Anomaly Detection in Health Data Streams Using Autoencoders and Recurrent Neural Networks

Abstract

Anomaly detection in health data streams is critical for early diagnosis, timely intervention, and effective healthcare management. This study presents a hybrid deep learning framework that combines Autoencoders (AEs) and Recurrent Neural Networks (RNNs) for robust and real-time anomaly detection in health-related time-series data. Autoencoders are employed to learn compact representations and reconstruct the input data, effectively capturing normal patterns while minimizing noise. RNNs, particularly Long Short-Term Memory (LSTM) networks, are integrated to model temporal dependencies and enhance sensitivity to irregularities across time. The reconstruction error from the autoencoder and prediction deviations from the RNN are jointly analyzed to detect anomalies. The proposed model is validated on benchmark physiological datasets, including heart rate, blood pressure, and ECG signals, demonstrating superior performance in identifying rare and subtle health events compared to traditional methods. This approach not only improves detection accuracy but also adapts dynamically to evolving patient conditions, making it suitable for continuous health monitoring in real-world settings. The model’s scalability and low false alarm rate highlight its potential for deployment in wearable devices and telemedicine applications.