Predicting invasive mechanical ventilation in COVID 19 patients: A validation study.

INTRODUCTION: The decision to intubate and ventilate a patient is mainly clinical. Both delaying intubation (when needed) and unnecessarily invasively ventilating (when it can be avoided) are harmful. We recently developed an algorithm predicting respiratory failure and invasive mechanical ventilation in COVID-19 patients. This is an internal validation study of this model, which also suggests a categorized "time-weighted" model. METHODS: We used a dataset of COVID-19 patients who were admitted to Rabin Medical Center after the algorithm was developed. We evaluated model performance in predicting ventilation, regarding the actual endpoint of each patient. We further categorized each patient into one of four categories, based on the strength of the prediction of ventilation over time. We evaluated this categorized model performance regarding the actual endpoint of each patient. RESULTS: 881 patients were included in the study; 96 of them were ventilated. AUC of the original algorithm is 0.87-0.94. The AUC of the categorized model is 0.95. CONCLUSIONS: A minor degradation in the algorithm accuracy was noted in the internal validation, however, its accuracy remained high. The categorized model allows accurate prediction over time, with very high negative predictive value.

A novel machine learning model to predict respiratory failure and invasive mechanical ventilation in critically ill patients suffering from COVID-19.

In hypoxemic patients at risk for developing respiratory failure, the decision to initiate invasive mechanical ventilation (IMV) may be extremely difficult, even more so among patients suffering from COVID-19. Delayed recognition of respiratory failure may translate into poor outcomes, emphasizing the need for stronger predictive models for IMV necessity. We developed a two-step model; the first step was to train a machine learning predictive model on a large dataset of non-COVID-19 critically ill hypoxemic patients from the United States (MIMIC-III). The second step was to apply transfer learning and adapt the model to a smaller COVID-19 cohort. An XGBoost algorithm was trained on data from the MIMIC-III database to predict if a patient would require IMV within the next 6, 12, 18 or 24 h. Patients' datasets were used to construct the model as time series of dynamic measurements and laboratory results obtained during the previous 6 h with additional static variables, applying a sliding time-window once every hour. We validated the adaptation algorithm on a cohort of 1061 COVID-19 patients from a single center in Israel, of whom 160 later deteriorated and required IMV. The new XGBoost model for the prediction of the IMV onset was trained and tested on MIMIC-III data and proved to be predictive, with an AUC of 0.83 on a shortened set of features, excluding the clinician's settings, and an AUC of 0.91 when the clinician settings were included. Applying these models "as is" (no adaptation applied) on the dataset of COVID-19 patients degraded the prediction results to AUCs of 0.78 and 0.80, without and with the clinician's settings, respectively. Applying the adaptation on the COVID-19 dataset increased the prediction power to an AUC of 0.94 and 0.97, respectively. Good AUC results get worse with low overall precision. We show that precision of the prediction increased as prediction probability was higher. Our model was successfully trained on a specific dataset, and after adaptation it showed promise in predicting outcome on a completely different dataset. This two-step model successfully predicted the need for invasive mechanical ventilation 6, 12, 18 or 24 h in advance in both general ICU population and COVID-19 patients. Using the prediction probability as an indicator of the precision carries the potential to aid the decision-making process in patients with hypoxemic respiratory failure despite the low overall precision.