Optimizing supervised machine learning algorithms with practical applications

Machine learning (ML) and artificial intelligence (AI) are rapidly growing fields with applications in many scientific domains. In this work we focus on supervised learning algorithms for prediction tasks in practical applications. Specifically, we focus on two applications, predicting adverse events in low-carbon energy production and screening patients' eligibility for implantable defibrillators. We present optimised pre-processing methods, optimise hyperparameter selections and present our own bilevel optimisation model for simultaneous training and hyperparameter tuning. We first consider the problem of predicting infrequently occurring adverse events from time series data (Provided by Andigestion Ltd, a UK-based anaerobic digestion company, and a civil nuclear power plant in the UK) for which we propose a framework for modelling this problem as an imbalanced classification task and we construct and compare numerous models and sampling techniques. The models developed here could be integrated into a decision support tool for providing advanced warning of adverse events which can lead to significant commercial benefits. We then propose an AI tool for automated, prolonged screening of patients' implantable defibrillator eligibility which is created using real ECG data provided by our partners at the University Hospital Southampton. As we demonstrate in our experiments, this tool is capable of predicting patients' T:R ratios (a major indication of implantation eligibility) to within 0.0461 of their true values. This level of accuracy is sufficient to facilitate the automation of the measurement process. We show how this tool can enable cardiologists to perform 24-hour automated screenings, thus allowing them to better determine patients' eligibility for implantation. Finally, we formulate the bilevel problem of hyperparameter selection for non-linear kernel support vector machines via k-fold cross validation and propose an algorithm for solving it, for which we demonstrate some convergence properties. We provide a number of examples of this algorithm in use on a number of real data sets from the UCI repository.