Federated Anomaly Detection on Wearable Sensor Data Streams for Real-Time Alarm Fatigue Mitigation in Intensive Care Units

Abstract

Background: Alarm fatigue in Intensive Care Units (ICUs) is a critical patient safety issue, driven by a high frequency of non-actionable alarms from monitoring devices. This phenomenon leads to desensitization among clinical staff, potentially causing delayed responses to true critical events. While Edge AI and wearable sensors offer a promising avenue for real-time, continuous patient monitoring, concerns over data privacy and model personalization remain significant barriers.

Objective: This study aims to design, develop, and evaluate a novel framework that leverages Edge AI and Federated Learning (FL) to mitigate ICU alarm fatigue. The proposed system performs real-time anomaly detection on physiological data streams from wearable sensors, training a robust, personalized model without centralizing sensitive patient data.

Methods: We architected a three-tiered system comprising wearable sensor nodes, edge computing gateways, and a central FL server. A lightweight autoencoder model was deployed on edge devices for real-time anomaly detection of a patient's physiological data (e.g., heart rate, SpO2). We implemented a Federated Averaging protocol to collaboratively train a global anomaly detection model. The edge models learn from local data streams to capture individual patient baselines, while the global model benefits from the diverse patterns across the entire patient cohort. The system's performance was evaluated against traditional centralized and non-federated models using metrics of alarm reduction rate, detection accuracy, and model convergence speed.

Results: The proposed federated edge framework demonstrated a significant reduction in false alarms by over 40% compared to baseline thresholding systems, while maintaining a high sensitivity to true adverse events (F1-Score > 0.95). The FL model converged within 50 communication rounds, showing efficient learning. The approach outperformed both centralized models (which lack personalization) and isolated local models (which lack diverse training data) in overall accuracy and adaptability.

Conclusions: Our findings indicate that a federated, edge-based approach can effectively and safely reduce alarm fatigue in ICUs. By preserving data privacy and enabling model personalization, this framework presents a viable solution for developing intelligent, real-time clinical decision support systems.