Identifying prevalent multimorbidity and its correlates in a low- and middle-income country: a machine learning analysis

Multimorbidity, the coexistence of two or more chronic conditions, presents growing challenges for health systems in low- and middle-income countries. Neighbourhood socioeconomic conditions are widely assumed to contribute to multimorbidity risk beyond individual characteristics, but this contribution has not been quantified in sub-Saharan Africa. We estimated how much neighbourhood socioeconomic context adds to the identification of prevalent multimorbidity, over and above individual characteristics and place of residence, and compared machine learning with conventional regression for this task. We analysed cross-sectional data from 6,177 adults aged 15 years and older in 691 primary sampling units in the 2016 South Africa Demographic and Health Survey. Multimorbidity was defined as two or more of twelve chronic conditions ascertained by biomarker measurement, self-reported diagnosis or reported treatment. Predictors were classified by temporal status as structural, behavioural, or illness-concurrent; the primary model excluded illness-concurrent variables because current medication use is itself a criterion for ascertaining two of the outcome conditions. We compared 18 supervised algorithms with three regression comparators using repeated ten-fold cross-validation grouped by primary sampling unit, with sampling weights applied in the evaluation of all algorithms and in the fitting of those that accept observation weights. Discrimination was estimated from pooled out-of-fold predictions with cluster-bootstrap confidence intervals. We assessed subgroup sensitivity, false-negative rate and calibration across wealth, age, sex, residence and province. Multimorbidity affected 2,628 participants (weighted prevalence 40.0%). The primary model achieved an area under the curve of 0.776 (95% CI 0.760–0.793), sensitivity 0.599 (0.568–0.631) and specificity 0.785 (0.764–0.804). Adding neighbourhood socioeconomic indicators to a model already containing individual characteristics, province and urban/rural residence changed discrimination by 0.001 (95% CI -0.001 to 0.002); the neighbourhood-level intraclass correlation was 0.027. Machine learning offered no advantage over conventional regression: multilevel logistic regression with a primary-sampling-unit random intercept achieved an identical area under the curve of 0.776 (0.761–0.793), and the total spread across the five best-performing algorithms was 0.004. Including illness-concurrent predictors inflated discrimination by 0.031 (0.025–0.036). Although accuracy was similar across subgroups, sensitivity ranged from 0.00 in adults aged 15–24 years to 0.95 in those aged 65 and over. In a weighted analytic sample drawn from a nationally representative South African survey, neighbourhood socioeconomic context contributed negligibly to identifying prevalent multimorbidity once individual characteristics and place of residence were accounted for, and flexible machine learning offered no discriminative advantage over an appropriately specified multilevel regression. Reported performance of multimorbidity models is materially inflated by predictors that reflect existing disease. Because these data are cross-sectional, the model identifies prevalent rather than incident multimorbidity, and comparable subgroup accuracy conceals near-total failure to identify affected younger adults.