External ICU tests expose low positive predictive value for an AF model
Source check: Source checked. Not independent scientific confirmation.
Finding: A CatBoost model trained on Amsterdam ICU records reached ROC areas of0.77 in MIMIC-IV and0.84 at Ghent when transferred directly. In the respective imbalanced evaluation sets, positive predictive values were0.07 and0.02. Local retraining improved some metrics, but these retrospective tests do not establish a clinically useful alert.
Technology: CatBoost gradient boosting
Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.
- AI contribution
- Learned CatBoost classifiers use12-hour clinical-feature windows ending before recorded AF and are tested for hospital transfer; Powershap selects features.
- Test and result
- Development used18163 admissions, including2000 with nursing-chart AF. Matched AF/no-AF pairs stayed together in an80:20 split, and unmatched controls formed the imbalanced test set. External evaluation covered archived MIMIC-IV and Ghent cohorts; reported direct-transfer precision is specific to those sampled event/control constructions.
- Limitation / still unknown
- Nursing-chart timing may lag actual arrhythmia; prior AF history was unavailable, so this is not securely new-onset AF. Admission-level splitting does not establish isolation of repeated people. External feature availability influenced the common feature set. What the source calls recalibration includes local model retraining, not merely a probability adjustment. Calibration/ROC metrics do not prove a calibrated patient-specific AF risk or clinical benefit; event-aligned retrospective windows are not live alarms. Online supplementary procedures were not independently read.
Evidence sources and annotation context
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
Evidence sources: https://biblio.ugent.be/publication/01H113A35SMY5KN5ZFDMD3SA3S/file/01H11SBMXA8H225D2R2PKH44NS.pdf · https://doi.org/10.1016/j.ijmedinf.2023.105086
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: paper. Paper: linked separately.
Review date: 21 Nov 2026. Review scope: Actual original relevant clinical/computational Methods, reference labels, evaluation Results and limitations read.
No follow-up recorded; this is not a claim that no follow-up exists. Follow-up review scope and date are not recorded. No source-supported relationships recorded.
Editorial assessment: Impact 2/5 · Validation 2/5 · Novelty 2/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.