Multimorbidity representation via graph learning: A population-based study on hepatosplenic conditions in schistosomiasis-endemic areas of rural Uganda

BackgroundThe global burden of multimorbidity is increasing yet poorly understood, owing to insufficient methods available for modelling complex systems of conditions. In particular, hepatosplenic multimorbidity has been inadequately investigated. MethodsFrom 17 January to 16 February 2023, we examined 3186 individuals aged 5-92 years from 52 villages across Uganda within the SchistoTrack Cohort. Point-of-care B-mode ultrasound was used to assess 45 hepatosplenic conditions. Three graph learning methods for representing hepatosplenic multimorbidity were compared including graphical lasso (GL), signed distance correlations (SDC), and co-occurrence. Graph kernels were used to identify thresholds of relevant condition inter-dependencies (edges). Graph neural networks were applied to validate the quality of the graphs by assessing their predictive performance. Clinical utility was assessed through medical expert review. FindingsMultimorbidity was observed in 54{middle dot}65% (1741/3186) of study participants, who exhibited two or more hepatosplenic conditions. Conditions of mildly fibrosed vessels were most frequently observed (>14% of individuals). Percentage thresholds were found to be 50{middle dot}16% and 64{middle dot}46% for GL and SDC, respectively, but could not be inferred for co-occurrence. Thresholded GL and SDC graphs had densities of 0{middle dot}11 and 0{middle dot}17, respectively. Both thresholded graphs were similar in predictive utility, although GL produced marginally higher AUCs under certain experiments. Both GL and SDC had significantly higher AUCs than co-occurrence. Numerous conditions were predicted with perfect sensitivity using both GL and SDC with graph convolutional network with five input conditions. InterpretationThe most common method for multimorbidity (co-occurrence) provided an uninformative representation of hepatosplenic conditions with respect to sparsity and predictive performance. More clinically useful graphs were computed when algorithms consisted of statistical assumptions, such as graphical lasso. Future work could apply the pipeline developed here for clinically relevant multimorbidity representations. FundingNDPH Pump Priming Fund, John Fell Fund, Robertson Foundation, UKRI EPSRC (EP/X021793/1).