Leveraging the power of data to provide better healthcare services

Utilizing healthcare data can significantly enhance service delivery, improve patient outcomes, and optimize operational efficiency. This PhD project focuses on developing and applying advanced statistical and machine learning models to analyze complex datasets for accurate prediction and improved decision-making. In Study 1 – Part I (Chapter 2), we predicted Length of Stay (LoS) and Prolonged LoS for emergency department inpatients at Sant’Orsola Malpighi University Hospital. Notably, Gradient Boosting excelled in PLoS prediction, while Ridge and XGBoost performed well in LoS prediction. Study 1 – part II (Chapter 3), we specifically targeted the General Medicine department, characterized by high patient volume and heterogeneity. We compared nine ML regression models in predicting hospital LoS. Feature Importance plots and SHAP (SHapley Additive exPlanations) were employed to identify the top important features and enhance interpretability. The eXtreme Gradient Boosting Regression model had the lowest prediction error. Study 2 (Chapter 4), conducted, when COVID-19 was at its peak, aimed to analyze the spatio-temporal patterns of the diffusion of SARS-CoV-2. The study aimed to derive a model for infection risk and identify place-specific factors, revealing varied impacts across city areas during the first three epidemic waves, with an estimated area-to-area influence within a 4.7 km radius. Study 3 (Chapter 5) aimed to identify admission risk factors associated with Length of Stay using Poisson, negative binomial, and Hurdle regression models. Hurdle–NB provided the best fit model. The ICU setting and long-term hospitals significantly influenced LoS, and age, wave periods, and hospital types also played crucial roles. Finally, Study 4 (Chapter 6) is an ongoing project targeting childhood cancer survivors, aiming to predict the risk of nonsurgical premature menopause using survival models and machine learning. The emphasis is on developing a landmark predictive model based on treatment-induced toxicities to improve long-term health outcomes for female childhood cancer survivors.