medicine decision making

Data Analytics

How to best develop individualized prediction algorithms?

Our Digital Twin captures all medical-related life-course information about a patient. Think of genomic details, MR images, medical records, treatment outcome information, and social and economic information. The former two work packages focussed on data collection and storage; we now work with these data to build personalized diagnosis, monitoring, and prediction models.

Life-course data analysis

This project studies novel statistical and learning approaches that can leverage the opportunities of such data. Life-course data analysis can also reveal new subtypes of patients with an unexpected disease outcome or therapy response. We will build learners that can deal with these new, unexpected classes.

Finally, we will test how these individualized predictions can help patients, being inspired by the Digital Twin roadmap of a patient journey in the context of cardiovascular disease.

medicine decision making

Current Projects

Online learning data to robustly predict cardiovascular outcomes with a dynamic model

In the clinic, data will get added to the digital twin over time in batches. That means that the model will learn only from one part of the data but not from the new unprocessed data. This could hurt the model fit as found relationships may change over time. We want to test updating these models to make them dynamic and better fit the current situation.

We will online train in the ALLHAT & TOHP, which are two large longitudinal cardiovascular trials. We will investigate how we can update a model using a decision-based strategy. The lessons learned can help us build a dynamic digital twin which can be maintained over time.

Describing the early development in the womb with AI to predict and prevent adverse outcomes

We will use Artificial Intelligence, specifically: deep learning techniques such as a variational autoencoder on echographic images to construct a lower-dimensional feature vector. This representation can be used for downstream stream tasks such as:

  • Landmark detection (detecting parts of the fetus).
  • Birthweight prediction (which could signal growth retardation).

This could influence clinical decisions by identifying which patients are at high risk for bad outcomes. Subsequently, the digital program may be helpful for doctors to screen patients’ images to allocate care more efficiently.

Digital Twin position paper: defining the state of the art

We define the state-of-the-art digital twin in healthcare. This cross-domain position paper allows many specialists to combine their knowledge and viewpoints to generate consensus on how best to move forward to practically construct a digital twin.

We show that a virtual instantiation of a human relies on knowledge in the medical, technical and ethical domain and needs to have clearly defined values and trade-offs to ensure optimal outcomes. We collaborate with the work packages Data Collection, Data Storage and Data Implementation.

 

Data missing at random (MAR) and deep learning

Handling missing data is a part of data analysis. In a clinical trial in which data is collected over time, a patient might be excluded from further participation if one of his measurements crosses a clinically relevant threshold. The missing data will result in his future measurements missing. We call this “missing at random” (MAR), and we could make a valid interference using multiple imputations.

We expect these imputation procedures for MAR data to improve the predictive performance of deep learning algorithms applied to clinical data. We will conduct simulation studies with MAR data and compare commonly used imputation methods in deep learning.

 

Prediction of hand functioning, pain and treatment satisfaction for patients with thumb osteoarthritis

This project aims to predict hand functioning, pain and satisfaction for patients with thumb osteoarthritis for two treatment choices, surgery versus no surgery.

Patients are initially treated with conservative treatment for 3 months and then have the choice to undergo surgery. During the 3 months of conservative treatment, a variety of measurements are recorded over time. We will train an artificial neural network to predict hand functioning, pain and satisfaction one year after treatment. Using these predictions, we will study the benefit of surgery. The treatment of missing data has a particular emphasis.