Network maps stable disease clusters in South African adults

A network analysis of South African adults found three broad clusters of chronic disease, including a notably stable grouping of tuberculosis, HIV and anemia, but also showed that complexity-based measures did not predict poor self-rated health any better than simply counting how many conditions a person had. The cross-sectional study used complete data from 5,650 adults in the nationally representative 2016 South Africa Demographic and Health Survey. Twelve conditions were represented as nodes in a network, with weighted connections indicating how often pairs occurred together. The researchers then tested whether the resulting clusters were reproducible, whether their structure differed by sex or urban-versus-rural residence, and whether network-derived summaries of an individual’s conditions improved upon a conventional disease count.

The three clusters divided the conditions into an infectious group of tuberculosis, HIV and anemia; a cardiometabolic and pain group comprising hypertension, diabetes and chronic pain; and a larger cardiovascular, respiratory and neoplastic group containing stroke, heart attack, high cholesterol, chronic bronchitis, asthma and cancer. Their stability was assessed in 1,000 bootstrap replicates that resampled primary sampling units within strata to respect the survey design. Conditions assigned to the same cluster in the main analysis co-clustered in an average of 89% of replicates, compared with 21% for cross-cluster pairs. The tuberculosis-HIV-anemia grouping was the most consistent: the tuberculosis-HIV and HIV-anemia relationships appeared together in every replicate. The infectious cluster and the ranking of conditions by eigenvector centrality also remained intact in sensitivity networks based on observed-to-expected co-occurrence ratios and survey-weighted phi coefficients.

Subgroup analyses indicated that the broader network was not identical across populations. The infectious cluster persisted in both men and women, as well as in urban and rural residents, but the organization of the other conditions varied. Among men, respiratory conditions clustered with chronic pain, metabolic conditions formed a separate group, and stroke, heart attack and cancer formed another. Among women, heart attack was grouped with hypertension, diabetes and chronic pain, while respiratory conditions clustered with stroke and high cholesterol. Permutation tests supported an overall difference between the male and female networks, with a summed absolute edge-weight difference of 2.67 and p = 0.001. Urban networks showed broader integration of metabolic, cardiac, respiratory and pain conditions, whereas rural networks separated the cardiometabolic-pain group from a group containing respiratory conditions, stroke, heart attack, high cholesterol and cancer. The urban-rural contrast was also supported by permutation testing, with a summed absolute edge-weight difference of 2.19 and p = 0.031.

Centrality analysis identified conditions that occupied especially connected positions in the network. Heart attack, chronic bronchitis, asthma and chronic pain each co-occurred with 10 of the other 11 conditions, giving each a degree centrality of 0.91. They also shared the highest eigenvector centrality value, 0.32, which reflects both the number of a condition’s connections and the connectedness of the conditions linked to it. Anemia had the highest betweenness centrality, 0.10, making it the condition most often positioned between different parts of the network. HIV occupied the opposite structural extreme: it had a degree centrality of 0.18, an eigenvector centrality of 0.06 and no betweenness, indicating that it connected with relatively few conditions and did not bridge the major clusters. These measures describe the organization of co-occurring conditions; because the study was cross-sectional, they cannot establish that centrally positioned conditions preceded or caused the accumulation of other illnesses.

The researchers also examined whether particular pairs were associated with poor self-rated health by more than the sum of their separate associations. Of 66 possible condition pairs, 47 had at least 20 participants with both conditions and were eligible for testing; 19 sparse pairs were suppressed. After false-discovery-rate correction, one pair remained statistically supported: stroke combined with anemia had an interaction contrast of +0.55 on the risk-difference scale, with a 95% confidence interval of 0.24–0.86 and an adjusted q-value of 0.027. The estimate was based on 22 participants with both conditions, and no other pair remained supported after correction, with all remaining q-values at or above 0.62. The researchers described this additive interaction descriptively and did not imply biological synergy. Rare combinations produced imprecise estimates, making the interaction analysis exploratory despite the one result surviving statistical correction.

Self-rated health was assessed with the survey question asking respondents whether their health was generally excellent, good, average or poor. For the primary analysis, poor health was contrasted with average, good or excellent health. Survey-weighted estimates showed that 7.7% of adults rated their health as poor, while 34.6% had two or more of the 12 conditions. In models adjusted for age, sex and residence, the strongest individual associations with poor health included stroke, with an odds ratio of 6.51, and heart attack, with an odds ratio of 3.14. Tuberculosis, asthma, high cholesterol, chronic pain and HIV were also associated with poor health after adjustment. Hypertension, diabetes, chronic bronchitis and cancer were not significantly associated with the outcome in those models. These were adjusted associations in cross-sectional data, not evidence that any condition caused respondents to rate their health poorly.

When the central comparison was placed on a common per-standard-deviation scale, disease count had an adjusted odds ratio of 1.66 for poor self-rated health, compared with 1.55 for the network score and 1.24 for maximum centrality. The network score was highly correlated with disease count, at r = 0.96, and the two predictors showed substantial collinearity when entered together. Model discrimination offered the clearest test of incremental value. Adding disease count to age, sex and residence raised the area under the receiver operating characteristic curve from 0.721 to 0.758. The network score alone produced an AUC of 0.752, while maximum centrality produced 0.729. Adding both network measures to disease count did not improve discrimination: the combined model had an AUC of 0.756, compared with 0.758 for disease count alone, and the difference was not significant by DeLong test, p = 0.26.

Ten-fold cross-validation, in which the network and individual-level measures were rebuilt within each training fold, reproduced the same ranking. Mean held-out AUC was 0.720 for covariates alone, 0.762 for disease count, 0.758 for network score and 0.738 for maximum centrality. A model containing disease count plus both network measures also achieved a mean held-out AUC of 0.762, confirming that the network summaries added no validated information beyond enumeration. Entering all 12 conditions individually produced the best performance, with an AUC of 0.769 and a cross-validated mean of 0.777. The result is a notable negative finding: a more elaborate representation of multimorbidity did not outperform the simpler measure for this outcome, even though the network remained useful for describing population-level patterns.

The study has several important limitations. Its cross-sectional design cannot establish causal relationships or reconstruct the order in which conditions developed. The analytic sample required complete ascertainment of all 12 conditions, which excluded 3,818 adults, largely because they were outside the survey’s biomarker subsamples; complete-case selection may therefore reduce representativeness even though the original survey used a national probability design. Mental health conditions were not measured, potentially understating overall multimorbidity complexity, while some diagnoses were self-reported and could be undercounted. Health behaviors, socioeconomic circumstances and access to care may have influenced both disease patterns and self-rated health. The outcome was also a single global self-assessment rather than a multidimensional measure of health-related quality of life. Rare pairs had limited statistical precision, and the network’s descriptive clusters varied in sensitivity to analytic choices, although the infectious grouping was consistently robust. Taken together, the findings support using disease counts as a pragmatic marker of poor self-rated health in this population, while treating network-based individual scores as research tools that require prospective and external validation before clinical adoption.

Subject of Research: Network analysis of chronic disease clustering and self-rated health among South African adults

Article Title: Beyond disease counts: multimorbidity patterns and self-rated health in South African adults–a network analysis of the 2016 demographic and health survey

Article References: Uthman, O. A., Hazell, M., Uthman, M. M. B., Wahab, K. W., Kengne, A. P., Gill, P., & Saravanan, P. (2026). Beyond disease counts: multimorbidity patterns and self-rated health in South African adults–a network analysis of the 2016 demographic and health survey. Frontiers in Public Health, 14, Article 1947329. https://doi.org/10.3389/fpubh.2026.1947329

Image Credits: AI Generated

DOI: 10.3389/fpubh.2026.1947329

Keywords: multimorbidity, network analysis, South Africa, self-rated health, chronic disease clustering, HIV, tuberculosis, anemia, non-communicable diseases, survey epidemiology, Beyond, disease