Covid severity prediction: Who cares about the data quality?

COVID-19 is an ongoing pandemic disrupting daily life and overwhelming the healthcare infrastructure. Since the outburst of the pandemic, researchers have used various techniques to predict many aspects of the disease, including mortality rate and severity. The reproducibility of this research is challenging due to varying methodologies used to collect data, data quality, vague description of methodological approach to training prediction models, over-relying on data imputation, and over-fitting. This paper focuses on these challenges and provides a short yet comprehensive review of research on COVID mortality and severity prediction. The emphasis is on the reproducibility of the results and data quality issues. To further elaborate on the issue, we report the development of severity prediction models using two data sets. CRISP-DM is used as a methodological approach. We analyze and criticize the quality of the used data sets and how they affect the performance and limitations of the trained models. We conclude this paper with comments on data quality issues, the importance of reproducibility, and suggestions to improve reproducibility.