When to and when not to use machine learning in risk prediction models

Machine learning for health data science, fuelled by proliferation of data and reduced computational costs, has garnered considerable interest among researchers. The debate around the use of machine learning or classical medical statistics for medical research is long standing.1,2 Given major methodological advances and increased policy attention, renewed consideration is warranted for when machine learning should and should not be used in risk prediction models. In this Comment, we offer practical guidance on when (and when not) to use machine learning for risk prediction.