Staff member

Home hospitalization (HH) is presented as a healthcare alternative capable of providing high standards of care when patients no longer need hospital facilities. Although HH seems to lower healthcare costs by shortening hospital stays and improving patient's quality of life, the lack of continuous observation at home may lead to complications in some patients. Since blood tests have been proven to provide relevant prognosis information in many diseases, this paper analyzes the impact of different sampling methods on the prediction of HH outcomes. After a first exploratory analysis, some variables extracted from routine blood tests performed at the moment of HH admission, such as hemoglobin, lymphocytes or creatinine, were found to unmask statistically significant differences between patients undergoing successful and unsucessful HH stays. Then, predictive models were built with these data, in order to identify unsuccessful cases eventually needing hospital facilities. However, since these hospital admissions during HH programs are rare, their identification through conventional machine-learning approaches is challenging. Thus, several sampling strategies designed to face class imbalance were herein overviewed and compared. Among the analyzed approaches, over-sampling strategies, such as ROSE (Random Over-Sampling Examples) and conventional random over-sampling, showed the best performances. Nevertheless, further improvements should be proposed in the future so as to better identify those patients not benefiting from HH.

JTD Keywords: Hospitals, Blood, Training, Standards, Diseases, Prognostics and health management