The urgency of the COVID-19 pandemic has led to many machine learning projects trying to help doctors. Now, researchers at Mount Sinai created an algorithm that looks at patients’ hospital data and predicts whether they’ll need invasive ventillation a.k.a. intubation.
The dataset
Five hospitals pooled together their data to train the model. Demographic, vitals, and laboratory data were retrospectively queried from electronic health records of 4087 patients. The index period was between February 2020 and April 2020. Given the novelty of COVID-19, researchers used all available data to maximize available training info for the model.
The model
Researchers developed this supervised-learning binary classification model with an aim of constantly supervising patient data. In this way, it simulated a real physician assistant that had access to patient information as they become available. “Alerts” were sent out the moment patient data became indicative of pending intubation. Indeed, only 20% of “alerted’ remained intubation-free during their hospital stay, compared with 65% of patients that received no alerts.
As it stands, the best tool physicians have to predict intubation, apart from clinical gestalt, is the ROX index. Using simple data about patients’ oxygen status, this index has shown potential in the COVID-19 population. The Mount Sinai model actually outperformed it by quite a margin. This should hardly come as a surprise – the ML model has access to far more data than the ROX Index. That said, the findings of this study are encouraging. A prospective study to assess whether actually using this model to direct treatment is beneficial to patients is the next step.
The future
As per the authors’ own admission, “COVID-19 …[is] a moving target with ongoing changes in clinical guidelines and even virus biology.” Frequent retraining of machine learning models may be necessary to stay up to date with changes in clinical practice and disease progression. That said, this paper is another indication that machine learning has a potential to help physicians with COVID-19 therapy efficacy, efficiency and prioritization.