The Impact of Multi-Institution Datasets on the Generalizability of Machine Learning Prediction Models in the ICU
Authors' conclusion
Our results emphasize the importance of diverse training data for DL-based risk prediction.
Study context
- Population
- Adults admitted to the ICU for at least 6 hours with good data quality, ICUs across Europe and the US
- Participants
- 334,812 ICU stays
- Exposure
- Single-center models vs Multicenter models
- Outcome
- Transferability (AUROC) of DL models for early detection of death, AKI and sepsis at previously unseen hospitals
Methodological detailsharmonized data from four public datasets, multicentre
- Harmonized data from four public datasets
- Multicentre
Main finding
As reported in the abstract
AUROC at training hospital and at new hospitals
Area under the receiver operating characteristic (AUROC) for mortality at the training hospital
0.838-0.869
AUROC for acute kidney injury (AKI) at the training hospital
0.823-0.866
AUROC for sepsis at the training hospital
0.749-0.824
AUROC drop when applied at other hospitals, sometimes by as much as
−0.200
Observational data: associations, not effects.
Original publication
The Impact of Multi-Institution Datasets on the Generalizability of Machine Learning Prediction Models in the ICU. Crit Care Med. Published 3 July 2024. doi:10.
Composed by AI from the published abstract and checked against it. How Briefs are made →