1Ph.D Researcher, Department of Computer Science, Kamban College of Arts and Science, Coimbatore.
2*Associate Professor, Department of Computer Applications, School of Science and Computer Studies, CMR University, Bengaluru.
3Assistant Professor, Department of Computer Science, Kamban College of Arts and Science, Coimbatore.
*Corresponding author: Janarthanam S
The Obesity prevalence in Tamil Nadu has escalated to 41.3% in males and 27.3% in females, representing a significant public health burden. Centralised data mining approaches for obesity prediction encounter substantial obstacles, including regulatory constraints, institutional data silos, and patient privacy concerns. This investigation presents a privacy-preserving federated deep learning framework for stratifying obesity risk across distributed hospital networks in Tamil Nadu using electronic health records, advancing data-driven public health surveillance while maintaining data sovereignty. A federated learning architecture was implemented across simulated multi-institutional healthcare networks representing diverse urban, semi-urban, and rural settings. The framework integrated data from 724,115 individuals derived from national health surveys, combined with synthesised clinical features. The architecture employed deep neural networks augmented with differential privacy mechanisms and cryptographically secure aggregation protocols. Obesity risk was stratified into four categories through pattern recognition and predictive modelling approaches inherent to advanced data mining. The federated model demonstrated superior performance metrics with 94.2% accuracy and AUC-ROC of 0.963, substantially exceeding centralised approaches (92.3%, AUC 0.948) and conventional federated averaging (91.8%, AUC 0.941). Model convergence occurred within 32 communication rounds compared to 45 rounds required for standard federated averaging. Feature importance analysis identified body mass index (18.5%), waist circumference (17.2%), and chronological age (14.8%) as principal predictive factors. Population-level risk stratification revealed 14.5% classified as very high risk and 24.7% as high risk, facilitating targeted clinical intervention strategies. Federated deep learning architecture successfully enables population-level obesity risk prediction while simultaneously preserving patient privacy and maintaining institutional data autonomy. This approach demonstrates considerable scalability potential for distributed health surveillance and evidence-based intervention programming across Tamil Nadu's decentralised healthcare infrastructure.
Keywords: Federated learning, Obesity prediction, Electronic health records, Deep learning, Risk stratification, Data mining, Differential privacy, Healthcare networks.
How to cite this article: Rajeswari R, Janarthanam S, Shanthakumar M. Privacy-Preserving Federated Deep Learning for Multi-Class Obesity Risk Stratification in Tamil Nadu Territorial Region. Int J Drug Deliv Technol. 2026; 16(8s): 360-372; DOI: 10.25258/ijddt.16.8s.48
Source of support: Nil.
Conflict of interest: None