Student for Dynamic AI-driven thrombosis risk prediction in ICU patients

This internship focuses on developing an AI-driven approach for dynamic thrombosis risk prediction in ICU patients using both structured and unstructured clinical data. Thrombosis is a major healthcare challenge with complex and multifactorial risk factors. While thrombosis can often be prevented with timely administration of low dose anticoagulants, this requires accurate risk prediction.

What will you do?

In this project, you will contribute to developing a novel pipeline for dynamic thrombosis risk prediction using electronic health record (EHR) data. Depending on your interests, the project can be more data engineering-focused or AI-focused.

Your tasks may include:

• Designing and building a data pipeline / platform for clinical data (raw EHR → de-identification → preprocessing)
• Integrating structured data (e.g. lab values) and unstructured data (e.g. clinical notes, radiology reports)
• Exploring LLM- or alternative methods-based feature extraction from unstructured data. A concrete example includes training a model to reliably classify radiology reports as VTE yes/no and, if yes, where, but there are many potential use cases here, in increasing difficulty.
• Developing a dynamic, time-dependent prediction model for VTE risk
• Evaluating how inclusion of unstructured data improves prediction performance

What is this research about?

In this project you will work in a clinical research environment closely related to the ICU, focused on improving prediction of patient outcomes through predictive modelling. This project builds on earlier work in thrombosis prediction.

What is the problem?

Thrombosis is a common and serious complication in hospitalized patients. Although prophylactic treatment with low dose anticoagulants can effectively prevent thrombosis, it may also cause harm through bleeding. Therefore, predicting who is at risk of developing thrombosis is a key first step in order to apply individualized prevention.

Current prediction models rely mainly on structured data and static snapshots, usually at hospital admission, whereas large amounts of valuable unstructured clinical EHR data that could describe the clinical course and improve predictions remain unused.

What do we ask?

You're in a relevant HBO or WO program, for example in Data Science, AI or Biomedical Engineering.

• Affinity with AI, data engineering, or healthcare
• Interest in working with clinical data
• Independent and proactive attitude
• Good communication skills
• Python experience

What we offer

• Internship agreement with UMCG
• Good supervision at UMCG
• Scientific working environment
• In consultation, you can partly work from home.