A data-driven approach to model health-related behavioural change

Modelling human behaviour during the spreading of infectious diseases is a complex task: while, on one hand, individuals change their behaviour as a response to the epidemic, on the other the behavioural change of people heavily influences the spreading itself. Because of the high impact of this feedback loop on models' reliability and results, a great effort has been done in order to close it. In particular, during the last decade, computational epidemiology has seen the introduction of coupled behaviour-disease models. However, many questions remain open, especially regarding the validation of models proposed due to the lack of data. A recent review on the subject (Verelst et al, 2016) indicates that only 15% of the relevant papers published between 2010-15 considered empirical data. In this work, we focus on determining which are the main factors driving behavioural change during seasonal flu epidemics. To achieve this, we analyse a unique dataset comprised of behavioural surveys submitted by Italian users to the participatory surveillance platform Influweb over the 2017-18 and 2018-19 flu seasons. The analysis is performed using a state-of-the-art machine learning technique: Gradient Boosted Trees (GBT). We leverage these algorithms to gather insights on the relationships between individuals' characteristics and their attitude towards behavioural change. Outputs are then deeply analysed exploiting SHAP, a unified approach that connects game theory with local explanations to interpret the output of any machine learning model. Findings show that individuals tend to form three distinct groups: those who do not change their behaviour during flu season, those who take moderate preventive measures in order to avoid contagion, and those who take strong measures such as social distancing. Furthermore, the GBT algorithm independently selects as fundamental drivers the constructs suggested in the Health-Belief Model, which is by far the most commonly used psychological theory to explain and predict health-related behaviours (Hochbaum, 1958). Starting from these results, we implement innovative data-driven compartmental models that account for possible behavioural change of individuals. We extend the classic susceptible-infected-recovered model introducing new compartments to represent individuals who engage in behavioural defensive measures. The proposed changes seriously complicate the dynamics. In fact, an extensive exploration of the parameters space reveals complex phenomena such as phase transitions or multiple waves of infection. We propose also an economic-based model in which behavioural change is introduced as the result of the dynamic optimization of individuals who aim at maximizing the current and future utility they derive from contacts with others. This second approach is introduced to model the adaptive behavioural response to the epidemic that we observe from data. In conclusion, this work aims to advance our understanding of self-initiated behavioural change which represents a key challenge in modelling fundamental socio-economic phenomena, such as epidemics.