Wearable Bioimpedance-Based Deep Learning Techniques for Live Fish Health Assessment under Waterless and Low-Temperature Conditions

• Published in: Sensors
• Main Authors: Yongjun Zhang, Longxi Chen, Huanhuan Feng, Xinqing Xiao, Marina A. Nikitina, Xiaoshuan Zhang
• Format: Article in Journal/Newspaper
• Language: English
• Published: MDPI AG 2023
• Subjects: live fish health monitoring; waterless and low-temperature conditions; deep learning; wearable bioimpedance monitoring; stress evaluation; Chemical technology; TP1-1185; Turbot
• Online Access: https://doi.org/10.3390/s23198210; https://doaj.org/article/ef87a0271db74f4394f71ae5e781c8aa

(1) Background: At present, physiological stress detection technology is a critical means for precisely evaluating the comprehensive health status of live fish. However, the commonly used biochemical tests are invasive and time-consuming and cannot simultaneously monitor and dynamically evaluate multiple stress levels in fish and accurately classify their health levels. The purpose of this study is to deploy wearable bioelectrical impedance analysis (WBIA) sensors on fish skin to construct a deep learning-based stress dynamic evaluation model for precisely estimating their accurate health status. (2) Methods: The correlation of fish (turbot) muscle nutrients and their stress indicators are calculated using grey relation analysis (GRA) for allocating the weight of the stress factors. Next, WBIA features are sieved using the maximum information coefficient (MIC) in stress trend evaluation modeling, which is closely related to the key stress factors. Afterward, a convolutional neural network (CNN) is utilized to obtain the features of the WBIA signals. Then, the long short-term memory (LSTM) method learns the stress trends with residual rectification using bidirectional gated recurrent units (BiGRUs). Furthermore, the Z-shaped fuzzy function can accurately classify the fish health status by the total evaluated stress values. (3) Results: The proposed CNN-LSTM-BiGRU-based stress evaluation model shows superior accuracy compared to the other machine learning models (CNN-LSTM, CNN-GRU, LSTM, GRU, SVR, and BP) based on the MAPE, MAE, and RMSE. Moreover, the fish health classification under waterless and low-temperature conditions is thoroughly verified. High accuracy is proven by the classification validation criterion (accuracy, F1 score, precision, and recall). (4) Conclusions: the proposed health evaluation technology can precisely monitor and track the health status of live fish and provides an effective technical reference for the field of live fish vital sign detection.