Imputation of PaO2 from SpO2 values from the MIMIC-III Critical Care Database Using Machine-Learning Based Algorithms

Background The partial pressure of oxygen (PaO2)/fraction of oxygen delivered (FIO2) ratio is the reference standard for assessment of hypoxemia in mechanically ventilated patients. Non-invasive monitoring with the peripheral saturation of oxygen (SpO2) is increasingly utilized to estimate PaO2 because it does not require invasive sampling. Several equations have been reported to impute PaO2/FIO2 from SpO2 /FIO2. However, machine-learning algorithms to impute the PaO2 from the SpO2 has not been compared to published equations. Research Question How do machine learning algorithms perform at predicting the PaO2 from SpO2 compared to previously published equations? Methods Three machine learning algorithms (neural network, regression, and kernel-based methods) were developed using 7 clinical variable features (n=9,900 ICU events) and subsequently 3 features (n=20,198 ICU events) as input into the models from data available in mechanically ventilated patients from the Medical Information Mart for Intensive Care (MIMIC) III database. As a regression task, the machine learning models were used to impute PaO2 values. As a classification task, the models were used to predict patients with moderate-to-severe hypoxemic respiratory failure based on a clinically relevant cut-off of PaO2/FIO2 ≤ 150. The accuracy of the machine learning models was compared to published log-linear and non-linear equations. An online imputation calculator was created. Results Compared to seven features, three features (SpO2, FiO2 and PEEP) were sufficient to impute PaO2/FIO2 ratio using a large dataset. Any of the tested machine learning models enabled imputation of PaO2/FIO2 from the SpO2/FIO2 with lower error and had greater accuracy in predicting PaO2/FIO2 ≤ 150 compared to published equations. Using three features, the machine learning models showed superior performance in imputing PaO2 across the entire span of SpO2 values, including those ≥ 97%. Interpretation The improved performance shown for the machine learning algorithms suggests a promising framework for future use in large datasets.