Predictive Modelling Approach to Data-Driven Computational Preventive Medicine

This thesis contributes novel predictive modelling approaches to data-driven computational preventive medicine and offers an alternative framework to statistical analysis in preventive medicine research. In the early parts of this research, this thesis presents research by proposing a synergy of machine learning methods for detecting patterns and developing inexpensive predictive models from healthcare data to classify the potential occurrence of adverse health events. In particular, the data-driven methodology is founded upon a heuristic-systematic assessment of several machine-learning methods, data preprocessing techniques, models’ training estimation and optimisation, and performance evaluation, yielding a novel computational data-driven framework, Octopus.

Midway through this research, this thesis advances research in preventive medicine and data mining by proposing several new extensions in data preparation and preprocessing. It offers new recommendations for data quality assessment checks, a novel multimethod imputation (MMI) process for missing data mitigation, a novel imbalanced resampling approach, and minority pattern reconstruction (MPR) led by information theory. This thesis also extends the area of model performance evaluation with a novel classification performance ranking metric called XDistance.

In particular, the experimental results show that building predictive models with the methods guided by our new framework (Octopus) yields domain experts' approval of the new reliable models’ performance. Also, performing the data quality checks and applying the MMI process led healthcare practitioners to outweigh predictive reliability over interpretability. The application of MPR and its hybrid resampling strategies led to better performances in line with experts' success criteria than the traditional imbalanced data resampling techniques. Finally, the use of the XDistance performance ranking metric was found to be more effective in ranking several classifiers' performances while offering an indication of class bias, unlike existing performance metrics

The overall contributions of this thesis can be summarised as follow. First, several data mining techniques were thoroughly assessed to formulate the new Octopus framework to produce new reliable classifiers. In addition, we offer a further understanding of the impact of newly engineered features, the physical activity index (PAI) and biological effective dose (BED). Second, the newly developed methods within the new framework. Finally, the newly accepted developed predictive models help detect adverse health events, namely, visceral fat-associated diseases and advanced breast cancer radiotherapy toxicity side effects. These contributions could be used to guide future theories, experiments and healthcare interventions in preventive medicine and data mining.