Close
Supervisors: Professor Neil Sebire, Professor Payam Barnaghi
Modern machine learning models provide powerful ways to analyse large-scale fixed and stationary datasets. They often outperform human-level ability in extracting patterns and creating predictive outcomes using large-scale data. However, these models fail to imitate humans' basic ability to continually learn in changing and non-stationary environments. The existing machine learning models are also often trained on a specific task and using single modality data.
The real world is, however, non-stationary. Healthcare records analysis requires continuous learning models that can analyse unseen data and adapt to the changes in the data (e.g. new treatments, new analysis, new experiences). Designing generalisable machine learning models that can analyse continuously changing electronic healthcare data requires new approaches to designing the learning methods. Continually learning models will provide novel ways to develop clinically applicable solutions to transform healthcare delivery by creating preventative and predictive measures. This will enhance patients' quality of treatment and care, reduce time, and improve the efficiency of clinical decision making.
The aim of this PhD research is to create an online digital consultation system with insights that could be extracted from large-scale electronic healthcare records. The unique digitally mature Great Ormond Street Hospital’s (GOSH) electronic record analysis environment, which has recently been achieved, combined with the research data analysis advancements in (deep) machine learning and scalable, high-performance computing, makes this approach possible. The research will initially focus on specific disease areas; however, the approach will be scalable across different conditions. By using standard ways of storing health data, it could also be used for children in multiple other centres across the country and beyond.
This technical research will focus on the following steps:
i) designing models that can incrementally learn new knowledge and adapt to emerging changes without significant deterioration in their performance.
ii) creating backwards and forward transfer in learning models to adapt to the changes without forgetting what has been learned in the past.
iii) investigating biological learning models and developing sparse latent representation, and pathway-guided networks to create dynamic learning models that can process high-dimensional, multimodal, and continuous remote monitoring data.
Validation studies will be carried out using a large-scale electronic healthcare record at Great Ormond Street Hospital (GOSH).
In April 2019, GOSH adopted the Electronic Patient Record System (EPR) EPIC. At the same time, a database for Electronic Health Data was established, using data from the EHR in combination with historical data and data held in legacy systems. To date, over 280 million events have been uploaded into the database for research, and this database continues to grow as data is continuously pulled directly from the EPR. This digital advancement was instrumental in GOSH becoming the first UK hospital to achieve the HIMSS Stage 7 digital maturity benchmark.
The datastore is a bespoke secondary-use data warehouse maintained and run by GOSH and falling within all of its usual IG and data security guidelines and policies. The datastore is managed by a specific team DRE (Digital Research Environment team, https://www.goshdre.com) who curate and provision data for researchers using the Aridhia DRE platform. The Aridhia platform is a secure cloud-based research environment with access to analytics tools (e.g., R, MatLab, Python) to allow advanced data analysis without the requirement for data export.
The PhD student will be based in Great Ormond Street Hospital Institute of Child Health at the University College London. The student will work at the GOSH DRIVE Unit within a team of Data Engineers and Data Scientists from the NHS, Academia, and Industry.
Related reading:
- Banerjee A, Pasea L, Harris S, Gonzalez-Izquierdo A, Torralbo A, Shallcross L, Noursadeghi M, Pillay D, Sebire NJ, Holmes C, Pagel C, Wong WK, Langenberg C, Williams B, Denaxas S, Hemingway H. Estimating excess 1-year mortality associated with the COVID-19 pandemic according to underlying conditions and age: a population-based cohort study. Lancet. 2020 May 30;395(10238):1715-1725. doi: 10.1016/S0140-6736(20)30854-0. Epub 2020 May 12. PMID: 32405103; PMCID: PMC7217641.
- H. Li, P. Barnaghi, S. Enshaeifar, F. Ganz, “Continual Learning Using Bayesian Neural Networks”. IEEE Transactions on Neural Networks Learning Systems 2021 Sep;32(9):4243-4252. doi: 10.1109/TNNLS.2020.3017292; 2021 Aug 31. PMID: 32866104.
- R. Hadsell, D. Rao, A. A. Rusu, R. Pasxanu, “Embracing Change: Continual Learning in Deep Neural Networks”. Trends in Cognitive Sciences 24 (2020): 1028-1040.
- Li, Y., Rao, S., Solares, J.R.A. et al. BEHRT: Transformer for Electronic Health Records. Sci Rep 10, 7155 (2020). https://doi.org/10.1038/s41598-020-62922-y