Professor Mihaela van der Schaar, John Humphrey Plummer Professor of Machine Learning and AI in Medicine at the University of Cambridge, explains why given the virtually infinite number of different diseases, variables, and patient needs, it’s not possible to hand-craft an ML model to analyse each disease clinically. Instead, we need a principled method, such as Automated ML, to leverage ML effectively in clinical analysis.
In this short excerpt from her full keynote session delivered at the 2020 International Conference on Learning Representations (ICLR), Mihaela outlines the limitations of existing AutoML methods when applied to clinical datasets and compares these to the frameworks developed by her lab in recent years, including AutoPrognosisis; a tool for crafting clinical scores.
The key topics Mihaela covers are:
- Complex combined pipeline selection and hyperparameter optimisation – a hard learning and hard optimisation problem
- Bayesian Optimisation does not work well where there are many dimensions – the curse of dimensionality
- Developing a novel structured kernel learning model to overcome the curse of dimensionality
- The results: her team have demonstrated this to be a powerful method of building clinical analytics for asthma, Cystic Fibrosis, breast cancer, and a variety of cardiovascular diseases.
For more information on van der Schaar lab’s recent collaboration with Public Health England and NHS Digital, click here.
The full 55-minute keynote session can be found here