn from the data as if it were real data. In many cases, synthetic patients are generated by the AI models themselves and built independently at each centre.
There is a further critical issue with regard to AI, and it is relevant also in the medical field —possible “bias” in the data. For example, biases related to gender, race, and unbalanced source populations can compromise the accuracy and equity of diagnoses. One example of this is an algorithm used in the United States for managing chronic patients which proved to be less accurate for black patients due to bias in the training data (Obermeyer et al., 2019).
Unfortunately, the data available (not only medical) are often based on white and male populations. This can create problems for future personalised medicine, so it is important to know what databases have been used for the models to be employed in the medical field. Doctors should treat these tools with the same caution they apply when making decisions about surgery or chemotherapy treatments.
Finally, public research, including medical research, faces a major challenge. Currently, certain types of AI “experiments” can only be performed by large private companies. No single research centre can compete with their hardware resources. This is unprecedented in research, and it creates fundamental problems for checking on what is being developed, produced and “sold”. One possible solution could be the creation of common European or international resources (like CERN in the physics field). This would prevent public AI research from being limited to theory or marginal applications.
References
- Esteva A. et al. (2017). Dermatologist-level classification of skin cancer with deep neural networks. Nature, 542(7639), 115.
- McKinney S. M. et al. (2020). International evaluation of an AI system for breast cancer screening. Nature, 577(7788), 89.
- Jumper J. et al. (2021). Highly accurate protein structure prediction with AlphaFold. Nature, 596(7873), 583.
- Zhavoronkov A. et al. (2019). Deep learning enables rapid identification of potent DDR1 kinase inhibitors. Nature Biotechnology, 37(9), 1038.
- Sharma A. et al. (2022). Artificial Intelligence-Based Data-Driven Strategy to Accelerate Research Development and Clinical Trials of COVID Vaccine. Biomed Res, 6, 7205241.
- Sharma S., Guleria K. (2023). A comprehensive review on federated learning based models for healthcare applications. Artif Intell Med., 146, 102691.
- Obermeyer Z. et al. (2019). Dissecting racial bias in an algorithm used to manage the health of populations. Science, 366(6464), 447.
Image: Photo by the (US) National Cancer Institute on Unsplash
