Machine learning models aimed at identifying risk factors for reducing morbidity and mortality still need to consider confounding related to calendar time variations

Abstract Machine learning models applied to health data may help health professionals to prioritize resources by identifying risk factors that may reduce morbidity and mortality. However, many novel machine learning papers on this topic neither account for nor discuss biases due to calendar time variations. Often, efforts to account for calendar time (among other confounders) are necessary since patterns in health data – especially in low- and middle-income countries – may be influenced by calendar time variations such as temporal changes in risk factors and changes in the disease and mortality distributions over time (epidemiological transitions), seasonal changes in risk factors and disease and mortality distributions, as well as co-occurring artefacts in data due to changes in surveillance and diagnostics. Based on simulations, real-life data from Guinea-Bissau, and examples drawn from recent studies, we discuss how including calendar time variations in machine learning models is beneficial for generating more relevant and actionable results. In this brief report, we stress that explicitly handling temporal structures in machine learning models still remains to be considered (like in general epidemiological studies) to prevent resources from being misdirected to ineffective interventions.