Descriptive characteristics of continuous oximetry measurement in moderate to severe COVID-19 patients

Abstract

Background: Non-invasive oxygen saturation (SpO2) measurement is a central vital sign that supports the management of COVID-19 patients. However, reports on SpO2 characteristics (patterns and dynamics) are scarce and none, to our knowledge, has analysed high resolution continuous SpO2 in COVID-19. Methods: SpO2 signal sampled at 1Hz and clinical data were collected from COVID-19 departments at the Rambam Health Care Campus (Haifa, Israel) between May 1st, 2020 and February 1st, 2021. Data from a total of 367 COVID-19 patients, totalling 27K hours of continuous SpO2 recording, could be retrieved, including 205 non-critical and 162 critical cases. Desaturations based on different SpO2 threshold definitions and oximetry derived digital biomarkers (OBMs) were extracted and compared across severity and support levels. Findings: An absolute SpO2 threshold at 93% was the most efficient in discriminating between critical and non-critical patients without support or under oxygen support. Under no support, the non-critical group depicted a fold change (FC) of 1,8 times more frequent desaturations compared to the critical group. However, the hypoxic burden was 1,6 times more important in critical versus non-critical patients. Other OBMs depicted significant differences, notably the percentage of time below 93% SpO2 (CT93) was the most discriminating OBM. Mechanical ventilation depicted a strong effect on SpO2 by significantly reducing the frequency (1,85 FC) and depth (1,21 FC) of desaturations. OBMs related to periodicity and hypoxic burden were markedly affected up to several hours before the initiation of the mechanical ventilation. Interpretation: This is the first report investigating continuous SpO2 measurements in hospitalized patients affected with COVID-19. SpO2 characteristics differ between critical and non-critical patients and are impacted by the level of support. OBMs from high resolution SpO2 signal may enable to anticipate clinically relevant events, monitoring of treatment response and may be indicative of future deterioration.

Building Trust: Lessons from the Technion-Rambam Machine Learning in Healthcare Datathon Event

“…Four challenges proposed by clinicians who had collected large datasets in recent years and who presented challenging scientific questions which could be tackled by ML were selected. The projects were 1) Prediction of newborn birthweight by maternal parameters and previous newborn siblings birthweights [15], 2) ML-based predictive model for bloodstream infections during hematopoietic stem cell transplantation [16], 3) Prediction of recurrent hospitalization in heart failure patients [17] and 4) Risk factor and severity prediction in hospitalized Covid-19 patients [18,19]. Each challenge came from a different hospital department and was presented with a database of approximately 2 Gb data.…”