Comparing risk factors in severe COVID-19 using machine learning and non-machine learning methods: analysis from 2 international randomized controlled trials.

内容摘要

To compare differences in risk factors and 90-day mortality prediction from 2 machine learning (ML) models with a previously published non-ML model and investigate their validity in an external cohort. Prospectively collected data from 2 separate randomized controlled trial (RCT) cohorts from 2020 to 2021, the Therapeutics for Inpatients with COVID-19 (TICO/ACTIV-3) Trial (derivation and internal validation cohort) and the Inpatient Treatment with Anti-Coronavirus Immunoglobulin (ITAC) Trial (external validation cohort) were used. Data were collected from 114 sites in 10 countries (TICO/ACTIV-3) and 63 sites in 11 countries (ITAC). A ML pipeline including 5 classification models, and 1 survival model was used for risk factor identification and clinical outcome prediction. Risk factors were compared between a ML-based classification model, a ML-based survival model and a previously published Cox model. Performance of the ML-based classification model was compared across TICO/ACTIV-3 and ITAC. A total of 2625 (TICO/ACTIV-3) and 579 (ITAC) adults hospitalized for COVID-19 were included. Some overlap of risk factors was identified across models. Five were identified in all models, 3 only in ML models, and 4 only in the non-ML model. The ML model showed good predictive performance in TICO/ACTIV-3. Internal validation showed no overfitting. Lower model performance was observed in ITAC (-15.8%), but performance remained above chance level. Differences in methods for risk factor identification using ML and non-ML complicates the comparison of results derived from each approach, but using multiple approaches may unveil overlooked risk factors. Risk factor identification may benefit from integrating both ML and non-ML methods, but external validation is necessary, even in RCTs.