João Henriques, Machine Learning Researcher at Adentis, reflects on the potential of artificial intelligence, namely in the progress of medicine. Read below, the article published in Sapo TEK.
When properly used, AI can help and facilitate our lives to levels never reached before. AI is one of the areas that in the future will increase the production and efficiency of our factories, improve our services, assist and increase medical screening, etc. So AI came to replace us? Yes and no. Yes, it will replace humans in many routine tasks, it will perform them much more efficiently and at a much lower cost. No, because we will continue to need people capable of playing roles that require our decision-making skills. For example, agriculture did not stop because tractors appeared, nor the manufacturing process, because there are robots. It's just the natural evolution of society and technology. An important point that is overlooked when talking about AI is that we will only use smart technology when it is sufficiently studied and safe. Regardless of the potential of the application of AI systems in medicine, for example, we will only see applications to be used with more investment if some challenges are overcome.
Building a multidisciplinary research community
The importance of building multidisciplinary communities is crucial, as each member is specialized in one area and is the best person to identify problems and assist in the development process. By bringing together data scientists, systems developers and dermatology experts, for example, we could create a system that helps dermatologists identify critical signs. But why is this still not happening in practice? Currently, much of the research is focused on developing methods rather than implementing these methods in clinical practice. Research centers and companies must come together to find effective ways of understanding uncertainty and validating AI approaches, educating users about the technology's strengths and weaknesses, and rigorously evaluating its benefits in terms of outcomes, user experience, patient, and costs. These systems are 100% dependent on data that exists but is not yet available to everyone.
Access to quality data
Access to data is in itself a major difficulty, as data is a valuable intangible asset, but a large, well-annotated, quality, and publicly available dataset is clearly a significant barrier to research and development of algorithms. AI Having data stored in databases without anyone being able to work with them is like having money in the bank without being invested – in the long run we will be losing it. Data quality has the power to make algorithms more accurate and reduce reliance on large amounts of annotated data, which may not be available.
Understanding what the machine sees
One of the great challenges of AI, specifically of Deep Learning (DL), is to be a “black box”, that is: we do not know which variables the algorithm used in its decision-making process. Therefore, we need to carefully consider how biases affect the data being used to develop models. Making the process transparent for scientists, clinicians and patients is achieved by showing how algorithms arrive at their decisions.
What can companies do?
Companies can invest more in research and development projects focused on artificial intelligence in partnership with academic research institutes and centers. The objective of the companies is to create products and services to generate profits and the one of the institutes and centers is the scientific discovery that originates knowledge, articles, theses. Fortunately, both objectives do not conflict, on the contrary, they are an excellent synergy. At the company where I work, we started with a project with the objective of developing an algorithm capable of identifying the degree of blindness in diabetic patients. Diabetic retinopathy can be irreversible, so it is necessary to monitor it so that it does not get worse. Our objective was to develop a system capable of assisting physicians in the diagnosis of this disease. Publicizing the projects in which we participate to our partners also brought us added value, as we learned, for example, that dogs and cats suffer from the same pathology. This information is very valuable, as we can extend the knowledge acquired to the world of veterinary medicine as well.
The success of the applicability of AI models depends on investing in data quality, in the communication and sharing of data and information, in the implementation of practices for the detection and monitoring of biases, and also in the measurement and monitoring of the performance of these applications in the real world.