Abstract
The digitization of hospital information systems is becoming widespread, enabling the increasing integration of interactive visualization methods and machine learning into decision support systems. This development facilitates the anticipation of critical risks in monitored patients and helps reduce the workload of healthcare providers. However, Electronic Health Records (EHRs) contain large, heterogeneous, and temporal data. Applying these methods to such complex data presents two main challenges: (1) Designing models that can adapt to these specific characteristics ; (2) Providing tools to understand these complex data and the results produced by these models.In this thesis, we introduce GLIMPSE-Med, an interactive timeline-based visualization interface for temporal and heterogeneous events in the EHR, incorporating a score generated by a predictive model. The evaluation of this interface, conducted with healthcare professionals, confirmed that it meets two essential needs: (1) Assess the quality of data collected in an electronic health record ; (2) Estimate the patient's condition over time.We also present an original modeling approach based on recent advancements in graph convolution. In our approach, the patient journey is modeled as a graph where each node represents an event, and temporal proximities are represented by weighted edges. This modeling has been tested using graph convolution-based learning methods to predict death within 24 hours, employing a graph classification task on a real dataset. The obtained results were successfully compared with state-of-the-art methods.The work of this thesis is organized into two parts: (1) The first part focuses on visualization. It provides a state-of-the-art review on the visualization of multivariate timelines, both in medical and non-medical fields, and includes the presentation of GLIMPSE-Med along with its evaluation by healthcare professionals ; (2) The second part addresses the use of machine learning methods. It begins with a state-of-the-art review of graph convolution methods, followed by the presentation of a new method for representing patient journeys as graphs. This part concludes with a study of the generated latent representations, aiming to assess their ability to reveal common characteristics among patients.