ML clinical prediction models that are trained on a specific patient population may not offer reliable prediction if the data distribution changes after the model is deployed. This points to the need for:
i) solid theoretical understanding of models' behaviour under data shift
and
ii) implementation of OOD detection models to determine to what extent the model is reliable.
The candidate will contribute to the solution to this problem by:
- Investigating the desired properties for OOD detection models and improving on current techniques;
- Carrying out experimentation to test which OOD detection methods work in practice on medical data.
The project will focus on Electronic Health Records from various medical sources. The goal of the project is to improve our understanding of the OOD detection problem and enable safe(r) deployment of medical AI tools by providing concrete and implementable solutions.
For examples of papers on the topic see
here and
here.
We are looking for a PhD candidate who is eager to contribute to the safer implementation of ML methods in healthcare settings.You will investigate how to build a suitable OOD detection model that is meant to act as a failsafe layer, enabling the user of a ML model in healthcare to know to what degree the model in question is reliable. This exploration will address a number of methodological topics; a non-exhaustive list includes:
- Robustness under data shift.
- Density estimation.
- Interpretability and explainability for OOD detectors.
- Uncertainty estimation and model confidence.
- Counterfactual and/or adversarial methods for model robustness.
- Domain adaptation.
You will provide efficient and scalable implementations of your methods and will integrate them with popular open-source systems. Via your affiliation with the Medical Informatics department, you will have access to several large datasets comprising tens of thousands of patient records.
You will be integrated in an ongoing collaboration with industrial partners, with concrete possibilities for your research to influence the development of existing medical AI products. Finally, you will be integrated in the vibrant community of the Medical Informatics department, with the opportunity to interact with talented researchers and to contribute to top-rated education programs.