Speaker
Description
After a short introduction of the artifial intelliogence research group of the HUN-REN Alfréd Rényi Institute of Mathematics, including application of AI to theoretical (“Erdős”) problems, applications in energy management and applications to modernize the processing of archival documents using artificial intelligence we will focus on our flagship AI application project, on the patient journey analysis.
High-fidelity patient journeys are essential for advancing predictive modeling in healthcare. In Hungary, large volumes of electronic health records are available from diverse sources, including administrative and clinical data. As a first step, we focus on leveraging structured administrative records to build temporally ordered patient histories suitable for predictive modeling. In parallel, we are developing tools and workflows to extract structured representations from unstructured patient documentation, with the long-term goal of integrating these modalities into richer, unified patient trajectories, enabling their use in downstream predictive tasks.
the reserach and development project aims to utilize comprehensive patient journey datasets to enable the development of sequence-based predictive models for healthcare. Specifically, we focus on applying state-of-the-art transformer architectures to predict clinically relevant future health events, such as mortality or hospitalization. We also aim to investigate how such models can be adapted to the peculiarities of local healthcare data and explore their capacity to generate insights in real-world clinical settings.
We constructed structured patient timelines from health data, and used these datasets to train a variety of machine learning models, including gradient boosting machines and deep learning architectures. Particular emphasis was placed on sequential models, such as transformer-based architectures, to capture the temporal dynamics of patient history. We evaluated the models’ performance using standard classification metrics and applied model interpretation techniques to understand feature relevance and temporal dependencies.
Neural models consistently outperformed classical approaches in key prediction tasks, such as estimating six-month mortality risk. These models learned from complex, time-aware patient data and achieved high predictive accuracy. Furthermore, our interpretability analyses reveal clinically plausible patterns, such as the predictive role of recent diagnoses supporting the trustworthiness and usability of the models.
.
