Perioperative medicine & outcomes
Multi-source data integration through pooling and transfer learning improves generalizability and specialization of deep learning models for ICU mortality and length of stay prediction: a four-database external validation study
Authors' conclusion
Single-source training compromises clinical applicability. Data pooling across harmonized international cohorts is the most effective strategy for both generalization and specialization, supporting a "Data-Centric AI" approach.
Study context
- Population
- ICU patients in four harmonized international databases (eICU-CRD, MIMIC-IV, AmsterdamUMCdb, HiRID)nearly 20,000 per database
- Participants
- 3,574 / 10,915 / 14,489 / 14,489
- Comparison
- Single-source training vs data pooling vs transfer learning
- AmsterdamUMCn = 3,574
- MIMIC-IVn = 10,915
- Data poolingn = 14,489
- Transfer learningn = 14,489
- Outcome
- Generalizability and specialization for ICU mortality (AUROC, AUPRC) and remaining LOS (MAPE) prediction; composite metric (Mcomposite)
Methodological detailsmulticentre, four-database external validation, harmonized BlendedICU databases
- Multicentre
- Four-database external validation
- Harmonized BlendedICU databases
Performance
As reported in the abstract
- Internal validation overestimated external performance: MAPE increase up to
- 51.8%
- Internal validation overestimated external performance: AUROC drop up to
- 13.8%
- Data pooling best generalization: Mcomposite improvement over single-source models up to
- 8.0%
Original publication
Multi-source data integration through pooling and transfer learning improves generalizability and specialization of deep learning models for ICU mortality and length of stay prediction: a four-database external validation study. Crit Care. Published 18 May 2026. doi:10.
Composed by AI from the published abstract and checked against it. How Briefs are made →