Pediatric Early Warning Score
Transformer-based models for predicting clinical deterioration in pediatric patients.
In collaboration with Children’s Healthcare of Atlanta (CHOA) and the Georgia Tech Pediatric Technology Center, we develop machine learning systems for pediatric early warning — predicting clinical deterioration before it becomes life-threatening, enabling timely intervention in hospital settings.
Problem: Existing Pediatric Early Warning Scores (PEWS) rely on hand-crafted rules applied to discrete vital sign thresholds. They miss subtle temporal patterns in continuous physiological data and generate high false-alarm rates that contribute to alarm fatigue.
Our approach: We apply transformer-based sequence models to continuous streams of physiological signals (heart rate, respiratory rate, SpO₂, blood pressure) from pediatric patients in CHOA’s electronic health records. The model learns to identify deterioration signatures hours before clinical events, providing interpretable risk trajectories to clinicians.
Key directions:
- Transformer architectures for irregularly-sampled clinical time series
- Calibrated uncertainty quantification for high-stakes clinical decisions
- Integration of structured (vitals) and unstructured (nursing notes) EHR data
- Alarm optimization to balance sensitivity against false-alarm burden
- Prospective evaluation and deployment at CHOA
Impact: Improved early warning could reduce preventable PICU admissions and cardiac arrests in pediatric wards, directly affecting patient outcomes at one of the largest pediatric healthcare systems in the United States.
This is an active collaboration. Results forthcoming.