463 - Pulse Oximetry Warning System (POWS): Machine learning to predict cardiorespiratory deterioration events in very low birth weight infants

Publication Number: 463.6104

Sara Manetta, University of Virginia School of Medicine, Charlottesville, VA, United States; Douglas E. Lake, University of Virginia, Charlottesville, VA, United States; Annabelle Leung, University of Virginia School of Medicine, Charlottesville, VA, United States; Sherry Kausch, University of Virginia, Charlottesville, VA, United States; Brynne Sullivan, University of Virginia School of Medicine, Charlottesville, VA, United States

Background: The Pulse Oximetry Warning System (POWS) predictive machine learning model was previously developed and externally validated to analyze continuous heart rate and oxygen saturation data for early detection of late-onset sepsis in very low birth weight (VLBW) infants. Retrospective chart review showed that high POWS scores occur in cardiorespiratory events with or without sepsis. To optimize POWS for future implementation as an early warning system, we prospectively evaluated POWS performance to predict clinically significant deterioration events.

Objective: 1) Determine the sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) of POWS to detect deterioration events
2) Quantify and characterize deterioration events detected by POWS

Design/Methods: In June 2024, three independent and blinded clinicians recorded daily logs of VLBW infants and identified significant deterioration events in near-real-time. Events were defined as those documented, intervened on, or discussed during hand-off. Events were labeled with diagnosis categories such as respiratory decompensation, suspected sepsis, etc. Hourly POWS scores were calculated from continuous pulse oximetry data sampled at 0.5 Hz (Figure 1). POWS alerts were defined as (1) a POWS score >2 AND POWS:baseline >2, or (2) a POWS score >5 AND a POWS score >95th percentile from baseline. POWS alerts occurring within 24 hours were grouped. True positive (TP) alerts had events occurring within 24 hours, and false positive (FP) alerts had none. Patient days without POWS alerts were designated as true negatives (TN) or false negatives (FN).

Results: We monitored 20 VLBW infants over 472 patient-days, documenting 129 clinically significant deterioration events (Table 1, Figure 2). Respiratory deterioration was most common (33%). Of 151 POWS alerts, 110 were TP and 41 FP. Among patient-days without alerts, 286 were TN and 19 FN. POWS demonstrated a sensitivity of 85%, specificity of 88%, PPV of 73%, and NPV of 94%. Review of FN events revealed many were non-specific or inconsequential, including gaseous abdominal distention and non-invasive respiratory support changes.

Conclusion(s): POWS detects clinically significant cardiorespiratory deterioration in VLBW infants with high sensitivity and specificity. The high NPV suggests utility in ruling out imminent deterioration when scores are low. POWS predicts various causes of deterioration beyond sepsis, highlighting its broader use in monitoring cardiorespiratory stability in VLBW infants and its potential as a real-time early warning system.

Figure 1. Visual Representation of POWS Scores
A visual display of POWS as a “heatmap” for all patients; blue color gradient represents low POWS scores, and yellow color gradient represents high POWS scores. An individual patient’s POWS score over time (red box) is visualized by row. Hourly POWS scores were calculated from continuous pulse oximetry data sampled at 0.5 Hz. POWS scores are calculated by a machine learning model that uses features derived from oxygen saturation and heart rate (green box) to make predictions.

Table 1. Patient Demographics
VLBW infant demographics are grouped as the primary June VLBW cohort and all VLBWs born in 2024 for comparison. Categorical variables are summarized as N (%) and continuous variables as median (IQR) or N (%). Groups were compared using Mann-Whitney U test and Pearson’s Chi-squared tests.

Figure 2. Clinically Significant Events by Categories
Clinically significant events by characteristic categories including respiratory concerns, infectious concerns, abdominal concerns, neurologic concerns, anemia, apnea, acidosis, hypotension, intubation and failed extubation, surgery, and death. Events are divided into TP and FN. Counts of events (above) are compared to percentages by type (below), TP (left) and FN (right).

Figure 1. Visual Representation of POWS Scores
A visual display of POWS as a “heatmap” for all patients; blue color gradient represents low POWS scores, and yellow color gradient represents high POWS scores. An individual patient’s POWS score over time (red box) is visualized by row. Hourly POWS scores were calculated from continuous pulse oximetry data sampled at 0.5 Hz. POWS scores are calculated by a machine learning model that uses features derived from oxygen saturation and heart rate (green box) to make predictions.

Table 1. Patient Demographics
VLBW infant demographics are grouped as the primary June VLBW cohort and all VLBWs born in 2024 for comparison. Categorical variables are summarized as N (%) and continuous variables as median (IQR) or N (%). Groups were compared using Mann-Whitney U test and Pearson’s Chi-squared tests.

Figure 2. Clinically Significant Events by Categories
Clinically significant events by characteristic categories including respiratory concerns, infectious concerns, abdominal concerns, neurologic concerns, anemia, apnea, acidosis, hypotension, intubation and failed extubation, surgery, and death. Events are divided into TP and FN. Counts of events (above) are compared to percentages by type (below), TP (left) and FN (right).