Background: Blood pressure is one of the most important modifiable risk factors for stroke. Although the influence of an individual’s average blood pressure (BP) on their overall stroke risk is well established, visit-to-visit blood pressure variability (BPV) - variation in blood pressure from one clinic visit to the next - may be an independent risk factor for stroke. The influence of BPV on stroke risk in the general population is not fully understood, nor is it known whether associations with BPV vary by pathological stroke type. Very large prospective studies, including exposure measurements of BP and BPV as well as accurate identification, confirmation and sub-classification of large numbers of stroke cases during follow-up, are needed to test the associations between BP parameters, stroke and its main pathological types. UK Biobank (UKB) is a very large prospective cohort study of ~500,000 middle aged adults recruited from England, Scotland, and Wales between 2006 and 2010. Participants completed a detailed baseline assessment at recruitment (which included self-report of prior stroke and BP measurement). Follow up for health-related outcomes (including new occurrences of stroke) in UKB relies on linkages to routine coded datasets for hospital admissions, death registrations and primary care data. Coded primary care data could also be used to capture novel exposures, like blood pressure variability (BPV). In this thesis, I aimed to investigate how large prospective epidemiological studies such as UK Biobank might be used to investigate the influence of BP, and in particular BPV, on stroke and its types and subtypes. I did this through advancing understanding of the identification and characterisation of stroke cases in large prospective studies, and of obtaining measures of BPV from linked primary care data. Specifically, I aimed: (1) to evaluate the accuracy of patient self-report of stroke, the accuracy of routinely available coded healthcare data for stroke, and the reliability and feasibility of ischaemic stroke classification systems for large epidemiological studies such as UKB; (2) to identify prevalent and early incident stroke cases in UKB using multiple overlapping sources of coded data, and determine the proportions of cases classified into main pathological types of stroke; (3) to explore the feasibility of using coded primary care data to obtain measures of BPV in UKB. Methods: (1) I performed a series of systematic reviews of published data on (i) the accuracy of patient self-report of stroke, (ii) the accuracy for stroke and its main pathological types (ischaemic stroke, intracerebral haemorrhage, subarachnoid haemorrhage) of International Classification of Diseases (ICD) coded hospital admissions and death certificates, and Read coded primary care records, and 3) the inter-rater reliability of ischaemic stroke classification systems. (2) Informed by this work I identified prevalent and early incident stroke cases in UKB using linked coded hospital and death registration data as well as self-report data. In a sub cohort of participants, I was able to assess the additional role in case identification of linked coded primary care data. I compared the numbers of potential stroke cases ascertained by multiple overlapping combinations of these data and examined the proportions classified into the main pathological stroke types. (3) Finally, I analysed data from about 10,000 Welsh UKB participants with linked coded primary care data to identify those in whom visit-to-visit BPV could be measured using coded systolic blood pressure values (BP). I explored the association between frequency of visits with coded BP values and: participant characteristics; time between visits; mean BPV; standard deviation of BPV (SD BPV). I also calculated within-individual agreement between coded BP and UKB baseline assessment BP. Results: (1) From my systematic reviews I found that self-report accuracy was strongly influenced by characteristics of the study population. In populations with low stroke prevalence up to 75% of self-reported strokes were false positives. ICD codes for cerebrovascular diseases had a broad range of accuracy for stroke and its main pathological types, but appropriately selected, ‘stroke specific’ ICD codes were consistently >70% accurate when compared to an independent reference standard for stroke. Few studies assessed the accuracy of either primary care data or combinations of data sources for stroke. The overall inter-observer reliability of ischaemic stroke classification systems ranged from moderate to almost perfect. Study characteristics other than classification system accounted for much of the variation in reliability. Additional features which enhanced reliability included use of clear rules, data abstraction protocols, computerised assignment, and reduced number of subtype categories. (2) The prevalence of stroke in UK Biobank based on linked ICD coded hospital admissions data and participant self-report was ~1.7%. The majority of these prevalent stroke cases were of ‘unspecified’ stroke type. Incident strokes captured by ICD codes were mostly hospital admitted cases, but a smaller additional proportion were fatal cases not detected in hospital admissions data. The majority (~89%) of ICD coded incident strokes were a specified pathological type. In the sub-cohort of UKB participants with additional primary care data linkage ~20% of potential incident stroke cases were detected by coded primary care data alone. (3) Among Welsh UKB participants with linked primary care data, around two thirds had sufficient coded data to estimate visit-to-visit BPV any time before recruitment, and just under half had sufficient coded data to estimate BPV during the 5 years before recruitment. Selecting participants with more visits reduced generalizability, but there was good variability in BPV amongst those selected (standard deviation in BPV range ~5mmHg to ~7mmHg), and reasonable agreement between coded BP and BP recorded at the UKB baseline assessment (intra class correlation coefficient 0.53, 95% CI 0.52 to 0.55). Conclusions: This work will inform the approaches to stroke outcomes ascertainment and the measurement of a novel exposure, blood pressure variability, in UK Biobank. This will enable future exploration of the associations between blood pressure parameters, stroke, and its main types and sub-types in UK Biobank. Investigating these associations will improve our understanding of causal pathways for the different pathological types and sub-types of stroke and underpin increasingly targeted strategies to modify BP for stroke prevention.

Yes / Cross-sectional and longitudinal studies have consistently reported an association between education and myopia. However, conventional observational studies are at risk of bias due to confounding by factors such as socioeconomic position and parental educational attainment. The current study aimed to estimate the causal effect of education on refractive error using regression discontinuity analysis. Methods: Regression discontinuity analysis was applied to assess the influence on refractive error of the raising of the school leaving age (ROSLA) from 15 to 16 years introduced in England and Wales in 1972. For comparison, a conventional ordinary least squares (OLS) analysis was performed. The analysis sample comprised 21,548 UK Biobank participants born in a nine-year interval centered on September 1957, the date of birth of those first affected by ROSLA. Results: In OLS analysis, the ROSLA 1972 reform was associated with a −0.29 D (95% confidence interval [CI]: −0.36 to −0.21, P < 0.001) more negative refractive error. In other words, the refractive error of the study sample became more negative by −0.29 D during the transition from a minimum school leaving age of 15 to 16 years of age. Regression discontinuity analysis estimated the causal effect of the ROSLA 1972 reform on refractive error as −0.77 D (95% CI: −1.53 to −0.02, P = 0.04). Conclusions: Additional compulsory schooling due to the ROSLA 1972 reform was associated with a more negative refractive error, providing additional support for a causal relationship between education and myopia. / Global Education program of the Russian Federation government (DP) and an NIHR Senior Research Fellowship award SRF-2015-08-005 (CW), The Department for Health through an award made by the NIHR to the Biomedical Research Centre at Moorfields Eye Hospital NHS Foundation Trust, and UCL Institute of Ophthalmology, London, United Kingdom (grant no. BRC2_009). Additional support was provided by The Special Trustees of Moorfields Eye Hospital, London, United Kingdom (grant no. ST 12 09)

Introduction Vitamin D status is an important public health issue due to the high prevalence of vitamin D insufficiency and deficiency, especially in high latitude areas. Furthermore, it has been reported to be associated with a number of diseases. In a previous umbrella review of meta-analyses of randomized clinical trials (RCTs) and of observational studies, it was found that plasma/ serum 25-hydroxyvitamin D (25(OH)D) or supplemental vitamin D has been linked to more than 130 unique health outcomes. However, the majority of the studies yielded conflicting results and no association was convincing. Aim and Objectives The aim of my PhD was to comprehensively explore the association between vitamin D and multiple outcomes. The specific objectives were to: 1) update the umbrella review of meta-analysis of observational studies or randomized controlled trials on associations between vitamin D and health outcomes published between 2014 and 2018; 2) conduct a systematic literature review of previous Mendelian Randomization studies on causal associations between vitamin D and all outcomes; 3) conduct a systematic literature review of published phenome wide association studies, summarizing the methods, results and predictors; 4) create a polygenic risk score of vitamin D related genetic variants, weighted by their effect estimates from the most recent genome wide association study; 5) encode phenotype groups based on electronic medical records of participants; 6) study the associations between vitamin D related SNPs and the whole spectrum of health outcomes, defined by electronic medical records utilising the UK Biobank study; 7) explore the causal effect of 25- hydroxyvitamin D level on health outcomes by applying novel instrumental variable methods. Methods First I updated the vitamin D umbrella review published in 2015, by summarizing the evidence from meta-analyses of observational studies and meta-analyses of RCTs published between 2014 and 2018. I also performed a systematic literature review of all previous Mendelian Randomizations studies on the effect of vitamin D on all health outcomes, as well as a systematic review of all published PheWAS studies and the methodology they applied. Then I conducted original data analysis in a large prospective population-based cohort, the UK Biobank, which includes more than 500,000 participants. A 25(OH)D genetic risk score (weighted sum score of 6 serum 25(OH)D-related SNPs: rs3755967, rs12785878, rs10741657, rs17216707, rs10745742 and rs8018720, as identified by the largest genome wide association study of 25(OH)D levels) was constructed to be used as the instrumental variable. I used a phenotyping algorithm to code the electronic medical records (EMR) of UK Biobank participants into 1853 distinct disease categories and I then ran the PheWAS analysis to test the associations between the 25(OH)D genetic risk score and 950 disease outcome groups (i.e. outcomes with more than 200 cases). For phenotypes found to show a statistically significant association with 25(OH)D levels in the PheWAS or phenotypes which were found to be convincing or highly suggestive in previous studies, I developed an extended case definition by incorporating self-reported data collected by UK Biobank baseline questionnaire and interview. The possible causal effect of vitamin D on those outcomes was then explored by the MR two-stage method, inverse variance weighted MR and Egger's regression, followed by sensitivity analyses. Results In the updated systematic literature review of meta-analyses of observational studies or RCTs, only studies on new outcomes which had not been covered by the previous umbrella review were included. A total of 95 meta-analyses met the inclusion criteria. Among the included studies there were 66 meta-analyses of observational studies, and 29 meta-analyses of RCTs. Eighty-five new outcomes were explored by meta-analyses of observational studies, and 59 new outcomes were covered by meta-analyses of RCTs. In the systematic review of published Mendelian Randomization studies on vitamin D, a total of 29 studies were included. A causal role of 25(OH)D level was supported by MR analysis for the following outcomes: type 2 diabetes, total adiponectin, diastolic blood pressure, risk of hypertension, multiple sclerosis, Alzheimer's disease, all-cause mortality, cancer mortality, mortality excluding cancer and cardiovascular events, ovarian cancer, HDL-cholesterol, triglycerides and cognitive functions. For the systematic literature review of published PheWAS studies and their methodology, a total of 45 studies were included. The processes for implementing a PheWAS study include the following steps: sample selection, predictor selection, phenotyping, statistical analysis and result interpretation. One of the main challenges is the definitions of the phenotypes (i.e., the method of binning participants into different phenotype groups). In the phenotyping step, an ICD curated phenotyping was widely used by previous PheWAS, which I also used in my own analysis. By applying the ICD curated phenotyping, 1853 phenotype groups were defined in the participants I used. In PheWAS, only phenotype groups with more than 200 cases were analysed (920 phenotypes). In the PheWAS, only associations between rs17216707 (CYP24A1) and "calculus of ureter" (beta = -0.219, se = 0.045, P = 1.14*10-6), "urinary calculus" (beta = -0.129, se = 0.027, P = 1.31*10-6), "alveolar and parietoalveolar pneumonopathy" (beta = 0.418, se = 0.101, P = 3.53*10-5) survived Bonferroni correction. Nine outcomes, including systolic blood pressure, diastolic blood pressure, body mass index, risk of hypertension, type 2 diabetes, ischemic heart disease, depression, non-vertebral fracture and all-cause mortality were explored in MR analyses. The MR analysis had more than 80% power for detecting a true odds ratio of 1.2 or larger for binary outcomes. None of explored outcomes were statistically significant. Results from multiple MR methods and sensitivity analyses were consistent. Discussion Vitamin D and its association with multiple outcomes has been widely studied. More than 230 outcomes have been linked with vitamin D by meta-analyses of observational studies and RCTs. On the contrary, evidence from Mendelian Randomization studies is lacking. In particular I identified only 20 existing MR studies and only 13 outcomes were suggested to be causally related to vitamin D. In the systematic literature review of previous PheWAS studies, I summarized the applied methods, predictors and results. Although phenotyping based on ICD codes provided good performance and was widely applied by previous PheWAS studies, phenotyping can be improved if lab data, imaging data and medical notes can be incorporated. Alternative algorithms, which takes advantage of deep learning and thus enable high precision phenotyping, needs to be developed. From the PheWAS analysis, the score of vitamin D related genetic variants was not statistically significantly associated with any of the 920 phenotypes tested. In the single variant analysis, only rs17216707 (CYP24A1) was shown to be associated with calculus outcomes statistically significantly. Previous studies reported associations between vitamin D and hypercalcemia, hypercalciuria, nephrolithiasis and nephrocalcinosis, may be due to the role of vitamin D in calcium homeostasis. In the MR analysis, I found no evidence of large to moderate (OR > 1.2) causal associations of vitamin D on a very wide range of health outcomes. These included SBP, DBP, hypertension, T2D, IHD, BMI, depression, non-vertebral fracture and allcause mortality which have previously been proposed to be influenced by low vitamin D levels. Further, even larger studies, probably involving the joint analysis of data from several large biobanks with future IVs that explain a higher proportion of the trait variance, will be required to exclude smaller causal effects which could have public health importance because of the high population prevalence of low vitamin D levels in some populations.

UK Biobank is a British clinical study containing over 40 000 Magnetic Resonance Images (MRI) with 100 000 MRI planned of participants aged 44-82 as well as a large amount of related medical data. Analyzing these images with a neural network to find relations between the information in an MRI image and various medical data could lead to interesting medical revelations. While other studies usually focus on improving the network architecture, we instead propose a method to get targeted information out of full body MRI images. This is done by sampling various sub-volumes of the full body images and making a collage specifically tailored to the problem at hand before feeding them to a ResNet50 based network. The images are further analyzed using saliency analysis in order to gain information on what regions the network found important. This method was attempted on a variety of medical data including age, kidney volume, liver fat percentage, and heart volume. The method is used both as a way to increase information density in the input images as well as restricting information, such that we can see how well the network can predict about some medical data point from only some part of the body.The collages are able to increase the information in the images while the more complex representation and non-continuous representation does not cause problems for the network. These collages are also conducive to getting clearer and sharper saliency maps, which may give interesting medical information by showing what regions the network considers relevant. This may reveal otherwise difficult to notice relations between the information in the MRI images and medical information.

A major health concern for subjects with diabetes is weaker bones and increased fracture risk. Current clinical assessment of the bone strength is performed by measuring Bone Mineral Density (BMD), where low BMD-values are associated with an increased risk of fracture. However, subjects with Type 2 Diabetes (T2D) have been shown to have normal or higher BMD-levels compared to healthy controls, which does not reflect the recognized bone fragility among diabetics. Thus, there is need for more research about diabetes-related bone fragility to find other factors of impaired bone health. One potential biomarker that has recently been studied is Bone Marrow Fat (BMF). The data in this project consisted of whole-body water-fat Magnetic Resonance Imaging (MRI) volumes from the UK Biobank Imaging study (UKBB). Each subject in this data has a water volume and a fat volume, allowing for a quantitative assessment of water and fat content in the body. To analyze and perform quantitative measurements of the bones specifically, a Deep Learning (DL) model was trained, validated, and tested for performing fully automated and objective skeleton segmentation, where six different bones were segmented: spine, femur, pelvis, scapula, clavicle and humerus. The model was trained and validated on 120 subjects with 6-fold cross-validation and tested on eight subjects. All ground-truth segmentations of the training and test data were generated using two semi-automatic pipelines. The model was evaluated for each bone separately as well as the overall skeleton segmentation and achieved varying accuracy, performing better on larger bones than on smaller ones. The final trained model was applied on a larger dataset of 9562 subjects (16% type 2 diabetics) and the BMF, as well as bone marrow volume (BMV) and cortical bone volume (CBV), were measured in the segmented bones of each subject. The results of the quantified biomarkers were compared between T2D and healthy subjects. The comparison revealed possible differences between healthy and diabetic subjects, suggesting a potential for new findings related to diabetes and associated bone fragility.

Human ageing is accompanied by characteristic metabolic and endocrine changes, including altered hormone profiles, insulin resistance and deterioration of skeletal muscle. Obesity and diabetes may themselves drive an accelerated ageing phenotype. Untangling the causal web between ageing, obesity and diabetes is a priority in order to understand their aetiology and improve prevention and management. The role of biological ageing in determining the risk of obesity and associated conditions has often been examined using mean leukocyte telomere length (LTL), a marker of replicative fatigue and senescence. However, considering phenotypes which represent different domains of biological and functional ageing as exposures for obesity and related traits could allow the elucidation of new understudied phenotypes relevant to cardio-metabolic risk in the wider population. This PhD considers the causal role of (1) hand grip strength (HGS), a marker of overall strength and physical functioning, and (2) resting energy expenditure, an indicator of overall energy metabolism and the major component of daily energy expenditure, in cardio-metabolic risk. I also characterise a new and readily-quantifiable marker of age-related genomic instability, mosaic loss of the Y chromosome (mLOY). Observational evidence implicates each of these phenotypes in cardio-metabolic conditions and intermediate phenotypes. However, it is not possible to infer causality from these observational associations due to confounding and reverse-causality. Mendelian randomisation offers a solution to these limitations and can allow the causal nature of these relationships to be investigated. Using population-based data including UK Biobank, this thesis presents the first large-scale genetic discovery effort for each trait and provides new biological insight into their shared and separate aetiology. I used identified variants to investigate the bidirectional causal associations of each trait with cardio-metabolic outcomes, intermediate phenotypes and other related traits such as frailty and mortality. In total I identified 16 loci for hand grip strength, 19 for mLOY, and one signal for REE. I have shown that HGS is likely to be causally linked to fracture risk, and I have identified the important shared genetic architecture between mLOY, glycaemic traits and cancer. I have also demonstrated that at least one known genetic variant contributing to obesity risk acts partially via reduced REE. Overall the findings of my PhD contribute to our wider understanding of the aetiological role of ageing processes in metabolic dysfunction, and have implications for both basic science and translational applications.

Peripheral blood biomarkers can inform clinical care and drug development. Establishing causality between biomarker and disease is often critical for such applications, but epidemiological studies are limited due to biases from confounding and reverse causation. Mendelian randomization analysis leverages random inheritance of genetic variants at conception to mimic properties of randomized studies and estimate unconfounded effects between biomarker and disease, or vice-versa. This thesis demonstrates the utility of Mendelian randomization as a complementary tool to elucidate observational studies, predict drug safety and repurposing opportunities, and improve diagnostic biomarkers for cardiometabolic diseases. First, we characterized the hypothesized relationship between lipoprotein(a) and atrial fibrillation. We demonstrated both observed and genetically predicted lipoprotein(a) levels were associated with higher risk of atrial fibrillation across multiple independent cohorts. Importantly, risk was partly mediated independent of atherosclerotic cardiovascular disease, a known consequence of elevated lipoprotein(a) and itself a risk factor for atrial fibrillation. Next, we explored the lifelong effects of endogenous testosterone across a comprehensive set of 461 health outcomes in 161,268 males from the UK Biobank cohort. Using Mendelian randomization analysis, we found higher testosterone had beneficial effects on body composition and bone mineral density but adverse effects on prostate cancer, androgenic alopecia, spinal stenosis, and hypertension. Finally, we applied Mendelian randomization with the intention of discovering biomarkers caused by disease, which are expected to represent markers of early disease. As a proof-of-concept, we applied this framework to identify biomarkers associated with genetic predisposition to kidney function among 238 biomarkers measured in the ORIGIN trial. We discovered reduced kidney function caused increased trefoil factor 3 and showed its addition to models with known risk factors improved discrimination of incident early-stage chronic kidney disease. Taken together, Mendelian randomization identified biomarkers that warrant further study, with promising implications for screening, prevention, and treatment of different cardiometabolic diseases. / Thesis / Doctor of Philosophy (PhD) / Biological markers associated with disease can inform novel therapeutics or diagnostics but distinguishing causation from correlation is challenging. Mendelian randomization – a technique that leverages random inheritance of genetic variation to infer causality – was used to examine the role of biomarkers in cardiometabolic diseases. First, we implicated lipoprotein(a) as a risk factor for atrial fibrillation that acts independent of atherosclerotic cardiovascular disease. Second, we comprehensively characterized the lifelong effects of testosterone on health outcomes in males, where we found evidence of both beneficial and adverse effects on disease. Finally, we discovered trefoil factor 3 as a diagnostic marker for early-stage chronic kidney disease. Altogether, this thesis demonstrated different applications of Mendelian randomization that showcase its utility as a complementary tool to reveal causal biomarkers, and served to identify biomarkers for cardiometabolic diseases that merit further studies to evaluate their potential benefit on patient care.