Predicting low oxygen in patients with acute COVID-19 infection isolating at home: a clinical prediction model

Comprehensive Summary

This research presented by Wu et al. examined the use of prediction models on anticipating silent hypoxia in COVID-19 patients isolating at home. In a prospective cohort study, adults with the confirmed respiratory illness were monitored at home across three hospitals in Canada via an app, COVIDFree@Home. Factors such as SpO2, symptoms, temperature, and heart rate were reported by patients twice daily for approximately a ten day period. Random forest model and SHAP values were utilized to estimate the most significant predictors in oxygen saturation. Results showed that a minority of the cohort developed silent hypoxia with a higher risk of hospitalization while key predictors were seniority, comorbidities, and changes in vital signs. The machine learning model displayed moderate predictive accuracy given that some patients experienced ‘happy hypoxia.’ Wu et al. acknowledged the limitation of patient compliance on data collection and a singular healthcare system on geographical diversity.

Outcomes and Implications

Artificial intelligence can help prevent life-threatning escalations of a disease by promoting physician awareness of a patient's at home health status. However, given that some respiratory illnesses can be asymptomatic, clinical evaluation still remains the primary determinant of health and associated complications. Machine learning model application to potentially at-risk patients could be the next revolutionary step in identifying which individuals need closer observation and faster care.