Characterizing and Predicting End-of-Life Patient Trajectories Using Routine Clinical Data

Understanding the biological processes that precede death is critical for making informed clinical decisions and facilitating care transitions. Here, we analyzed routine clinical data of 292,576 patients from two large hospital cohorts in Germany and the United States to identify temporal patterns at the end of life.

Integrating comprehensive laboratory values, vital signs, and ICD codes, we identified consistent, multisystem trajectories of terminal decline that were conserved across age, sex, and diagnostic subgroups. Changes largely occurred in an orchestrated pattern, beginning with electrolyte imbalances, followed by hepatic and renal dysfunction, progressive vital sign deterioration, and culminating in coagulation failure. Based on internal data from over 80,000 deceased and non-deceased patients, we developed a diagnosis-agnostic machine learning model to predict 90-day mortality risk using 27 routine clinical parameters and 15 disease groups. The prediction model had a high accuracy on internal data (AUC: 0.86) and generalized well to an external cohort of 211,527 hospitalized patients (AUC: 0.79).

Our results provide a data-driven foundation for understanding end-of-life pathophysiology and demonstrate the potential of routine hospital data to inform individualized care planning.