Martin Rune Hassan Hansen
Position:
Improved detection of diabetes mellitus among African adults using machine learning
Categories:
Fellows, Postdoc Fellows 2024
Location:
Steno Diabetes Center Aarhus
Purpose:
The objective of the project is to develop risk scores that can be used for detection of undiagnosed diabetes mellitus in the general adult population of Africa, using readily available information on e.g. gender, age, physical activity level, diet, blood pressure and weight status. We will develop a risk score suitable for self-administration, as well as risk scores suitable for administration by health workers.
Methods:
The outcome to be predicted is undiagnosed diabetes mellitus, defined as elevated fasting blood glucose (either fasting plasma glucose or fasting capillary blood glucose), with no self-reported previous diagnosis of diabetes mellitus, and no consumption of glucose-lowering drugs. We will use cross-sectional data from 43 population-wide surveys that were conducted as part of the World Health Organization (WHO) STEPS program. The surveys were conducted from 2003 to 2020 to monitor non-communicable diseases in 32 African countries and cover the entire continent. We have already received the data from the WHO, and no further data collection is necessary. 125,538 individuals participated in the STEPS surveys and can be assessed for undiagnosed diabetes (fasting blood glucose measured, no pregnant women). We will create risk scores using both regression-based (Lasso) and tree-based models (decision tree, random forests), and validate them by k-fold cross-validation. Performance will be evaluated using measures of calibration (predicted vs. observed risk of diabetes mellitus) and discrimination (Area Under the Receiver Operating Curve), and models will be assessed for algorithmic fairness. We will also compare the performance of the models with that of risk scores developed in other populations.
Perspectives:
A validated risk score for diabetes mellitus has the potential to considerably improve the management of diabetes mellitus in Africa, as it will allow targeted screening of high-risk individuals, thus reducing the cost of case-finding. The STEPS surveys were conducted in the general population and the final risk scores will be suitable for administration in the same setting.
The risk scores will be implemented in community-based intervention programs against diabetes mellitus, coordinated by the East African NCD Alliance in Kenya, Rwanda, Burundi, Uganda and Tanzania, and will also be disseminated to African health ministries and other stakeholders.