Enhancing Causes of Death Prediction from Electronic Health Records through Multi-Modal Integration of Structured and Unstructured EHR Data

Rapid identification of death and causes of death is vital for medical product safety studies. In the United States, structured Electronic Health Record (EHR) data and unstructured notes offer valuable clinical information, but death data is often incomplete, and Cause of Death (CoD) details are typically missing. This study addresses these challenges by using a multi-modal approach that combines structured and unstructured EHR data to predict CoD more accurately. Integrating textual data with structured features improved the weighted Area Under the Curve (AUC) from 0.86 to 0.90. Significant AUC increases were observed for conditions like chronic lower respiratory disease (9%), cerebrovascular disease (8%), essential hypertension (9%), and intentional self-harm (9%). Unstructured notes were particularly beneficial for conditions with fewer samples. This study was presented at the American Medical Informatics Association (AMIA) 2024 Annual Symposium on November 13, 2024.