SWIFT: A Deep Learning Approach to Prediction of Hypoxemic Events in Critically-Ill Patients Using SpO ₂ Waveform...

Area of Science:

• Critical Care Medicine
• Biomedical Engineering
• Artificial Intelligence in Healthcare

Background:

• Hypoxemia significantly increases mortality and adverse outcomes in critically ill patients, including those with COVID-19.
• Early identification of patients at risk for hypoxemia is crucial for timely and effective clinical interventions.
• Existing methods lack sufficient predictive power for anticipating hypoxemic events in intensive care settings.

Approach:

• Developed SWIFT (SpO2 Waveform ICU Forecasting Technique), a deep learning model utilizing prior SpO2 values to predict future SpO2 waveforms.
• The model forecasts blood oxygen saturation (SpO2) at 5 and 30-minute intervals.
• Validated SWIFT on novel datasets comprising critically ill and COVID-19 patient data.

Key Points:

• SWIFT accurately predicts over 80% of hypoxemic events in critically ill patients and 60% in COVID-19 patients.
• The model achieves a mean squared error (MSE) below 0.0007 for SpO2 waveform prediction.
• Provides 30-minute advance warning of both the occurrence and severity of potential hypoxemic events.

Conclusions:

• SWIFT offers valuable predictive insights into future hypoxemia, enabling proactive clinical management.
• The model can enhance clinical decision support systems for patient triaging and resource allocation.
• SWIFT has the potential to significantly improve patient care and outcomes in critical care, particularly during pandemics.