BackgroundThe rapid spread of coronavirus disease COVID 19 calls for early screening and monitoring of these patients to distinguish those that are likely to worsen from stable patients that may be directed to intermediate care facilities. We designed a score for COVID-19 patients severity assessment, dynamic intubation and prolonged stay prediction using the Breathing Frequency (BF) and oxygen saturation (SPO2) signals.MethodsWe recorded BF, and SPO2 signals of confirmed COVID-19 patients admitted during the first and second outbreak of the pandemic in France (March to May 2020 and September 2020 to February 2021) in an ICU of a teaching hospital. We extracted four features from the signals that represent the four last hours before intubation for intubated patients and the mean of the four hours before the median intubation time for non-intubated patients. These data were used to train AI algorithms for intubation recognition. Algorithm robustness was checked on a validation set of patients. We selected the best algorithm that was applied every hour to predict intubation, thus a severity evaluation. We performed a 24h moving average of these predictions giving a S24 severity score that represent the patient's severity during the last 24 h. MS24, the maximum of S24 was confronted with the risk of intubation and prolonged ICU stay (>5 days).ResultsWe included 177 patients. Among the tested algorithms, the Logistic regression classifier had the best performance. The model had an accuracy of 88.9 % for intubation recognition (AUC=0.92). The accuracy on the validation set was 92.6 %. The S24 score of intubated patients was significantly higher than non-intubated patients 48h before intubation and increased 24 hours before intubation. MS24 score allows distinguishing three severity situations with an increased risk of intubation: green (3%), orange (30%) and red (76%). A MS24 score superior to 20 was highly predictive of an ICU stay greater than 5 day with an accuracy of 88.8% (AUC=0.95).ConclusionsThe score we designed uses simple signals and seems to be efficient to visualize the patient's respiratory situation and may help in decision-making. Real-time computation is easy to implement.

No abstract

No abstract