Personalized Oxygen Saturation Targets Compared to the Oxygen Saturation Experienced in Care Among Mechanically Ventilated Adults

RATIONALE: For critically ill adults receiving invasive mechanical ventilation, randomized trials have found no significant average treatment effect of a higher (96-100%) versus lower (88-92%) peripheral oxygen saturation (SpO2) target. The effect of SpO2 targets on outcomes, however, may differ for patients with different characteristics. Machine learning methods have been used recently to derive and validate a statistical model capable of predicting which SpO2 target will result in lower mortality for an individual patient, as a personalized SpO2 target. Personalized targets can only improve patient outcomes, however, if the SpO2 target predicted to be best for them differs from the SpO2 values they are already experiencing in care. Whether the SpO2 values experienced in clinical care differ for patients predicted to benefit from a higher vs a lower SpO2 target is unknown. METHODS: Among consecutive patients receiving invasive mechanical ventilation in an intensive care unit, we used the previously validated machine learning model to calculate a personalized SpO2 target (predicted to benefit from a higher SpO2 target (96-100%) vs predicted to benefit from a lower SpO2 target (88-92%)) from baseline characteristics for each patient. We compared the personalized SpO2 target as the primary exposure to the observed SpO2 values experienced in clinical care as the primary outcome using a proportional odds model accounting for within-subject correlation. FiO2 values received in clinical care were analyzed as the secondary outcome. MEASUREMENTS AND MAIN RESULTS: Among 615 patients (median age, 58 years; 41% female), 325 (53%) were predicted by the machine learning model to benefit from a higher SpO2 target and 290 (47%) were predicted to benefit from a lower SpO2 target. Patients predicted to benefit from higher and lower targets experienced similar SpO2 values (mean 96.2% vs 95.6%; median 97% vs 96%; adjusted median difference 0.3%; 95%CI: -0.1% to 0.6%; P = 0.18) and FiO2 values (mean 0.45 vs 0.47; median 0.40 vs 0.40; adjusted median difference -0.02; 95%CI: -0.07 to 0.03; P = 0.38) in clinical care. In subgroups of interest, differences in the SpO2 or FiO2 values between groups were found only among patients with shock. CONCLUSIONS: In current clinical care, patients do not experience SpO2 values consistent with those predicted to result in the best outcomes for them, based on a machine learning model derived and validated in data from prior clinical trials. This lack of difference in care suggests that using evidence-based personalized oxygen saturation targets to guide oxygen administration in clinical care has the potential to improve outcomes.