Predicting Multimorbidity in 50,000 People Living with Obesity: A Machine Learning Model Applied to Understand Obesity Progression in Two Health Care Systems

Line Egerod, Rikke Linnemann Nielsen, Zahra McVey, Joseph Katigbak, Thomas Monfeuga, Frederik Steensgaard Gade, August T. H. Schreyer, Luis G. Leal, William G. Haynes, Alex Greenfield, Ella Nkhoma, Robert R. Kitchen, Michael L. Wolden, Kasper S. Matthiessen, Laurent Gautier, Abd A. Tahrani, Ramneek Gupta

Preprint at SSRN (2025)

DOI: 10.2139/ssrn.5095147

Abstract

Background: Obesity is a major driver of multimorbidity. There is a clinical and societal need to identify and target individuals at high risk of future multimorbidity, to reduce the health and economic burdens.

Methods: We conducted a retrospective study in 50,804 people living with obesity (PLwO) in two real-world cohorts (UK Biobank (UKB), n=11,646; Clalit Health Services (CHS), n=39,158). We mapped the progression to 15 obesity-related comorbidities (ORCs) over a ten-year follow-up period. Machine learning (ML) was used to identify prognostic biomarkers of obesity-related multimorbidity.

Findings: We characterised the timing and order of ORCs in PLwO. The longest prevention window was before development of the first ORC, which was most commonly hypertension. We developed an ML model integrating clinical, lifestyle, socioeconomic, and genetic data from UKB, with robust prediction of multimorbidity (area under the receiver operating characteristic curve (ROC-AUC: 0·74), superior to individual adiposity or genetic risk scores (ROC-AUC: 0·57–0·62). The top prognostic biomarkers included age, haemoglobin A1c (HbA1c), self-reported fair health rating, waist-to-height ratio (WHtR), genetic risk of hypertension and ischaemic stroke, smoking and blood pressure. Using only parameters easily attainable in clinical practice, robust predictive performance was maintained (ROC-AUC: 0·71). Finally, we replicated these findings, and the predictive capabilities of the top prognostic biomarkers, in the CHS cohort.

Interpretation: These tools may influence obesity management and the prevention of multimorbidity and emphasise the relevance of prognostic biomarkers beyond BMI (e.g. WHtR). Replication of these prognostic biomarkers in a diverse population supports the generalisability and utility of these findings in addressing the burden of multimorbidity.