Multistream LSTM for Artifact Detection in Impedance Cardiography

Monitoring cardiac hemodynamic parameters, such as cardiac output and pre-ejection period, is critical for assessing cardiovascular function, particularly in critically ill patients. Impedance cardiography (ICG) offers a noninvasive approach to measuring these parameters; however, its utility is often compromised by motion artifacts and electrode displacement. Many traditional artifact detection methods rely on rigid waveform templates, which may struggle to adapt to individual variations in ICG morphology, potentially resulting in limited generalization and higher misclassification rates in certain scenarios. In this study, we propose a deep learning-based framework that combines a Multi-Stream Long Short-Term Memory (LSTM) network, attention mechanisms, and ensemble learning to automatically detect corrupted ICG cycles. The model concurrently processes raw ICG signals and their derivatives to capture both temporal dynamics and morphological transitions. Attention layers highlight diagnostically relevant regions, while data augmentation and ensemble post-processing improve generalization and robustness. The proposed method was validated on a dataset of 2000 ICG cycles from 20 individuals, achieving an accuracy of 96.42% against human expert visual detection, significantly outperforming traditional methods and single-stream LSTM models. This method enhances artifact detection and supports more reliable noninvasive cardiac monitoring.