An explainable machine learning framework for obesity classification with optimized support vector machines

Post-COVID obesity has emerged as a significant public health challenge in Dhaka, driven by a complex interplay of socio-economic, lifestyle, and environmental factors that were exacerbated during the COVID-19 pandemic through increased sedentary behavior, unhealthy dietary patterns, and reduced access to healthcare services. This situation has led to a growing number of individuals being medically classified as overweight and obese, thereby accelerating the prevalence of chronic diseases, increasing pressure on the healthcare system, and hindering public health progress in the city. To address this challenge, the present study introduces a novel analytical framework based on a Relative Error Support Vector Machine (RE-SVM) optimized using the Tasmanian Devil Optimiser (TDO), a recently developed metaheuristic particularly effective f, robust algorithm specifically designed transparency and derive actionable insights, eIn order to maintain transparency and also derive actionable insights, explainable AI (XAI) methodologies are incorporated to offer interpretable models that identify the major contributing factors of obesity within the urban fabric of Dhaka.182 The proposed model was trained and tested on a dataset based on the population of Dhaka city, comprising 2,182 records and 17 original attributes. In the full-feature setting, the optimized RE-SVM model achieved an accuracy of 99.06%, with precision, recall, and F1-score values of 0.9906. However, because obesity-level labels are closely related to BMI, which is directly determined by weight and height, an additional leakage-controlled ablation experiment was conducted by entirely removing Weight (kg) and Height (m) before model training. The revised analysis distinguishes between anthropometric-inclusive classification and non-anthropometric obesity-risk prediction. The most critical factor contributing to obesity in Dhaka is weight, particularly extreme weight ranges (< 58.00 kg or > 95.70 kg), followed by height and family history of overweight. These findings indicate that the proposed framework can identify predictive patterns in structured obesity-level data. However, because the dataset should not be interpreted as an independently collected population-representative Dhaka cohort, the results should be viewed as methodological evidence rather than direct epidemiological evidence. Locally collected and clinically validated Dhaka data are required before deriving targeted public health interventions from the model outputs.