Predictive analytics of integrated genomic and clinical data using machine learning and complex statistical approaches

Large-scale electronic health record (EHR) data has the potential to transform our understanding of disease aetiology and clinical risk, providing important information for clinical decision making and health policy at population scale. Aligning multiple sources of data, including genetic, epigenetic and clinical data can substantially improve our understanding of disease risk. Such big data can provide unique opportunities for development of algorithms for precision medicine as well as for novel discovery of candidate genes associated with disease; providing a framework for integration of genetic discovery into clinical applications in medicine. While large-scale EHR resources have been utilised for risk prediction, and identification of genetic associations, analyses of these data have been limited and have not harnessed the multi-dimensional and longitudinal richness of data. The recent development of large-scale biodata resources within the UK necessitates the parallel development of flexible analytic methods that can realise the full potential of these rich datasets.

Machine learning methods provide a framework whereby the relationships among different variables and types of data are learnt from the data itself by identifying and utilising features that best predict outcomes. Such approaches may provide an advantage over classical statistical methods, where the relationships among variables require pre-specification, and only a limited number of factors can be modelled at a time. By contrast, machine learning methods not only allow model-free prediction of clinical risk, but also help better understand which factors, among large numbers of potential predictors, influence clinical risk.

This proposal focuses on development and validation of statistical, and machine learning approaches that utilise complex data to flexibly predict the risk of patient outcomes across multiple disease areas. Additionally, application of such methods also allows a better understanding of clinical and genetic risk factors associated with disease. This work will specifically focus on systematically evaluating and extending current statistical methods for risk prediction to incorporate machine learning approaches that can model complex patterns of risk using large-scale genetic and clinical data flexibly, while appropriately accounting for the time-dependent context of risk factors (e.g. repeated measurements).

The first phase of the project will involve integration, curation and harmonisation of publicly available biodata resources, such as the UK Biobank, Genomics England and INTERVAL study. Further data, including gene expression, and functional data will be layered to develop a rich multi-dimensional dataset. These data can be reasonably predicted from sequence data, when not directly measured, using published imputation and deep learning approaches. Developing on previous work with EHRs, complex statistical approaches will be used to develop predictive algorithms for specific disease areas. The next stage will involve evaluation of existing machine and deep learning approaches that can model risk. These approaches will then be extended to model longitudinal data incorporating repeated measurements over time. The predictive accuracy of these approaches will be evaluated using independent datasets. To understand the genetic aetiology of disease, classical GWAS approaches will be compared with approaches that integrate machine learning to allow prioritisation of the most important genetic and clinical predictors of risk.

This project will provide a broad analytic framework for clinical risk prediction and identification of genetic associations with disease in the context of big data analytics. In the longer term, this will contribute to a programme of developing research capacity and expertise in high throughput analytics of multi-dimensional data with the aim of supporting clinical decision making, and improving patient health.

Although the utility of complex statistical, machine and deep learning (ML and DL) approaches in the context of multi-dimensional data has been clearly demonstrated, these methods have not been widely utilised to improve novel drug discovery and clinical risk prediction. This proposal aims to harness the potential of large-scale integrated genetic and health data to spur innovation, and develop predictive algorithms to improve clinical decision making and patient health. Specifically, this will focus on the development and evaluation of ML and DL frameworks for GWAS, and clinical risk prediction using publicly available large-scale EHR and genomics biodata resources, including UK Biobank, Genomics England and INTERVAL studies. Transcriptomic and functional data will be integrated into these using predictive approaches, where this has not been directly measured.

This will be implemented in three stages: 1) assessment of complex time-dependent statistical approaches for modelling of hazard; 2) optimisation and assessment of existing ML and DL approaches for modelling of clinical risk; 3) development of novel approaches, specifically using recurrent neural networks (RNNs) to incorporate temporality and missingness in clinical data, including time varying covariates to accurately model complex hazard functions; the objective of this project will be to develop approaches that appropriately leverage the rich longitudinal and time-dependent data on individuals shown by us and others to substantially improve clinical risk prediction.

In addition to risk prediction, this proposal will also focus on improving our understanding of genetic aetiology of disease. In addition to standard GWAS approaches, hybrid ML and GWAS approaches for prioritisation of candidate genes, and genetic variants associated with disease will also be applied, potentially improving the power to identify novel associations, with important implications for prioritisation of therapeutic targets.