Abstract
Non-invasive oxygen saturation (SpO2) is a central vital sign that supports the management of COVID-19 patients. However, reports on SpO2 characteristics are scarce and none has analysed high resolution continuous SpO2 in COVID-19. We provide the first analysis of high resolution SpO2 across the spectrum of COVID-19 disease severity and respiratory support. A total of 367 COVID-19 patients’ recordings, comprising 27K hours of continuous SpO2 data, could be retrieved from patients hospitalized at the Rambam Health Care Campus. Using oximetry digital biomarkers (OBM), we quantified SpO2 characteristics and showed that the percentage of time under 93% oxygen saturation threshold is the best single OBM discriminating between critical and non-critical patients. OBMs traditionally used in the field of sleep medicine research, were informative for assessing the patient’s response to respiratory support. In addition, periodicity and hypoxic burden biomarkers were affected up to several hours before the initiation of the mechanical ventilation. Characteristics from high resolution SpO2 signal may enable to anticipate clinically relevant events, monitoring of treatment response and may be indicative of future deterioration.
Introduction
In the clinical management of COVID-19, blood oxygen saturation (SpO2) serves as the primary "north star" for decision-making. Despite its importance, most hospital protocols rely on intermittent "spot checks" or single values recorded in electronic health records. These discrete snapshots often miss the underlying dynamics of a patient's respiratory health—the "movie" of fluctuations, desaturation depth, and frequency. This study addresses the gap between raw, high-resolution monitor data and actionable clinical insights by identifying specific patterns that correlate with patient deterioration.
Methods
Researchers conducted a retrospective cohort study at the Rambam Health Care Campus, analyzing 27,000 hours of 1Hz (one sample per second) SpO2 data from 367 adult patients (162 critical, 205 non-critical). To quantify these signals, the team utilized Oximetry Digital Biomarkers (OBMs), mathematical features traditionally used in sleep medicine to study apnea. These OBMs were categorized into five domains: general statistics, signal complexity (e.g., Approximate Entropy), periodicity (frequency domain analysis), desaturation measures (depth and duration of "dips"), and hypoxic burden (area under the curve below specific thresholds).
Results
The analysis revealed several key physiological signatures that distinguish patient severity and predict clinical transitions.
The 93% Threshold and Hypoxic Burden
While international guidelines often use 90% as a threshold for severity, this study found that 93% (CT93) was the most discriminative absolute threshold for identifying critical patients. Interestingly, non-critical patients showed more frequent desaturations, but critical patients experienced significantly deeper and longer "hypoxic dives." Critical patients without support exhibited a significantly higher hypoxic burden (p < 0.001) compared to non-critical patients, with a Fold Change (FC) of 1.6.
Signal Complexity and Early Warning Signs
As patients deteriorated, the "complexity" or entropy of their SpO2 signal decreased. While a healthy respiratory system displays variability and "noise," a failing system becomes more rhythmic and predictable. Crucially, shifts in periodicity and hypoxic burden biomarkers were detectable several hours before the clinical initiation of mechanical ventilation, suggesting a vital window for early intervention.
Impact of Mechanical Ventilation
While mechanical ventilation successfully stabilizes SpO2 levels, the data indicated a tendency toward "over-oxygenation," with prolonged periods at 100% saturation. This highlights the potential for continuous monitoring to help clinicians optimize ventilator settings and avoid physiological detriment.
Discussion and Limitations
High-resolution oximetry provides a rich dataset that intermittent checks cannot capture. Integrating OBMs into AI-driven monitoring systems could provide clinicians with a "countdown" to respiratory failure, allowing for proactive adjustments in treatment.
However, the study is limited by its single-center, retrospective design. The findings require validation across diverse populations and hospital systems. Additionally, the study did not fully account for specific ventilator settings (e.g., PEEP or FiO2) that influence the SpO2 signal. Future work should focus on prospective trials to determine if real-time clinical action based on these digital biomarkers directly improves patient outcomes.
Summary updated