Evaluating the SWIFT algorithm's efficacy in predicting hypoxemia across multiple critical care datasets.

Area of Science:

• Artificial Intelligence in Medicine
• Clinical Informatics
• Critical Care Medicine

Background:

• Machine learning models can predict patient hypoxia, enabling timely interventions.
• Limited generalizability of current models necessitates external validation.

Purpose of the Study:

• To validate the generalizability of the SpO2 Waveform ICU Forecasting Technique (SWIFT), an LSTM algorithm.
• To assess SWIFT's ability to predict SpO2 5 and 30 minutes in advance on external datasets.

Main Methods:

• The SWIFT model was trained on the eICU Collaborative Research Database (eICU-CRD).
• Validation was performed on the Medical Information Mart for Intensive Care IV (MIMIC-IV) and Amsterdam University Medical Centers Database (UMCdb) datasets.
• Performance was evaluated for both ventilated and non-ventilated patient populations using SWIFT-5 and SWIFT-30.

Main Results:

• Population size reduction occurred in MIMIC-IV and UMCdb due to SpO2 measurement frequency differences.
• SWIFT performed well on eICU-CRD but showed reduced performance on MIMIC-IV, especially SWIFT-30.
• UMCdb validation showed promise, with performance comparable to eICU-CRD for ventilated patients. High specificity and NPV were observed across datasets.

Conclusions:

• Generalizing prediction models across diverse ICU populations presents challenges, underscoring the need for external validation.
• Future research should enhance model adaptability and interpretability for clinical settings.
• Ensuring trust in clinical alarms requires high specificity and negative predictive value (NPV).