Presently, the primary focus in emotion recognition through ECG revolves around the utilization of classification algorithms alongside the selection of pertinent features and the scope of identifiable emotions. Machine learning algorithms have proven to be effective for emotion recognition based on ECG data. Axel and his team employed the wavelet scattering method for the extraction of unique characteristics from ECG data. This method facilitated the acquisition of signal features spanning varying time scales to assess their performance. The outcomes of this study demonstrated that the proposed feature-extraction and signal classification algorithm, within the realm of emotional dimensions, achieved an accuracy rate of 88.8% for arousal, 90.2% for valence, and an impressive 95.3% for two-dimensional classification [16]. Theekshana et al. presented a machine learning approach utilizing an ensemble learning method to identify core emotions, including anger, sadness, happiness, and the combination of happiness with electrocardiogram (ECG) data. They utilized spectral analysis as a novel approach to extract features and assessed the efficacy of a widely recognized collection of ensemble learners for emotion classification through a machine learning process. The ensemble learners exhibited a 10.77% enhancement in accuracy when compared to the top-performing individual biosensor-based model [17]. Yan et al. introduced the X-GWO-SVM approach to analyze sentiments from ECG data. They conducted single-subject cross-validation and achieved an impressive average accuracy rate of 95.93% when utilizing the WESAD dataset. This result demonstrates its superior reliability compared to previous implementations of supervised machine learning techniques [18].

Deep learning algorithms are commonly employed in the realm of emotion recognition as well. Pritam et al. devised a novel approach using a self-supervised deep multitask learning framework to detect emotions based on electrocardiogram (ECG) data. In this process, the convolutional layer remained fixed, while the dense layer was fine-tuned using labeled ECG data. Notably, this innovative solution yielded state-of-the-art results in classifying arousal, mood, emotional state, and stress across four distinct datasets [19]. Xu et al. introduced an emotion detection technique rooted in deep learning tailored for healthcare data analysis. Their approach incorporated a multichannel convolutional neural network to extract distinctive features from ECG data and textual content related to emotions, particularly for detecting emotional fatigue. Ultimately, this method amalgamated the features derived from multiple data sources to ascertain emotional states. The experimental findings demonstrated that the proposed model consistently achieved an accuracy rate exceeding 85% in the prediction of emotional fatigue [20]. Chen et al. introduced a novel approach to emotion recognition, which involved combining multiple sensory modalities using the Dempster–Shafer evidence theory. They utilized an SVM classifier to categorize EEG signal features. The results of the experiment demonstrated that the multimodal fusion model outperformed the unimodal emotion detection method, resulting in a significant increase in accuracy by 7.37% and 8.73% as compared to the emotion recognition model based on ECG signals [21]. Hammad et al. applied the PETSFCNN model in their research on emotion recognition, achieving an impressive maximum classification accuracy of 97.56%. They employed a deep neural network in combination with grid search optimization to enhance accuracy in classification tasks. The results of their study indicated that the suggested method surpasses existing techniques in precisely identifying emotions from ECG signals. This implies the possibility of utilizing it as an intelligent system for emotion recognition [22].

When it comes to selecting features for emotion recognition, HRV-related characteristics have gained widespread usage. Guo et al. conducted an analysis of ECG signals to extract heart rate variability (HRV) features, employing techniques encompassing frequency domain, time domain, statistical methods, and Poincaré. The HRV characteristics were later utilized for categorizing different emotional states, utilizing principal component analysis, and then employing a support vector machine to reduce the feature set. Notably, they achieved classification accuracies of 71.4% for distinguishing two emotional states (positive/negative) and 56.9% for classifying five emotional states [23]. Ferdinando et al. employed K-nearest neighbors (KNN) as a classifier in their study. They combined supervised dimensionality reduction using neighborhood component analysis (NCA) with feature-extraction techniques, which included capturing standard HRV features and statistical distributions of instantaneous frequencies. These methods were applied to address the classification of three distinct emotion and arousal categories. The findings suggested that, in most instances, integrating NCA resulted in a significant performance boost of 74% when compared to the absence of NCA in the implementation [24]. Singson et al. harnessed a ResNet-based CNN to analyze both facial expressions and physiological data, with a particular focus on heart rate variability (HRV) features. Their aim was to discern and validate emotions, achieving an accuracy rate of 68.42% through the analysis of ECG signals [25].