A Two-Stage Federated Learning Framework for Human Activity and Anomaly Detection in IoMT

Abylay Salimzhanov, Ayaulym Tynyssova, Temirlan Tursynbekov, Saida Tulebayeva, Ayazhan Yelmagambetova, Enver Ever, Adnan Yazıcı

2025

Full text Open access: Hybrid Metadata (CSL-JSON)
DOI 10.1145/3770501.3770510
Data confidence 70 medium
How this score is calculated

It is built from checkable properties of the record itself: whether the article has a DOI, how many independent sources confirmed its data, whether the required fields are filled in, and how recent the latest confirmation is. Disagreement between sources and a missing identifier lower it. A value of 80 or above means the record was published without further checking; 60 to 79 means it is published at reduced confidence, with a manual check advisable. A record verified by an editor by hand is never pushed below 90 by a recalculation. This scores our data about the article, not the article itself.

Abstract

Recent advances in the Internet of Things (IoT) and communication technologies have enabled the Internet of Medical Things (IoMT), which integrates IoT into healthcare for real-time patient monitoring and smart interventions. IoMT applications are increasingly used in continuous health monitoring. However, conventional machine learning (ML) with centralized data aggregation raises concerns about privacy, data security, and compliance. To address these, federated learning (FL) has emerged as a promising solution. FL enables collaborative model training across multiple entities (hospitals, homes, etc.) without sharing raw data, preserving privacy and promoting decentralized intelligence in IoMT. This paper proposes a privacy-preserving health monitoring and emergency detection framework, integrating IoT sensing, FL, and expert feedback for model refinement. The two-stage system first captures physiological and contextual data (heart rate, SpO2, ECG, age, sex) using wearables, then uses multimedia devices for human activity recognition (HAR) to validate anomalies. Model training is distributed through FL, ensuring compliance. Experimental results show anomaly detection accuracy of 0.89 with SensorECGCNN-Light, and HAR accuracy of 0.85 with a Long-term Recurrent Convolutional Network (LRCN). Both model outputs are combined through federated decision-level fusion, where a meta-classifier makes the final prediction. User confirmation and caregiver feedback enables iterative model refinement through expert feedback, improving adaptability, performance, and resilience.