Characterization of Behavioural Patterns of Portuguese Blood Donors using Supervised and Unsupervised Learning

Abstract

Blood donations are essential to save innumerous lives on a global scale on a daily basis. Without blood donations, many medical procedures cannot take place. Thus, the study of what motivates blood donors to donate and how they behave is important to ensure a stable and safe blood supply. Several studies tried to understand the most important factors for blood donor return, by using mainly logistic regression models. Those studies identified several donor demographic characteristics as impor tant factors to describe donors’ future behaviour. However, in this dissertation it is argued that if models have a poor performance in the task for which they are trained for, the conclusions taken from them may be erroneous. Thus, this dissertation presents a contribution for the study of understanding blood donor behaviour by using the most recent machine learning, evaluation and interpretability techniques. In this dissertation, several machine learning experiments are implemented aiming to predict blood donors return one year following a given donation, gaining insights about blood donors future behaviour and which factors influence it the most. Primarily, the blood donations dataset is split according to several geographic characteristics. Each segment is further split into blood donations from new and experienced donors (i.e. those who donated more than once). For experienced donors several features regarding their past behaviour are computed. Finally, different machine learning models are trained on top of each segment. Our results suggest that donor’s demographics, as well as features regarding the donation, are not enough to predict donor return. As such, it is not possible to estimate the impact that donor’s demo graphics have on donor’s future behaviour. However, models trained over experienced donors performed significantly better than those trained over new donors data, due to the impact of past behaviour features. However, even with past behaviour features the machine learning models do not achieve outstanding scores in predicting donor future behaviour, and, as such, this work demonstrates that both demographics and past behaviour features are insufficient to accurately explain future behaviour.