Research flagships
Our flagship research programmes address core AI research challenges and drive scientific breakthroughs. While currently in their early stages, we are focusing on unlocking the value-of real-world complex data for real-world impact in the healthcare landscape.
Scroll down to see what our researchers are working on.
Problem Fingerprinting
Problem
Actual outputs can differ from our predicted outcomes due to patient demographics, resources, staffing and operations across institutions.
Methods to account for when outputs differ from the prediction are too generalised, strict, or unrealistic. Fairness and reliability are important in clinical decision-making especially when using large medical image datasets.
Research goals
Educate the community on the benefits of causal AI over non-causal AI.
Demonstrate how real-world situations give rise to data shifts and model drifts. Suggest how we can adapt to these changes.
Collaborate with clinical communities to create tools that model dataset cause-effect relationships.
This research is being conducted across the whole of the CHAI Hub.
Active Causal Discovery
Problem
Methods for collecting data, determining what data are required, and discovering cause-effect relationships in that data are disconnected and inefficient. Addressing this is integral for experimental design for uncovering biological networks.
Research goals
Increase data collection ease starting with a toy setting.
Use what is learnt to suggest better experiments
Build a system that makes smart, real-time experiment suggestions.
This research is being conducted across The University of Edinburgh, The University of Manchester, UCL, and The Wellcome Sanger Institute.
Chronological Causal Discovery
Problem
Patient data and the healthcare environment change over time. For example, patients visiting hospitals at different times in different health states when hospitals have different equipment. This makes understanding causal relationships in health datasets much more complex.
Research goals
Develop a method for finding cause-effect relationships in irregular and non-stationary time-sensitive data.
Use the method to study large electronic health datasets.
This research is being conducted across The University of Edinburgh, UCL, and Canon Medical Systems Ltd..
Stress Test Medical AI with Counterfactuals
Problem
Obtaining large and diverse datasets is difficult and often the framework for checking medical AI is fair and unbiased is lacking. This is especially relevant when designing the population sample for clinical trials and processing comprehensive medical imagery datasets.
Research goals
Test medical AI algorithms with ‘what if’ scenarios with different data types.
Simulate real-world changes and challenges.
Productively engage with regulatory bodies and policymakers on regulation and policy issues.
This research is being conducted across The University of Edinburgh, Imperial College London, KCL, and Canon Medical Systems Ltd..
Meta-risk assessment
Problem
Developing accurate and reliable risk models is difficult and most AI methods are simpler non-causal linear models. This difficulty reduces adoption and real impact in everyday healthcare such as cardiovascular disease. The therapy area’s SCORE risk model is simple and doesn’t fully capture cause-effect relationships that could help doctors make smarter and actionable decisions based on the diseases’ underlying factors.
Research goals
Reduce risk prediction uncertainty.
Combine different risk models to increase reliability.
Test approach using Scottish health data.
This research is being conducted across The University of Edinburgh, UCL, NEURii, SCAN-DAN, and SCONe.