Vaid stories no related monetary disclosures. Please see the research for all different authors’ related monetary disclosures.
For sufferers hospitalized with COVID-19, a boosted determination tree mannequin extra precisely predicted AKI requiring dialysis and mortality in contrast with both normal fashions or different machine studying fashions.
The mannequin, often called XGBoost, confirmed greater precision (outlined by constructive predictive worth) and recall (outlined by sensitivity) than the opposite fashions, which researchers contended is necessary as a result of “a mannequin with excessive constructive predictive worth minimizes false positives and thus may help keep away from clinician fatigue and alert burnout … [while] excessive sensitivity means the mannequin can have fewer false negatives, thus maximizing the utility in figuring out sufferers at want for dialysis or in danger for loss of life.”
Based on Akhil Vaid, MD, of Icahn College of Medication at Mount Sinai, and colleagues, though a number of machine studying fashions have been revealed all through the COVID-19 pandemic, none have particularly addressed AKI and dialysis.
“Machine studying fashions can harness the disparate knowledge collected throughout medical care in digital well being information for correct final result predictors,” they wrote. “ … We aimed to develop and validate a machine studying mannequin to foretell a composite endpoint of AKI handled with dialysis or loss of life in sufferers hospitalized with COVID-19 early within the hospital course.”
To this finish, researchers included 6,093 sufferers who had been hospitalized throughout the Mount Sinai Well being System between March 2020 and December 2020. Researchers in contrast the efficiency between logistic regression, least absolute shrinkage and choice operator, random forest and XGBoost for predicting dialysis or mortality at 1 to 7 days following hospital admission. All fashions thought-about demographics, comorbidities, and laboratory and important indicators inside 12 hours of hospital admission.
Outcomes demonstrated XGBoost (with out imputation for prediction of a composite final result of both loss of life or dialysis) carried out higher than the opposite fashions, with the best space underneath the receiver curve on inner validation (vary of 0.93 to 0.98) and space underneath the precision recall curve (vary of 0.78 to 0.82 throughout) forever factors. XGBoost with out imputation additionally outperformed all fashions via its greater precision and recall (imply distinction within the space underneath the curve receiver working attribute of 0.04 and imply distinction within the space underneath precision recall curve of 0.15).
Serum creatinine, blood urea nitrogen, systolic blood strain, age and oxygen saturation had been the “main drivers” of the mannequin’s predictions.
“In conclusion, identification of sufferers in danger for acute dialysis and loss of life in COVID-19 presents quite a lot of challenges,” Vaid and colleagues wrote. “One such issue pertains to useful resource allocation in doubtlessly overcrowded hospitals. Our fashions might help with this problem and are being prospectively validated and deployed in a real-world setting to aide in administration of hospitalized sufferers with COVID-19.”