AUTHOR=Marrah Saio Alusine , Wang Jiahao , Koroma Abu Bakarr , Kamara Gibrilla Deen , Babatunde Ologun Sodiq , Marrah Mabinty , Stamber Ryvel Timothy , Saidu Sylvester Edmond 
  
TITLE=Federated deep learning for privacy-preserving disease detection in IoT-enabled healthcare systems
  
JOURNAL=Frontiers in Computer Science
  
VOLUME=Volume 8 - 2026
  
YEAR=2026
  
URL=https://www.frontiersin.org/journals/computer-science/articles/10.3389/fcomp.2026.1725597
  
DOI=10.3389/fcomp.2026.1725597
  
ISSN=2624-9898
  
ABSTRACT=AbstractThis research addresses the burgeoning tension between the predictive power of artificial intelligence and the imperative of data sovereignty in healthcare. By proposing a privacy-preserving Federated Deep Learning (FDL) framework, this study develops a decentralized paradigm for early disease detection tailored for IoT-enabled clinical environments. The overarching objective is to architect and validate a scalable system that facilitates the training of sophisticated deep learning models across distributed nodes, thereby obviating the security vulnerabilities associated with the centralized aggregation of sensitive patient information.MethodsThe methodological rigor of this study rests on a multi-dimensional data approach, utilizing established benchmark datasets—specifically the UCI Heart Disease and Pima Indians Diabetes repositories—complemented by a synthetic temporal dataset designed to emulate high-frequency IoT sensor streams. We implemented a hybrid Convolutional Neural Network–Long Short-Term Memory (CNN-LSTM) architecture within the federated ecosystem, fortified with differential privacy (DP) protocols to neutralize the risk of model inversion attacks. Although the benchmark datasets are primarily static, the selection of the CNN-LSTM architecture was a strategic design choice to ensure generalizability to real-world medical IoT deployments where data are inherently sequential. A rigorous ablation study was conducted to isolate the impact of the LSTM component, which revealed a substantial performance uplift of 6.5 percentage points on temporal sequences without introducing significant computational overhead on static data.ResultsEmpirical evaluation confirms that the proposed framework maintains high predictive fidelity while upholding rigorous data confidentiality standards. The training and validation trajectories demonstrated stable convergence across successive federated communication rounds. Robustness was further evidenced by consistent performance across a suite of metrics—including Accuracy, Precision, Recall, F1-score, and AUC—for diverse disease profiles. Notably, communication efficiency was optimized through refined aggregation cycles. The integration of privacy-preserving mechanisms resulted in a marginal accuracy degradation of only 1–2%, representing a sustainable trade-off between cryptographic security and diagnostic utility.DiscussionThis study provides compelling empirical evidence that federated deep learning can successfully reconcile the competing demands of diagnostic accuracy, systemic efficiency, and patient privacy. By demonstrating the efficacy of the framework across both temporal and static data features, the findings contribute a novel methodological benchmark for the deployment of secure AI in preventive medicine. Ultimately, this work offers a scalable blueprint for the next generation of decentralized, privacy-first healthcare informatics.