Supported by its ability to synthesize large amounts of data, transforming them into fast and efficient tools, Artificial Intelligence (AI) is permeating many areas of modern medicine. These tools are born with the aim of supporting healthcare professionals in their daily tasks and contributing to the equity of the healthcare system. This is evident in the contributions of AI to precision medicine, where it is bringing us closer to the dream of personalized treatments based on the patient's genetic profile.

However, among the various problems limiting these benefits, the problem of the interpretability of AI results stands out above all others. This problem is very simple to explain, but not so easy to solve. By interpretability, we understand the ease with which a practitioner can understand and assess the results of an AI program, i.e., the 'reasons' for its decision/prediction. Generally, the predictions of AI programs are very difficult to comprehend ('interpret'), which impedes the routine use of these programs in the medical environment, due to the possibility of undetectable errors, but with serious medical consequences. To avoid such a possibility, the international community is promoting the development of new AI methodologies that allow better assessment and understanding of their results.

In this context is framed the project recently awarded to the Clinical and Translational Bioinformatics group of the Vall d'Hebron Research Institute (VHIR), led by Dr. Xavier de la Cruz, in the call "Proyectos de Transición Ecológica y Transición Digital" of the Ministry of Science and Innovation with European funds from the Recovery, Transformation and Resilience Plan. The amount granted, €219,650 over two years, will allow Dr. Xavier de la Cruz's group to develop a new technology, based on the most recent advances in AI, which will allow reformulating the current tools for the identification of pathogenic variants.

The general scope of this project is the early diagnosis of hereditary diseases or diseases with a genetic component, such as breast cancer, based on the sequencing of the genome of patients. Such sequencing makes it possible to identify genetic variants whose pathogenic risk can be established by AI applications. In this context, the aim of the project is to design a new risk prediction system that is accompanied by graphical representations that facilitate the interpretation of the results by health professionals. “It is important to be able to have information to be able to judge the results and assess how reliable the predictions that artificial intelligence gives us are. Currently, this cannot be done as the available models are considered "black boxes"”, explains Dr. Xavier de la Cruz.

As mentioned above, this project is part of the movement that has emerged in the last four years in favor of the interpretability of AI tools so that users have the ability to judge their results. "In this way, clinicians can easily and quickly assess predictions and, from here, make the most appropriate medical decisions for patients", concludes Dr. Xavier de la Cruz.

A first sample of their results (generated by Dr. Luz Marina Porras and Dr. Natàlia Padilla) is available here, where the user can access the prototypes of the new graphic representations. These results, developed within the Spanish consortium for the interpretation of ATM gene variants associated with breast, prostate, pancreatic cancer, etc., have recently been presented to the scientific-clinical community for evaluation.