Work

Exploring a dynamic digital biomarker of illness severity in older ICU patients via interpretable time-series deep learning modeling

Year: 2026

Type: article

Abstract: Older adults admitted to intensive care units (ICUs) experience rapid physiological deterioration, pronounced heterogeneity, and limited physiological reserve, necessitating continuous and individualized assessment of illness severity. Dynamic digital biomarkers derived from high-frequency electronic health record data offer a promising solution for capturing evolving patient states in real time, yet their development and validation in elderly ICU populations remain limited, particularly with respect to interpretability and cross-institutional generalizability. In this international multicenter study, we developed an interpretable time-series deep learning model (Elder-DDB) to derive a dynamic digital biomarker of illness severity in critically ill older adults, operationalized as a continuously updated, model-derived mortality risk trajectory over time. Using longitudinal data from more than 200 hospitals across the United States, Europe, and Asia, the model was trained on 41,050 elderly ICU admissions and externally validated on 47,335 admissions from independent cohorts. Across all external validation cohorts, the model demonstrated robust and consistent discrimination (AUROC range: 0.801–0.834), significantly outperforming conventional clinical scores and state-of-the-art deep learning baselines. Interpretability analyses confirm the proposed dynamic digital biomarker reliably assesses real-time illness severity in elderly ICU patients, supporting precision critical care.

Source: npj Digital Medicine

DOI: 10.1038/s41746-026-03114-2