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

Area of Science:

• Medical Informatics
• Epidemiology
• Clinical Trials

Background:

• Accurate prediction of COVID-19 mortality is essential for patient management.
• Machine learning (ML) offers advanced analytical capabilities for identifying risk factors and predicting outcomes.
• Comparing ML models with traditional statistical methods is vital for understanding their respective strengths.

Purpose of the Study:

• To compare risk factor identification and 90-day mortality prediction between two ML models and a non-ML model.
• To investigate the validity of ML models in an external cohort of hospitalized COVID-19 patients.
• To assess the performance of ML models in predicting clinical outcomes.

Main Methods:

• Utilized prospectively collected data from two randomized controlled trials (RCTs): TICO/ACTIV-3 (derivation/internal validation) and ITAC (external validation).
• Employed an ML pipeline with classification and survival models for risk factor identification and outcome prediction.
• Compared risk factors identified by ML models against a traditional Cox model and evaluated model performance across cohorts.

Main Results:

• A total of 3204 hospitalized COVID-19 patients were included (2625 in TICO/ACTIV-3, 579 in ITAC).
• Identified overlapping and unique risk factors across ML and non-ML models, with five factors common to all.
• The ML model demonstrated good predictive performance in the derivation cohort and remained above chance level in the external validation cohort, despite a performance decrease.

Conclusions:

• Integrating ML and non-ML methods may enhance risk factor identification for COVID-19.
• External validation of ML models, even within RCTs, is critical for ensuring generalizability and reliability.
• The study highlights the potential of ML in clinical outcome prediction while emphasizing the need for rigorous validation.