Time\u2013frequency time\u2013space LSTM for robust classification of physiological signals

Tuan D Pham

doi:10.1038/s41598-021-86432-7

Abstract

Automated analysis of physiological time series is utilized for many clinical applications in medicine and life sciences. Long short-term memory (LSTM) is a deep recurrent neural network architecture used for classification of time-series data. Here time–frequency and time–space properties of time series are introduced as a robust tool for LSTM processing of long sequential data in physiology. Based on classification results obtained from two databases of sensor-induced physiological signals, the proposed approach has the potential for (1) achieving very high classification accuracy, (2) saving tremendous time for data learning, and (3) being cost-effective and user-comfortable for clinical trials by reducing multiple wearable sensors for data recording.

Highlights

Analysis and classification of clinical time-series data in physiology and disease processes are considered as a catalyst for biomedical research and education
The Long short-term memory (LSTM) used in this study was the bidirectional LSTM (bi-LSTM) (LSTM will be used as bi-LSTM subsequently)
The TF–TS LSTM outperformed conventional LSTM, classification results of gait in Parkinson’s disease in terms of accuracy, sensitivity, specificity, precision, and F1 score obtained from the TF–TS LSTM are higher than those previously reported in literature

Summary

Introduction

Analysis and classification of clinical time-series data in physiology and disease processes are considered as a catalyst for biomedical research and education. Deep-learning methods or deep neural networks have been reported to outperform many baseline time-series classification approaches and appear to be the most promising techniques for classifying temporal d ata[3]. As a state-of-the-art method for learning physiological models for disease prediction, many applications of LSTM and other deep-learning networks have recently been reported in literature, such as classifying electroencephalogram (EEG) signals in emotion, motor imagery, mental workload, seizure, sleep stage, and event related p otentials[5], non-EEG signals in Parkinson’s disease (PD)[6], learning and synthesis of respiration, electromyograms, and electrocardiograms (ECG) signals[7], decoding of gait phases using EEG8, and early prediction of stress, health, and mood using wearable sensor data[9]. Conditions such as PD and post stroke, long walk trials are recommended to obtain at least 370 s trides[10] Such long-distance walks result in long records of physiological measurements, cause discomfort to the patients, and may be impractical to perform in many clinical s ettings[11]

Results

Discussion

Conclusion