The field of electronic health record (EHR) analysis is moving towards the development of more sophisticated and multimodal models that can effectively harness the heterogeneity of EHR data. Recent research has focused on creating foundation models that can learn from multiple data modalities, including structured and unstructured data, to support clinical tasks such as prediction and narrative generation. These models have demonstrated superior performance in various clinical prediction tasks, including mortality prediction, readmission prediction, and disease diagnosis. Furthermore, there is a growing emphasis on developing interpretable models that can provide insights into the decision-making process, which is critical for building trust and transparency in clinical AI systems. Noteworthy papers in this area include:

• Generative Deep Patient, a multimodal foundation model that natively encodes structured EHR time-series and fuses it with unstructured EHRs, demonstrating superior performance in clinical prediction and narrative generation tasks.
• ProtoEHR, a hierarchical prototype learning framework that fully exploits the rich, multi-level structure of EHR data to enhance healthcare predictions, offering interpretable insights on code, visit, and patient levels.
• A deep learning approach using natural language processing techniques that integrates multimodal EHRs to predict mortality and resource utilization in critical care settings, showing strong resilience to data corruption within structured data.