How Causal Analysis can Help Improve Radiation Cancer Treatments

By Dr. Eliana Vasquez Osorio, The University of Manchester

Radiotherapy is the use of targeted radiation to treat cancer. It plays a central role in cancer treatment, with around half of cancer treatments including this modality. However, many clinical decisions, especially those connected to treatment side-effects, are still guided by associations from observational data, often without sufficient scrutiny of the underlying causal assumptions.

Join us for an insightful session that promises to advance our understanding of identifiable deep generative models in sequence modelling.

Keynote Speaker: Dr. Eliana Vasquez Osorio

Register and join us on Wednesday 11th June at 1pm!

CHAI at the CLeaR (Casual Learning and Reasoning) Conference, Lausanne, Switzerland

The CLeaR (Causal Learning and Reasoning) conference was held this year in the beautiful city of Lausanne, Switzerland at the SwissTech Convention center from the 6th to the 9th of May. CLeaR, while being quite a young conference (with 2025 being only its 4th iteration), has established itself as the premier conference for causal statistics and causal machine learning. The range of submissions spanned far ranging areas from causal discovery, causal inference, causal model calibration, causal representation learning and more.

Each day began with talks from leaders in different areas of causal statistics and learning, the first day, the speaker was David Blei from Columbia university, on the topic of Hierarchical Causal Models, a new way of modelling nested or tiered data, and how that can provide opportunities for causal identification and estimation that may be impossible in a “flat” causal model structure. The speaker on the second day was Erin Gabriel from the University of Copenhagen, who spoke about causal inference from latent variables in various biomedical settings. Finally, on day 3, the speaker was Elias Bareinboim from Columbia university, who spoke about developments needed to move towards Causal AI systems, what the current challenges are, where there is room for development and some of the current works from his lab. The talk was quite insightful to identify which areas to model your research program after for building next generation AI systems.

CHAI paper accepted at the Conference on Uncertainty in Artificial Intelligence (UAI)

We are delighted to share that the KCL team’s CHAI paper was accepted by the conference on Uncertainty in Artificial Intelligence (UAI) 2025. The UAI is one of the leading conferences on research related to knowledge representation, learning, and reasoning in the presence of uncertainty.

Title: Explaining Negative Classifications of AI Models in Tumor Diagnosis

Authors: David A. Kelly, Hana Chockler, Nathan Blake

CHAI Research Project - Domain Generalisation - Update from CHAI PDRA, Kurt Butler

Generalization is one of the fundamental goals of AI research. The idea is that given enough data, our machine learning algorithms should be able to discover enough useful patterns to make predictions in new, previously unseen scenarios. For example, consider a piece of AI software that can generate images from text, like ChatGPT, and now imagine asking it to generate an image of a horse on the moon. No such photos exist in real life, but given enough data, the program can develop enough of an understanding of the concepts of “horse” and “moon” to generate an image that depicts a horse on the moon.

At CHAI, we are investigating how causal AI can help us with generalization. Consider the image of a horse on the moon again. Horses tend to be photographed in grassy areas, but nothing about photographing a horse should cause the background to change. One advantage of causal AI is that during training, we encourage the AI to learn causal insights (an object doesn't cause it's background to change), which should be useful to generalize better. The generalization project at CHAI is a collaborative endeavour, bringing together researchers from the University of Edinburgh, the University of Manchester, King’s College London, and University College London.

In healthcare, generalization becomes a crucial challenge. An AI system developed during a clinical trial to predict a cancer diagnosis might find that it is more reliable to predict a disease from its symptoms, rather than predicting a disease from causes. As a result, if individuals in the general population are more likely to answer inaccurately about their symptoms, then the AI model that predicts cancer from its symptoms will lose its usefulness. Causal AI can help us make systems that are less vulnerable to these kinds of variations.