From:  Artificial intelligence strategies for emotion recognition in cancer pain research

 Physiological and wearable signal modeling for emotion recognition.

Approach/ModelInput signalsMain functionTypical applicationsKey challengesRef.
Recurrent and transfer learning models (LSTM, CNN-BiLSTM, transfer learning)ECG, PPG, GSR, temperature, respiration, accelerometryTemporal modeling and cross-modal generalization of physiological time seriesEmotion recognition, stress detection, cognitive load monitoring, wearable health inferenceMotion artifacts, inter-individual variability, non-stationarity, domain shift[4247]
Signal processing and feature extraction approachesPrimarily PPG and related wearable biosignalsSignal denoising, compression, and improvement of signal qualityPreprocessing for wearable-based monitoring and downstream classification tasksSensitivity to real-world noise, reduced robustness under motion, and low signal-to-noise conditions[48]
Multimodal foundation models (e.g., NormWear)Multivariate physiological signals (e.g., ECG, PPG, GSR, EEG)Learning transferable representations across tasks and populationsEmotion recognition, stress detection, sleep analysis, zero-shot/few-shot wearable sensingHigh training cost, data heterogeneity, potential bias, and limited clinical validation[49]
Self-supervised learning (SSL)Unlabeled physiological and wearable biosignals, especially PPGLearning latent representations through reconstruction-based or related pretext tasksScalable wearable biosignal modeling with limited annotation requirementsDefining suitable pretext objectives, interpretability, and downstream transferability[50]

LSTM: long short-term memory; CNN: convolutional neural network; BiLSTM: bidirectional LSTM; ECG: electrocardiography; PPG: photoplethysmography; GSR: galvanic skin response; EEG: electroencephalography.