Wearable Bioimpedance-Based Deep Learning Techniques for Live Fish Health Assessment under Waterless and...

Area of Science:

• Aquaculture
• Biomedical Engineering
• Machine Learning

Background:

• Current physiological stress detection in fish is invasive and time-consuming.
• Existing methods struggle with simultaneous, dynamic multi-stress monitoring and accurate health classification.
• There is a need for non-invasive, real-time fish health assessment tools.

Purpose of the Study:

• To develop a deep learning-based stress dynamic evaluation model using wearable bioelectrical impedance analysis (WBIA) sensors.
• To precisely estimate the accurate health status of live fish.
• To overcome limitations of traditional biochemical stress tests.

Main Methods:

• Grey relation analysis (GRA) weighted stress factors based on fish nutrient correlations.
• Maximum information coefficient (MIC) selected WBIA features relevant to stress.
• A deep learning model combining Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM), and Bidirectional Gated Recurrent Units (BiGRUs) was employed.
• Z-shaped fuzzy function for health status classification.

Main Results:

• The CNN-LSTM-BiGRU model demonstrated superior accuracy over other machine learning models (CNN-LSTM, CNN-GRU, LSTM, GRU, SVR, BP) using MAPE, MAE, and RMSE metrics.
• Accurate fish health classification was validated under challenging conditions (waterless, low-temperature).
• High accuracy was confirmed through classification metrics including accuracy, F1 score, precision, and recall.

Conclusions:

• The developed WBIA sensor and deep learning model offer precise, non-invasive monitoring of live fish health.
• This technology provides a valuable reference for vital sign detection in live fish.
• The system enables dynamic tracking and evaluation of fish health status.