E-Taste Signal Encoding and Classification via Multi- Attention Residual Networks Using sEMG Streams in Virtual Sensory Systems

Technical Summary: E-Taste Signal Encoding and Classification via Multi-Attention Residual Networks Using sEMG Streams

Problem Statement
The study addresses the challenge of accurately categorizing electronic taste (E-Taste) sensations using surface electromyography (sEMG) data. While taste perception is fundamental to human physiology, its loss due to medical conditions (e.g., anorexia, diabetes, COVID-19) and the growing interest in virtual sensory systems necessitate reliable, non-invasive detection methods. Previous approaches to sEMG classification for taste have relied on manually designed statistical features or standard deep learning architectures (such as CNNs and LSTMs) that often struggle with the subtle, highly overlapping muscle activation patterns associated with different taste categories. The specific difficulty lies in distinguishing between similar taste-induced sEMG signals (e.g., bitter vs. sour) without invasive procedures, requiring a model capable of extracting discriminative features from complex, multi-channel biosignals.

Methodology
The authors propose MACB-ResNet, a deep learning architecture designed to recognize six classes of E-Taste sensations: no taste, salty, sprite, bitter, mint, and sour.

• Data Acquisition: sEMG signals were recorded from the facial muscles (masseter, depressor anguli oris, and depressor labii inferioris) of eight healthy participants using a 10-electrode setup (six recording channels). Taste stimuli were delivered via a custom E-Taste device using silver electrodes and thermal/electrical stimulation at the tongue tip.
• Preprocessing: Raw sEMG signals were segmented into 7-second intervals, detrended, high-pass filtered (10 Hz cutoff), and subjected to adaptive notch filtering to remove power-line interference.
• Feature Representation: The preprocessed signals were converted into spectrograms using the Short-Time Fourier Transform (STFT). The resulting 6-channel spectrograms (size 6 × 129 × 63) served as the input tensor.
• Model Architecture: The MACB-ResNet model integrates a ResNet50 encoder with a Multi-Attention Channel Block (MACB).
  • Residual Learning: Utilizes residual connections to facilitate stable training and efficient gradient propagation in deep layers.
  • Multi-Attention Mechanism: The MACB module employs both channel-wise and spatial attention mechanisms. It calculates attention scores using average and max pooling followed by fully connected layers and sigmoid activation. This allows the network to weigh informative channels and spatial regions while suppressing irrelevant features, specifically targeting the weak and overlapping sEMG patterns of different tastes.
• Training: The model was trained for 30 epochs with a 90/10 train-test split, using a learning rate of 0.0001 and a batch size of 8.

Key Contributions

1. Novel Architecture: This is the first study to apply a residual dual-attention CNN framework (MACB-ResNet) specifically for six-class E-Taste recognition using sEMG data.
2. Feature Enhancement: The introduction of the MACB module, which jointly models channel and spatial dependencies, significantly improves the representation of discriminative sEMG features that are often lost in standard classification tasks.
3. Comprehensive Evaluation: The study provides a rigorous evaluation including class-level metrics, confusion matrix analysis, dimensionality reduction visualizations (PCA and t-SNE), and ablation studies to quantify the contribution of each architectural component.

Results

• Classification Performance: The MACB-ResNet model achieved a validation accuracy of 86.32%, outperforming KNN (65.11%), Random Forest (78.35%), and DenseNet-201 (81.93%).
• Class-Specific Analysis: The model demonstrated robust performance across most classes, with the highest F1 score (0.92) for the "no taste" class. However, confusion was observed between similar categories, particularly "bitter" and "sour," as well as "sprite," "mint," and "bitter," indicating that feature distributions for these classes remain partially overlapping.
• Ablation Study: Quantitative analysis confirmed the necessity of the proposed components. Removing residual connections reduced accuracy by 7.18%, removing channel attention reduced it by 5.09%, and removing the full MACB module resulted in a 9.74% drop in accuracy.
• Visualization: PCA and t-SNE visualizations supported the model's ability to separate classes, showing compact clustering for distinct tastes, though some overlap persisted for the more similar categories.
• Computational Cost: The model contains approximately 27.91 million parameters. Inference latency was estimated at approximately 5.4 ms per spectrogram window on a standard CPU, suggesting feasibility for real-time applications.

Significance and Claims
The paper claims that the MACB-ResNet framework offers a significant improvement in the precision of sEMG-based E-Taste classification compared to existing machine learning and deep learning baselines. By effectively handling the subtle and overlapping muscle activation patterns inherent in taste perception, the proposed model enhances the reliability of non-invasive taste assessment.

The authors position this work as a foundational step for future applications in:

• Personal Nutrition and Digital Health: Monitoring dietary intake and gustatory health.
• Food Technology and Neuromarketing: Analyzing consumer taste responses.
• Brain-Computer Interface (BCI) Systems: Enabling virtual sensory experiences and assistive technologies for individuals with impaired taste.

The study modestly acknowledges limitations, including the moderate dataset size, the specific demographic of participants (young adults), and the use of fixed time-frequency representations (STFT) which may not capture all dynamic temporal variations. The authors suggest that future work should involve larger, more diverse datasets and explore temporal models (e.g., Transformers, BiLSTMs) to further refine classification capabilities.