.png){kind=icon}
.png){kind=icon}
Return
Multi-scale temporal fusion transformer with adaptive uncertainty quantification for EHR-based infectious disease prognosis
.png){kind=icon}
PRE
AI
DOI：10.1007/s13042-026-03159-5.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
We present MSTFT, a transformer-based approach for infectious disease prognosis from multimodal electronic health records. Our work addresses key challenges in clinical time-series modeling: heterogeneous modalities with irregular sampling, missing values, and the need for calibrated uncertainty estimates suitable for clinical decision support. Key methodological contributions include: (1) hierarchical cross-modal attention specifically designed to learn interactions between modalities with different sampling frequencies and missingness patterns; (2) multi-scale temporal fusion that captures both short-term acute changes and long-term trends in clinical data, addressing the multi-timescale nature of sepsis progression; and (3) integrated uncertainty quantification and calibration that provides clinically actionable confidence estimates rather than treating uncertainty as post-hoc. On a large-scale ICU dataset of 5036 patients with 14.2 million clinical measurements, MSTFT achieves 94.2% AUC-ROC for mortality prediction (2.1% improvement over EfficientTransformer) and 0.546 AUC-ROC for infection detection. Calibration analysis shows an expected calibration error of 0.062, enabling reliable deployment in clinical settings. Results demonstrate that careful integration of transformer mechanisms with EHR-specific design choices yields interpretable, well-calibrated predictions suitable for high-stakes clinical applications.
Keywords:
Deep learning
Electronic health records
Infectious disease prognosis
Mortality prediction
Early detection
Journal
IF:
2.7
Papers:
3.1K
Citations:
5.6K
Organization
View more organizations