Today, we are witnessing an ever-increasing number of exchanges and migration flows of exchanges and migratory flows. These exchanges and natural disasters are among the most influential factors in spreading infectious diseases. This fact could be affirmed by the recent pandemic of COVID-19, which has caused an acute health crisis worldwide. In this context, we distinguish several sources that are crucial in the generation of health-related data, including open data, social networks, patient data, and IoTs. These data are characterized by a very dynamic aspect, heterogeneity, complexity, and a high growth factor. These characteristics may impact the data usefulness and handicap the data analysis process, especially in health crisis management systems which are the focus of the present thesis. Further, despite the immense technological advances, current health crisis systems cannot still treat such massive data with genuine autonomy and intelligence since they still need to check predictable and preprogrammed situations to generate outcomes. In addition, the users of such systems may use them in different chaotic situations that imply several constraints, like restricting time to make decisions. Accordingly, they may have changing preferences and requirements regarding the data quality and the desired recommendations according to their user roles and decision context. Thus, the challenge of the present thesis is to answer the following problem. "How to generate recommendations intelligently and autonomously on multi-source, heterogeneous, uncertain, and complex data gathered in a data lake without prior knowledge?"

For this purpose, we identified two sub-problems about the recommendation systems considering different users' needs in different contexts. More precisely, we focused on addressing the underlying sub-problems, namely (1) "How to ensure the management of heterogeneous data, and more specifically, the curation of data adaptively collected in batch and streaming while considering the functional and non-functional needs of the user?" and (2) "How to recommend preventive health measures while providing explanations adapted to user roles in different decision contexts?". Therefore, our main objective is to propose an approach integrating an intelligent system to recommend the appropriate preventive health measures according to the user requirements via analyzing data from multi-sources. Hence, we proposed contributions addressing each step involved in the prediction and recommendation to tackle our main objective. First, we proposed a service-based approach for adaptive data curation in data lakehouses by considering the user role, preferences, constraints, and decision context. Indeed, we relied on data lakehouses as a practical solution to overcome the big data integration challenges. Hence, we took advantage of semantic technologies and reinforcement learning techniques to constitute a multilayered framework for data curation. Subsequently, we focus on disease prediction and health measures recommendation problems by proposing a semantic-based approach for explainable health measures recommendations adapted for multiple users with different needs. The presented contributions are implemented and experimented on medical domain scenarios.