Which Features Are Useful in Machine Learning-Based COVID-19 Prognostication? A Meta-Analysis

COVID-19 brings devastating impacts on society and shows how fragile our current healthcare system is. To aid the healthcare system, machine learning (ML) models have shown strong potential in combating COVID-19 across several areas, including diagnosis, complication prediction, mortality risk prediction, and severity assessment. Feature selection is an essential step in developing efficient machine-learning models. In this research, we screened 143 COVID-19 research articles, where 26 articles had enough data to calculate the effect size. We used Cohen’s d to calculate the effect size of the features, which reveals several highly significant features yet explored in only a few studies and several less significant features yet explored in many studies. Lymphocyte (%) and hypersensitive c-reactive protein are comparatively less used features but showed a very large effect size. On the other hand, lymphocyte count, c-reactive protein, D-dimer, and creatinine are utilized in several studies but are not necessarily significant. Considering these high-effect-sized features and discarding the low-effect-sized features would benefit future ML model developments for combating COVID-19.