A popular approach to knowledge extraction from clinical databases is to first define an ontology of the concepts one wishes to model and subsequently, use these concepts to test various hypotheses and make predictions about a person's future health and wellbeing. The challenge for medical experts is in the time taken to map between their concepts/hypotheses and information contained within clinical studies. Presently, most of this work is performed manually. We have developed a method to generate links between Risk Factors in a medical ontology and the questions and result data in longitudinal studies. This can then be exploited to express complex queries based on domain concepts, to extract knowledge from external studies.

Abstract-Current estimates predict 1 in 3 people born today will develop dementia, suggesting a major impact on future population health. As such, research needs to connect specialist clinicians, data scientists and the general public. The In-MINDD project seeks to address this through the provision of a Profiler, a socio-technical information system connecting all three groups. The public interact, providing raw data; data scientists develop and refine prediction algorithms; and clinicians use in-built services to inform decisions. Common across these groups are Risk Factors, used for dementia-free survival prediction. Risk interactions could greatly inform prediction but determining these interactions is a problem underpinned by massive numbers of possible combinations. Our research employs a machine learning approach to automatically select best performing hyperparameters for prediction and learns variable interactions in a non-linear survival-analysis paradigm. Demonstrating effectiveness, we evaluate this approach using longitudinal data with a relatively small sample size.

Abstract. Agri-analytics is an emerging sector which uses data mining to inform decision making in the agricultural sector. Machine learning is used to accomplish data mining tasks such as prediction, known as predictive analytics in the commercial context. Similar to other domains, hidden trends and events in agri-data can be difficult to detect with traditional machine learning approaches. Deep learning uses architectures made up of many levels of non-linear operations to construct a more holistic model for learning. In this work, we use deep learning for unsupervised modelling of commodity price data in agri-datasets. Specifically, we detect how appropriate input signals contribute to, and interact in, complex deep architectures. To achieve this, we provide a novel extension to a method which determines the contribution of each input feature to shallow, supervised neural networks. Our generalisation allows us to examine deep supervised and unsupervised neural networks.

Abstract. Learning experiments are complex procedures which generate high volumes of data due to the number of updates which occur during training and the number of trials necessary for hyper-parameter selection. Often during runtime, interim result data is purged as the experiment progresses. This purge makes rolling-back to interim experiments, restarting at a specific point or discovering trends and patterns in parameters, hyper-parameters or results almost impossible given a large experiment or experiment set. In this research, we present a data model which captures all aspects of a deep learning experiment and through an application programming interface provides a simple means of storing, retrieving and analysing parameter settings and interim results at any point in the experiment. This has the further benefit of a high level of interoperability and sharing across machine learning researchers who can use the model and its interface for data management.