Available
Project number:
2025_73
Start date:
October 2025
Project themes:
Main supervisor:
Associate Professor in HealthCare AI
Co-supervisor:
Professor Richard Dobson
Additional Information:
Synthesising Realistic Patient Timelines Using LLMs for Counterfactual Modelling in Dynamic, Time-Varying Treatment Regimes
In medical decision-making, counterfactual prediction allows clinicians to estimate potential treatment outcomes by considering alternative therapeutic actions based on a patient's observed history. Recent advancements in machine learning, including deep learning techniques such as Transformers, provide new, data-driven approaches for estimating treatment effects from available data.
Transformers are particularly effective in modelling patient trajectories by leveraging timelines of covariates and treatment histories. However, the training data often suffer from sparsity (e.g., lack of evidence for specific treatments) and bias (e.g., underrepresentation of certain demographic groups), which can significantly impact the performance of data-hungry models like Transformers. A common strategy to address these challenges is the use of synthetic data generation. However, the quality of the data synthesised often suffers from poor accuracy and diversity. The problem of diversity can be alleviated by using the creative power of LLMs. However, the increase in diversity comes at the cost of decreased accuracy. The problem is underinvestigated and the assessment of clinical validity and utility of the generated data remains very limited.
The main aim of this project is to develop an LLM-based synthesis method to model realistic diverse patient timelines for improved counterfactual modelling.
The following objectives will allow to reach this aim: (1) Developing a Transformer-based counterfactual modelling method for predicting patient outcomes under varying treatment regimes over patient timelines and incorporating uncertainty estimations to improve reliability; (2) Developing a method to generate realistic patient timelines with the help of LLMs under clinical validity control; (3) Develop an assessment framework for counterfactual outcome prediction exploiting synthetic patient timelines.
Main project output will be software for assisting clinical experts with personalised treatment suggestions based on counterfactual Transformer-based algorithms.
We are now accepting applications for 1 October 2025
How to apply
Candidates should possess or be expected to achieve a 1st or upper 2nd class degree in a relevant subject including the biosciences, computer science, mathematics, statistics, data science, chemistry, physics, and be enthusiastic about combining their expertise with other disciplines in the field of healthcare.
Important information for International Students:
It is the responsibility of the student to apply for their Student Visa. Please note that the EPSRC DRIVE-Health studentship does not cover the visa application fees or the Immigration Health Surcharge (IHS) required for access to the National Health Service. The IHS is mandatory for anyone entering the UK on a Student Visa and is currently £776 per year for each year of study. Further detail can be found under the International Students tab below.
Next Steps
- Applications submitted by the closing date of Thursday 30 January 2025 will be considered by the CDT. We will contact shortlisted applicants with information about this part of the recruitment process.
- Candidates will be invited to attend an interview. Interviews are projected to take place in March 2025.
- Project selection will be through a panel interview chaired by either Professor Richard Dobson and Professor Vasa Curcin (CDT Directors) followed by informal discussion with prospective supervisors.
- If you have any questions related to the specific project you are applying for, please contact the main supervisor of the project directly.
For any other questions about the recruitment process, please email us at
