Data Augmentation and Machine Learning for Risk Assessment in Healthcare Associated Infections

Acquisition and analysis of extensive datasets is, today, a central tool in most research fields. Machine learning provides powerful methods to obtain descriptive and predictive models for the data in many applications. The acquisition of quality information is fundamental for the reliability and accuracy of predictive and classification models increasingly used in various applications. A correct and adequate use of A.I. models integrated with modern visual analytics techniques allows to extend and overcome the classical statistical methods, thus helping experts and professionals of different fields in decisions and policy-making. A key element for the success of machine learning models, beyond the continuous comparison with the experts in the field of application, is represented by the quality and the completeness of the data included in the analysis. The research reported in this Thesis focuses on the analysis, visualization and balancing of data collected in medical studies in the area of healthcare-associated infections (HAIs) to obtain useful classifier models. The results of this interdisciplinary work improve patient risk stratification and lead to targeted infection prevention and control interventions. This Thesis addresses these two issues with two main contributions: analytics technique designed to display pathways and common patterns in a sequence of events connected to associated outcome and a data augmentation method based on data imputation and oversampling of the minority classes to generate new records for training machine learning models and improve the visual analytics tools. The effectiveness of these methods is proved in selected real-world case studies, allowing to meet the performance requirements of Public Health, in particular with applications of visual analytics methods and machine learning models on medical datasets.