788 - How Much Medical Oxygen Is Needed?: A Novel Approach to Estimating Hypoxemia Incidence in Rural Malawi Using Anonymous Pulse-Oximetry Data

Publication Number: 788.5148

James Newton, University of North Carolina at Chapel Hill School of Medicine, Chapel Hill, NC, United States; Steven Donahoe, Unaffiliated, Pittsburgh, PA, United States; Eugene Katenga-Kaunda, Balaka District Council, Balaka, Balaka, Malawi; Juan Carlos Puyana, University of Pittsburgh School of Medicine, Pittsburgh, PA, United States

Background: Pneumonia remains the leading cause of death for children under five globally, with 99% of fatalities occurring in low-resource settings. Hypoxemia increases risk of death fivefold. However, rural health centers often lack access to supplemental oxygen, likely contributing to hundreds of thousands of deaths annually.

Objective: Motivated by an initiative to generate prehospital oxygen access in Balaka, Malawi, this study introduces a novel method to estimate oxygen demand for seven rural health centers utilizing anonymous pulse-oximetry data.

Design/Methods: Pulse-oximeters were used to screen for hypoxemia in each health center during routine patient care. Data were automatically stored and later extracted, representing 100 days of pulse-oximeter use. Because patient IDs were not recorded, an automated tool was developed (MATLAB) to assign a Retroactive Artificial Patient ID (RAPID) to each recording, as multiple recordings may have originated from the same patient. This RAPID assignment tool was calibrated against a subset of manual IDs, assigned independently by two raters through data inference based on comparisons in time of acquisition, heart rate, and blood oxygen saturation levels. A custom ‘percent agreement’ metric quantified reliability between raters (Fig. 1). A third rater served as tiebreaker to resolve disparities. The RAPID assignment tool was calibrated against this manually labeled subset, optimizing for percent agreement. Following calibration, all recordings received a RAPID assignment, enabling calculation of hypoxemia incidence and prevalence among those screened.

Results: The RAPID tool achieved 96.6% agreement with the calibration dataset. Applied to the full dataset, our tool estimated an average of 1.67 patients with hypoxemia per health center per day, with a hypoxemia prevalence of 21.1% amongst the 3,241 patients identified. Excluding two outlier health centers, the average incidence and prevalence adjusted to 0.60 and 9.4%, respectively.

Conclusion(s): This study offers valuable insights into prehospital oxygen needs in regions where hypoxemia incidence is seldom measured and oxygen rarely utilized. It does so using methods that minimize costs and personnel demands to optimize ease of replicability; future work would ideally address validity of the RAPID assignment tool against known patient data. More broadly, these methods demonstrate proof of concept that valuable public health or supply chain insights may be derived from routine and even automated sources, expanding the scope for cost-effective healthcare data collection in regions lacking reliable patient health registries.

Figure 1. Percent Agreement
Conceptual illustration of custom ‘Percent Agreement’ metric comparing two sample ID assignment lists (A & B). Each index represents one pulse-oximeter recording. Green indicates agreement, orange disagreement, at given indices.

Figure 1. Percent Agreement
Conceptual illustration of custom ‘Percent Agreement’ metric comparing two sample ID assignment lists (A & B). Each index represents one pulse-oximeter recording. Green indicates agreement, orange disagreement, at given indices.