Transformer + Ensemble fusion hits 98.9% accuracy in emotion recognition
New study from ACM BCB 2026 combines LSTM, TCN, and Transformer models...
A new paper accepted at ACM BCB 2026 presents a comprehensive evaluation of deep learning architectures for physiological emotion recognition. The team—Hagos, Aryal, Ymele-Leki, Andy, and Burge—tested Long Short-Term Memory (LSTM), Temporal Convolutional Networks (TCN), and Transformer models on the WESAD dataset, which contains wrist and chest sensor signals for stress and emotion detection. They conducted ablation studies to isolate the contribution of each sensor modality (wrist-only vs. chest-only) and implemented both early fusion (concatenating sensor signals at input) and late-fusion ensemble strategies (combining predictions from all three models).
Results show that Transformer models consistently outperform others in multimodal settings, while TCN models lead in wrist-only configurations. The late-fusion ensemble method achieved the highest overall accuracy of 98.91% (±0.13%) and macro-F1 score of 98.56% (±0.17%), demonstrating that combining diverse temporal architectures with sensor fusion yields a robust and generalizable emotion recognition system. The findings have direct implications for wearable health monitoring and affective computing applications.
- Transformer models achieve highest accuracy in multimodal settings; TCN best on wrist-only data
- Late-fusion ensemble of LSTM, TCN, and Transformer reaches 98.91% accuracy on WESAD dataset
- Early fusion at sensor level by concatenating wrist and chest signals improves overall performance
Why It Matters
Enables more reliable wearable health monitors that detect stress and emotions from physiological signals.