Applications of Deep Learning in Predicting the Risk of Metabolic Syndrome from Lifestyle and Behavioral Factors: A Scoping Review

Abstract

Background: Metabolic syndrome (MetS) affects roughly one quarter of the world’s adults and dramatically heightens cardiometabolic morbidity, mortality, and healthcare costs. Because the syndrome is heavily driven by modifiable lifestyle and behavioural factors, risk stratification that relies only on easily collected, non-invasive information would enable earlier, lower-cost intervention. Deep learning (DL) models are theoretically well-suited to capture the complex, non-linear interactions among these heterogeneous data types, yet the relevant evidence base remains diffuse. Objective: To map and critically appraise the ways in which DL has been applied to predict MetS risk from lifestyle and behavioural variables, identify the predictors and model designs that dominate current practice, and highlight gaps and opportunities for future work. Methodology: A PRISMA-ScR–guided search (PubMed, Google Scholar; 2010 – 2025) located empirical studies based on predefined inclusion and exclusion criteria. Seven (7) studies met all criteria. Data on setting, sample, predictor categories, DL design, validation approach, and performance metrics were charted and synthesised narratively. Results: Of the seven (7) studies included, most used feed-forward neural networks on cross-sectional cohorts (n = 468–70 370) in Iran, Mexico, Taiwan, and South Korea. Accuracies ranged from ~0.81 to 0.94; AUCs often surpassed 0.85 and peaked at 0.99 when polygenic risk scores were added. Waist circumference, BMI, and blood-pressure indices were the strongest predictors, while activity, diet, sleep, smoking, and alcohol intake improved performance. Conclusion: Despite small numbers and limited geographic reach, evidence suggests that DL can turn routine, non-invasive data into highly accurate MetS risk tools.