AI & Health Winter School: abstracts

Annette ten Teije (VU, Computer Science, Learning and Reasoning group)

In recent years there has been a strong increase in systems that combine data-driven machine learning with knowledge-driven reasoning. We have proposed a set of compositional design patterns to describe a wide variety of systems that combine statistical techniques from machine learning with symbolic techniques from knowledge representation. These design patterns help to understand the rapidly growing literature for such neuro-symbolic systems. They can also be used as guidelines for the design of such systems. In this talk we will introduce a notation for these design patterns, and illustrate these patterns with neuro-symbolic systems in the medical domain. We will dive into a guideline-informed reinforcement learning method for mechanical ventilation in critical care along the lines of these patterns.

The Next Frontier: Medical Foundation Models

Marike van der Leeden (Amsterdam UMC, Rehabilitation Medicine)

Automated recognition of (recovery in) functioning of COVID-19 patients with natural language processing

COVID-19 may result in impairments in functioning which may require rehabilitation. However, assessment of functioning based on the International Classification of Functioning, Disability and Health (ICF) is not easily obtained.  We tested the feasibility, reliability and internal validity of automated assessment of functioning of key ICF categories and their levels of hospitalised COVID-19 patients. The assessment is performed by applying state-of-the-art natural language processing (NLP) methods on unstructured text in electronic health records (EHR) from a large academic hospital.

In this talk, I will discuss the methods used to develop these NLP algorithms and what the next steps will be to use the algorithms in rehabilitation care and research. I will also share lessons learned to carry out such a project, in terms of technology, data and human factors.

Mohammad Rezazade Mehrizi (VU, Centre for Digital Innovation) & Willem Grootjans (Leiden UMC)

End-user adoption of AI in clinical practice: A socio-technical perspective

Like many other technologies, AI solutions witness a complex journey until they are actually used by the end users in organisations. Especially in the medical context, this journey is unpredictable and involves socio-technical factors beyond only the intentions and decisions of individual users. Metaphorically, the introduction of AI into clinical practice is like planting a seed into a jungle (not even a protected green house!). The eventual adoption and use of such a technology requires a wide range of conditions coming together to naturally nurture an effective implementation and use. In this session, we unpack such a process through a socio-technical change perspective. We learn how we can design, act, and assess the adoption of AI solutions through systematic and dynamic arrangements among actors, technologies, structures, and tasks. In the second part, we collectively discuss a real-life use-case of adopting an AI solution in radiology at the LUMC. The aim is to apply our learning into a practical example and learn how to from a critical and practical view on the adoption of AI solutions in clinical practice.

Matthan Caan (Amsterdam UMC, Biomedical Engineering & Physics)

What no eye has seen: how AI is revolutionizing medical imaging

For long time, human sense was the only diagnostic measure available in healthcare. Today, along the chain of clinical decision making and treatment planning and monitoring, medical imaging plays a pivotal role. When should these images be acquired and how should they be evaluated? How can we make Magnetic Resonance Imaging faster through the use of AI? This is for instance relevant in the context of stroke treatment. We will explore the integrated view on data analysis through so-called end-to-end deep learning for accurate delineation of white matter lesions in the brain. Major depression treatment outcome prediction on multimodal data can be realized and potentially provide a first treatment decision making tool to psychiatrists.

Organized challenges on a specific medical imaging task are an attractive and energizing way of competing and collaborating, with high impact. This holds for research and education.

Finally, we will reflect on involving and informing the general public about our work by presenting in the press and regular media.

Willem Jespers (Leiden University, Leiden Academic Centre for Drug Research)

Computational Methods in Drug Discovery: AI and Physics Based Approaches

Pharmaceutical science is changing; while perhaps not a paradigm shift, the influence and catalytic effect of data science on drug discovery cannot be denied. Scientific data is becoming public and even open access. Moreover, better computing capabilities and more data make it easier to use prior data to improve ongoing work.  Chemical Biology explores biology via chemical tools. In practice this means that the molecular interaction space of protein targets is probed. Computational Chemical Biology is the computational addition to these goals and is located in between the fields of medicinal chemistry, cheminformatics, bioinformatics, and computational biology. In this lecture, bio and cheminformatic approaches coupled to structure-based methods using crystal structures are discussed.