Learning From Past Respiratory Infections to Predict COVID-19 Outcomes: Retrospective Study.

摘要: Background For the clinical care of patients with well-established diseases, randomized trials, literature, and research are supplemented judgment to understand disease prognosis inform treatment choices. In void created by a lack experience COVID-19, artificial intelligence (AI) may be an important tool bolster decision making. However, data restricts design development such AI tools, particularly in preparation for impending crisis or pandemic. Objective This study aimed develop test feasibility "patients-like-me" framework predict deterioration COVID-19 using retrospective cohort similar respiratory diseases. Methods Our used COVID-19-like cohorts train models that were then validated on population. The included diagnosed bacterial pneumonia, viral unspecified influenza, acute distress syndrome (ARDS) at academic medical center from 2008 2019. total, 15 training different combinations ARDS exploratory purposes. this study, two machine learning developed: one invasive mechanical ventilation (IMV) within 48 hours each hospitalized day, all-cause mortality time admission. Model performance was assessed area under receiver operating characteristic curve (AUROC), sensitivity, specificity, positive predictive value, negative value. We established model interpretability calculating SHapley Additive exPlanations (SHAP) scores identify features. Results Compared (n=16,509), (n=159) significantly younger, higher proportion Hispanic ethnicity, lower smoking history, fewer comorbidities (P 0.90). Validating cohort, top-performing predicting IMV XGBoost (AUROC=0.826) trained pneumonia cohort. Similarly, all 4 without achieved best (AUROC=0.928) mortality. Important predictors demographic information (age), vital signs (oxygen saturation), laboratory values (white blood cell count, cardiac troponin, albumin, etc). had class imbalance, which resulted high low values. Conclusions provided feasible modeling patient existing technology address limitations during onset novel, rapidly changing