Empirical classification of fatigue-induced physiological tremor in robot-assisted manipulation tasks using BiLSTM-GRU network

TL;DR

This study uses a new neural network to classify fatigue-induced tremors during robotic surgery tasks, achieving high accuracy to help surgeons manage tremors effectively.

Contribution

A novel BiLSTM-GRU network is introduced for classifying fatigue-induced tremor stages with high accuracy using multimodal sensor data.

Findings

The proposed BiLSTM-GRU model achieved 99% classification accuracy for tremor stages.

Cross-sectional area changes from vision sensors improved tremor classification performance.

Fatigue progression was effectively tracked through repeated pattern-tracing tasks.

Abstract

Physiological tremor arises due to stress, anxiety, fatigue, alcohol or caffeine. Under conventional circumstances, the physiological tremor would not be detrimental. Still, the mere presence of such a tremor during any microsurgical procedure can be catastrophic. In these instances, it is necessary to predict the progression of the tremor. This article proposes a novel sensing methodology and adds a distinctive feature to aid in classification. The classification of the progressive stages of fatigue-induced physiological tremor (FIPT) is based on the hybrid bidirectional long short-term memory neural network with a Gated Recurrent Unit (BiLSTM-GRU) presented in this work.

Twenty healthy participants volunteered in the study, where a teleoperation stage was set up using the Geomagic Haptic device—Touch. On the master end, the participants were seated comfortably and asked to trace the patterns embedded over an image of an organ that was displayed on the screen. The EMG and MMGACC signals from the Mindrove Armband and cross-sectional area changes, MMGCSAC, calculated from area measurement using the vision sensor, were recorded. The pattern-tracing task (PTT) was carried out over five repetitions, with fatigue-inducing exercise occurring between task epochs, thus accumulating fatigue throughout the data collection process. The extracted features from human movement aid the classification of the stages of tremor using BiLSTM-GRU, showing the significance of a cross-sectional area informed model.

The stages of progression of tremor are classified into five levels in this study, and classified using BiLSTM GRU with four different input feature sets. The performance evaluation metrics, such as the accuracy, precision, recall and F1 score, have been reported to ascertain the efficiency of the proposed feature group. The proposed feature set and classification strategy are capable of estimating stages of FIPT with 99% classification accuracy. This can be used to design state-of-the-art movement training platforms for both experienced and novice surgeons that allow informed decision making to attend to their tremor condition, either by taking a break or including a limb support to minimize its effects. At the same time, the identification methodology can be extended to pathological tremor rehabilitation and any other movement disorder diagnostics.

Figures