MRC DTP 4 Year PhD Programme: Discovery of composite biomarkers and the role on the retina in risk prediction for systemic diseases using deep learning

Aims and objectives. The project will use deep learning (DL) techniques to discover combinations of phenotypic and genotypic features working as predictive risk scores for high-incidence conditions (e.g. cardiovascular, diabetic complications).

Experience and environment. Doney has extensive and internationally visible experience of research bioinformatics. Doney directs the development of the eClinical Phenome within the Health Informatics Centre which manages large bioresources in Tayside linked to electronic Medical Records (EMR). Trucco has been funded by EPSRC (eg recent £1.1M grant on multimodal biomarkers for vascular dementia), the EU, MRC, the Royal Society, the Wellcome Trust, Toshiba, OPTOS and various charities. Trucco’s group (CVIP/VAMPIRE) has extensive experience of DL applied to medical image and data analysis, and runs its own GPU resources (~10 NVIDIA cards). The student will join the friendly and active CVIP/VAMPIRE laboratory and be attached to the NIHR student cohort for effective interdisciplinary integration.

Materials available. We shall rely on the GoDARTS bioresource comprising more than 9,000 diabetic patients and 8,000 controls (http://diabetesgenetics.dundee.ac.uk/) with numbers set to more than double in the next few years. In addition to genome-wide genotyping GoDARTS is linked to 30 years of comprehensive longitudinal electronic medical records. This is excellent ground for data mining with artificial intelligence techniques. All GoDARTS participants have provided consent for research of this kind and has been used by Doney and Trucco in collaborative projects for about 10 years,

Training and work plan. (1) Initial induction: familiarization with interdisciplinary computing-medicine group; literature review (DL and risk prediction; related topics); training in machine / DL, clinical statistics, clinical context; short project (~2-3 months) leading to conference paper. (2) Identify, implement, test DL architectures for target problem; focus on explanation, i.e., what data features does a network discover as crucial for the target classification. (3) Definition of final specific problem expected after first year.

References

R Annunziata, E Trucco: Accelerating Convolutional Sparse Coding for Curvilinear Structures Segmentation by Refining SCIRD-TS Filter Banks. IEEE Trans on Medical Imaging, vol 35 no 11, Nov 2016, pp 2381-2392.

A E Fetit, S Manivannan, S McGrory, L Ballerini, A J Doney, T MacGillivray, I J Deary, J M Wardlaw, F Doubal, G J McKay, S J McKenna, E Trucco. Retinal Biomarker Discovery for Dementia in an Elderly Diabetic Population. Proc MICCAI International Workshop OMIA-4, Springer, 2017.

A Veluchamy, L Ballerini, V Vitart, K E Schraut, M Kirin, H Campbell, P K Joshi, D Relan, S Harris, E Brown , S K Vaidya, B Dhillon, K X Zhou, E R Pearson, C Hayward, O Polasek, I J Deary, T J MacGillivray, J F Wilson, E Trucco, C N A Palmer , A S F Doney: Novel locus influencing retinal venular tortuosity is also associated with risk of coronary artery disease. biorXiv, Apr. 14, 2017; doi: http://dx.doi.org/10.1101/121012.