Trends and Challenges of AI: Conclusions from THAIS

On June 11 and 12, the Alma Medical Imaging team attended the third Tauli Health Artificial Intelligence Symposium (THAIS), organized by Unidad Mixta del grupo de Nefrología Clínica, Intervencionista y Computacional (CICN), from Instituto de Investigación e Innovación del Parc Taulí (I3PT), and School of Engineering of the Universidad Autónoma de Barcelona.

Trends and Challenges of AI: Conclusions from THAIS
- - - - Access to Data, Key for AI Adoption in Medicine
      - Interpretability of Models, a Challenge to Overcome
        
        Heat Map
        
        SHAP Method (SHapley Additive exPlanations)
        
        NEAR Method
        
        Artificial Intelligence in Spain
        
        Alma Medical Imaging and the Central Role of AI

Led by Dr. José Ibeas, THAIS took place at the Parc Taulí Hospital in Sabadell, providing insights into the challenges and trends of AI in clinical practice. Undoubtedly, it was an enriching experience from which we have drawn valuable conclusions for Alma’s activities, summarized below.

Access to Data, Key for AI Adoption in Medicine

Access to data was one of the fundamental points reviewed during the symposium, as it is essential for adopting artificial intelligence in clinical practice.

Specifically, we discussed two types of data usage:

Primary or traditional use, for diagnostics.
Secondary use, related to research and model training.

Various models were discussed during the sessions to ensure the secondary use of information: debates included opt-in and opt-out approaches, the possibility of giving credit in scientific works to physicians who provide data to enhance motivation, and the European Health Data Space initiative, which aims to promote the sharing of medical data within the EU through the following lines of work:

Empower individuals to control their health data and facilitate its exchange for healthcare provision in the European Union
Promote a genuine single market for electronic health record systems.
Offer a consistent, reliable, and efficient system for the reuse of health data in research, innovation, policymaking, and regulation.

Interpretability of Models, a Challenge to Overcome

The interpretability of models is another critical issue that needs special attention to promote AI adoption in clinical practice.

In some models, the output is derived from nonlinear combinations of inputs, sometimes with multiple hidden layers, making interpretability challenging and consequently reducing the trust that physicians place in this technology. To overcome this barrier, several methods were discussed at THAIS:

Heat Map

AI-based automatic medical image analysis uses complex convolutional networks that are not easily interpretable at first glance. One proposed option to improve interpretability is to show a heat map on the original image that identifies the areas given the most weight.

SHAP Method (SHapley Additive exPlanations)

The SHAP method calculates the impact of different input variables on the model output, distinguishing which types of values lead to a particular output. This method can be combined with the XGBoost algorithm, a type of random forest where the decision trees are sequentially linked to improve the parts that fail in the previous tree.

NEAR Method

The NEAR method uses the SHAP method to find the variables with the most weight in the output and then constructs a model using only these variables, which has been shown to be significantly more explainable than the original.

Artificial Intelligence in Spain

After reviewing these topics, it is worth asking where Spain stands in this sector. Here are some AI research lines revealed at the symposium:

The Clinical, Interventional, and Computational Nephrology Group at Parc Taulí Hospital, led by Dr. José Ibeas, is working on massive data analysis to predict heart failure after kidney failure, among other things.
The Barcelona Supercomputing Center (BSC) is creating a massive language model from scratch that can be trained and validated in various sectors, including healthcare.
Federated or collaborative learning is another interesting line, where an algorithm is trained through a decentralized architecture comprising multiple devices, each with its own local and private data. This approach, developed by Google in 2017, differs from centralized machine learning and other classic decentralized approaches in that each device has its own local and private data instead of transferring it to a server as it is stored. In this way, the result of multiple decentralized models is centralized into a single global model, always preserving the integrity of the information.

Thus, research is on the right track to incorporate artificial intelligence into medicine. The challenge? Creating business models that allow for gradual and evaluable integration of these new technologies.

Alma Medical Imaging and the Central Role of AI

At Alma, we pride ourselves on the fact that artificial intelligence has become a central and strategic part of our solution, which is why training like THAIS enriches us as a team.

Here are some of our lines of work to improve AI implementation in the radiological workflow:

Developing AI algorithms to add new functionalities and enhance the current features of our DICOM viewers.
Integrating and implementing algorithms in the clinical practice of radiologists to optimize workflow.
Designing and developing custom AI algorithms through services to third parties.
Participating in national and European projects and consortia that drive research and development of innovations impacting the field of medical imaging.
Providing strategic support to professionals and centers for the incorporation of AI in healthcare digitalization processes.