Sub-population identification of multimorbidity in sub-Saharan African populations

Abstract

This work provides three contributions that straddle the medical literature on multimorbidity and the data science community with an interest on exploratory analysis of health-related research data. First, we propose a definition for multimorbidity as the co-occurrence of (at least) two disease diagnoses from a pre-determined list. This interpretation adds to a growing body of working definitions emerging from the literature. Second, we apply this novel outcome of-interest to two sub-Saharan populations located in Nairobi, Kenya and Agincourt, South Africa. The source data for this analysis was collected as part of the Africa Wits-INDEPTH Partnership for Genomic Studies project. Third, we stratify this outcome-of-interest across all possible sub-populations and identify sub-populations with anomalously high (or low) rates of multimorbidity. Critically, the automatic stratification approach emphasizes efficient, disciplined exploratory-based analysis as a complementary alternative to more commonly-used confirmation analysis methods. Our results show that high-risk sub-populations identified in one part of the continent transfer to the other location (and vice-versa) with the equivalent sub-population at the other location also experiencing higher rates of multimorbidity. Second, we discover a real-world scenario where a more-at risk sub-population existed beyond the simpler sub-populations traditionally stratified by age and sex. This is in contrast to existing literature which commonly stratifies disease diagnoses by sex when reporting results. Patterns in diseases, and healthcare more generally, are likely more nuanced than manual approaches may be able to describe. This work helps introduce public health researchers to data science methods that scale to the size and complexity of modern day datasets.

Keywords: Africa; Exploratory analysis; Multimorbidity; Subset scanning; Survey data.

Fig. 1

Flowchart illustrating the analysis pathways of manual and automatic stratification methods. Both approaches have the same initial data and end-goals. However, automatic stratification uses additional data science techniques to bring discipline and scalability to the otherwise slower manual approach. Also, manual stratification is more confirmatory in nature in that it measures evidence for a proposed hypothesis. Automatic stratification is more exploratory-driven in nature in that it finds hypothesis backed with the most evidence from the data. All analysis methods must recognize this balance.

Fig. 2

Upset Plots illustrating the co-occurrence of four health conditions from Agincourt and Nairobi.

Fig. 3

Venn Diagram for the sub-populations discovered by automatic stratification in Agincourt as described in Table 1. The left image is for sub-populations with High Risk Status and the right is for sub-populations with Low Risk Status. Each value in the Venn diagram represents the count of individuals in that section, the percentage of multimorbidity within that count in parenthesis, and the percentage of women in that subgroup is in square brackets. The blue text at the upper right corner describes the number of people that are not in any of the sub-populations participating in the Venn diagram and the percentage of multimorbidity within those.

Fig. 4

Venn Diagram for the sub-populations discovered by automatic stratification in Nairobi as described in Table 2. The left image is for sub-populations with High Risk Status and the right is for sub-populations with Low Risk Status. Each value in the Venn diagram represents the count of individuals in that section, the percentage of multimorbidity within that count in parenthesis, and the percentage of women in that subgroup is in square brackets. The blue text at the upper right corner describes the number of people that are not in any of the sub-populations participating in the Venn diagram and the percentage of multimorbidity within those.