Developing tools for the early detection of upper gastrointestinal cancers using traditional modelling and machine learning approaches.

Introduction

Outcomes for the commonest forms of upper gastro-intestinal (UGI) cancer (oesophagus, gastric, pancreatic, gallbladder and biliary tract) are poor worldwide [1]. However, most UGI cancer patients have relevant symptoms and multiple healthcare consultations in the two years before diagnosis [2,3]. Recent studies have also shown that changes in patient factors, such as blood test results and the medications they are prescribed, can occur months or years prior to cancer diagnosis [4,5]. This suggests earlier detection is possible, which could improve patient outcomes including survival for UGI cancers. Artificial Intelligence (AI)-based innovative tools which combine existing and novel cancer predictors, and which take account of changes in these predictors over time, could help doctors better identify those at increased risk of undiagnosed UGI cancer for urgent specialist investigation.

This studentship will form an integral part of a large multi-institution research programme (CanDetect) which aims to develop and evaluate a UGI Multi-Cancer Early Detection (MCED) Platform for use in primary care.

Research aims and objectives

The overall aim of this doctoral research is to develop and validate novel diagnostic risk prediction models for UGI cancer using large electronic healthcare datasets, which contain detailed longitudinal information on more than 15 million UK patients. This will involve the use of both conventional machine learning methods (e.g. deep neural networks, support vector machines and random forests) as well as more recent AI approaches utilising Bidirectional Encoder Representations from Transformer (BERT) and those developed specifically for the analysis of electronic healthcare records (e.g. TAPER, BEHRT). These approaches will be used to create bespoke solutions that take into account the specific characteristics of the dataset. Machine learning approaches will help to model the risk associated with changes in patient factors over time allowing for dynamic risk prediction. Ultimately these models are intended to help select patients at elevated risk of UGI cancer for further investigation.

Specific aims:

1)     To model longitudinal changes in key variables, such as test results, prescriptions and symptoms over time to identify patterns associated with subsequent UGI cancer diagnosis

2)     To develop, externally validate and compare the performance of diagnostic prediction models for UGI cancer developed using conventional statistical techniques and machine learning methods

3)     To assess the likely impact of these models on early diagnosis, as well as on socio-economic, ethnic and geographic inequalities

Support and training

The doctoral student will be supported by a team with expertise in statistics and machine learning, general practice, cancer research and prediction model development. Training and courses will be tailored to meet the particular needs of the applicant. The student will be offered the opportunity to spend time with our UK and international collaborators and to attend national and international conferences to build their research network and learn new skills. 

Applications are invited from graduates with a BSc (First or Upper Second) or MSc (Merit or Distinction), or equivalent, to work within the Wolfson Institute of Population Health. Applicants with experience in mathematics, statistics, data science or machine learning are particularly welcome. This 4-year studentship can commence in either May or September 2024 and will be based at the Charterhouse Square Campus in central London.

To apply please submit an application on Queen Mary's application site here, including:

1)     A covering letter (less then 2 pages A4) outlining why you are interested in this project and your suitability for the role

2)     A CV (less than 2 pages A4)

Prospective applicants are welcome to contact Dr Garth Funston ([Email Address Removed]) and Dr Oleg Blyuss ([Email Address Removed]) for informal discussion before applying.