Global Tech UpliftLinking Intelligence
Clinical AI
Led at Boston Children's Hospital. Senior-authored a continuous, noninvasive method to estimate arterial oxygen (PaO₂) from standard ICU monitor data, addressing the "SpO₂ = 100%" blind spot.
High-frequency vital signs and arterial blood gas measurements were retrieved from two hospital systems, standardized to 5-second intervals.
Records were merged into a PostgreSQL database, combining ~52,000 paired ABG samples with bedside monitor readings.
Derived metrics — recent SpO₂ averages, slopes, heart-rate disparities — were computed with Python (Pandas, NumPy).
Candidate approaches ranged from neural networks (Keras/TensorFlow) to an optimized conventional oxygenation equation.
Statistical assessment via Bland-Altman plots, intraclass correlation coefficients, and error analysis across SpO₂ ranges.
The empirically optimized "Sauthier ePaO₂" formula was chosen for clinical deployment.
The selected equation ingests streaming SpO₂ and heart-rate data, computing an estimated PaO₂ every 1-5 seconds.
Built-in noise detection compares pulse-oximeter and EKG heart rates; unreliable data is flagged or adjusted.
Estimated PaO₂ values display on bedside monitors with trend visualization and automated hypoxemia alerting.
Published in Critical Care Explorations (2021); designed for bedside monitor integration to catch dangerous oxygen drops between blood-gas draws.