PPGE: PPG Dataset for Emotion Recognition

• PPGE is a public PPG emotion dataset that provides raw signals, binary valence-arousal labels, and 5-minute video stimuli for robust cross-subject evaluation.
• It employs a controlled acquisition protocol with 100 Hz sampling, a 3rd-order Butterworth bandpass filter, and 60-second window segmentation to preserve waveform integrity.
• Advanced benchmarks using CNN-TCN-LSTM architectures demonstrate improved generalization and uncertainty-aware emotion prediction compared to consumer-wearable IBI data.

Searching arXiv for the specified papers and related PPGE work. The Photoplethysmogram Dataset for Emotional Analysis (PPGE) is a benchmark line of research in affective computing centered on emotion recognition from photoplethysmography-derived cardiovascular signals. In later work, PPGE is described as a publicly available PPG emotion dataset originally introduced in Jin et al. [22], providing raw PPG, binary emotion labels, and longer video stimuli of about 5 minutes for subject-independent evaluation (Alghoul et al., 10 Jul 2025). Closely related work in the same research area established a consumer-wearable PPG/IBI collection acquired from a Garmin vivosmart 3 and used it to study end-to-end Bayesian prediction of emotional valence from heartbeat time series, with explicit uncertainty quantification and abstention (Harper et al., 2019).

1. Dataset identity and scope

PPGE is described as a publicly available PPG emotion dataset used to study emotion recognition from raw photoplethysmogram signals. In the 2025 generalization study, the dataset is reported to contain 72 PPG signals from 18 participants, comprising 13 men and 5 women aged 23–31, sampled at 100 Hz. Each subject completed 4 trials, and each trial corresponded to a distinct emotional stimulus (Alghoul et al., 10 Jul 2025).

The same source emphasizes several properties that make PPGE methodologically distinctive within PPG-based affective computing. It provides binary emotion labels, uses longer video stimuli of about 5 minutes, and is explicitly treated as suitable for Leave-One-Subject-Out Cross-Validation (LOSOCV), because the central problem is not only emotion recognition from PPG but also generalization to unseen individuals (Alghoul et al., 10 Jul 2025). This suggests that PPGE occupies a specific niche: it is not merely a signal repository, but a benchmark for subject-independent affect modeling under small-sample conditions.

2. Acquisition protocol and signal preparation

In the PPGE benchmark as described in the 2025 study, raw photoplethysmogram is recorded from a fingertip pulse sensor while participants watch video stimuli. After each video, participants rate their emotional response, producing the valence-arousal labels used in the experiments. The study states that trial information is preserved so that data from different trials and subjects are not mixed in a way that could blur subject-specific physiological responses (Alghoul et al., 10 Jul 2025).

The preprocessing pipeline is deliberately constrained. The raw PPG is filtered with a 3rd-order Butterworth bandpass filter with cutoff frequencies 0.7–3.7 Hz, corresponding roughly to heart rates from about 40 bpm to 220 bpm. The filtered signal is segmented into 60-second windows with 5-second overlap, and each segment is normalized using Z-score normalization. The authors also note that class imbalance is important enough that training uses dynamic class weights and weighted categorical crossentropy (Alghoul et al., 10 Jul 2025).

These choices matter because they preserve a direct relationship between waveform-level PPG morphology and downstream learning. The use of raw PPG, rather than exclusively handcrafted HRV features, permits architectures that jointly model local morphology and temporal dependence. At the same time, the LOSOCV protocol and class-weighted training make the dataset a testbed for cross-subject robustness rather than within-subject memorization.

3. Consumer-wearable IBI acquisition lineage

A related dataset contribution in this literature was collected using a Garmin vivosmart 3 worn on the left wrist, with the device’s embedded PPG sensor used to extract heartbeat intervals locally on a mobile device via a custom Android Wear app. That app synchronized the extracted IBIs with the timing of the video stimuli and uploaded the resulting data to the cloud after the session. The experiment involved 17 participants in total, consisting of 5 female and 12 male participants (Harper et al., 2019).

The recording protocol was designed for emotion elicitation under relatively practical conditions. Participants were seated about 60 cm from a computer monitor, wore headphones to reduce distraction, and were left alone during recording after being briefed. They first received instruction on the Self-Assessment Manikin (SAM) emotion-reporting framework. The experimental design used 24 short emotion-inducing video stimuli shown in random order. Those videos were curated from 96 candidate videos independently annotated by 30 volunteers using SAM; the authors computed annotation variance and selected the top 25% lowest-variance clips. Of the final 24 stimuli, 8 were intended to induce pleasure and 16 displeasure, with an average duration of about 2:29. Between clips, participants viewed one minute of neutral content to reduce emotional carryover. Frontal face video was also recorded with a webcam, although it was not used in the reported emotion model (Harper et al., 2019).

This consumer-wearable line differs from the raw-PPG PPGE benchmark in signal representation and sensing hardware. PPGE, as described in later work, uses raw fingertip PPG. The Garmin-based study uses PPG-derived inter-beat interval sequences from a low-cost wrist tracker. A plausible implication is that the PPGE literature spans both waveform-centric and heartbeat-interval-centric formulations of PPG emotion analysis, with different trade-offs in deployment realism, temporal granularity, and susceptibility to acquisition artifacts.

4. Emotion labels and task formulations

The public PPGE benchmark frames emotion using the standard valence-arousal model. Valence ranges from unpleasant to pleasant, and arousal ranges from calm to excited. In the reported experiments, binary classification is performed for valence and arousal separately, using the labels provided by the dataset after participant self-reporting (Alghoul et al., 10 Jul 2025).

In the consumer-wearable IBI study, the target variable is emotional valence represented by SAM self-reports. The prediction task is framed primarily as binary high/low valence classification, built from an underlying valence regression output. The valence axis is treated as continuous, and a decision boundary is placed at its midpoint: values above the midpoint correspond to high valence, and values below correspond to low valence. To convert probabilistic regression outputs into class labels, the model uses a confidence threshold parameter $\alpha$. If enough posterior mass lies on one side of the midpoint, the model predicts the corresponding class; otherwise it abstains. With $N$ stochastic forward passes, the posterior predictive over valence is approximated by samples

$$\hat{y}^{(1)}, \hat{y}^{(2)}, \ldots, \hat{y}^{(N)},$$

and classification is made only when a sufficient fraction $\alpha$ of these samples lies on one side of the midpoint (Harper et al., 2019).

The two task definitions are compatible but not identical. One line uses binary valence and arousal classification from raw PPG segments; the other uses valence regression with thresholded Bayesian decision-making from IBI sequences. This suggests that PPGE-related research has treated emotional annotation both as a discrete classification problem and as a continuous latent variable with uncertainty-aware discretization.

5. Benchmark architectures and reported results

For raw-PPG modeling on PPGE, the 2025 study proposes a hybrid CNN-TCN-LSTM architecture with the pipeline raw PPG $\rightarrow$ CNN feature extractor $\rightarrow$ parallel LSTM and TCN branches $\rightarrow$ concatenation $\rightarrow$ softmax output. The CNN feature extractor has two convolutional layers: the first uses 8 filters, kernel size 64, stride 4, ReLU activation, same padding, max pooling, batch normalization, and dropout = 30%; the second uses 16 filters, kernel size 32, stride 2, followed by max pooling, batch normalization, and dropout = 30%. The temporal branches consist of one LSTM layer with 12 units and a Temporal Convolutional Network with 8 filters, kernel size 32, dilation rates $[1,2,4,8]$, causal padding, skip connections, dropout = 30%, implemented using KerasTCN. The fused representation is mapped to a fully connected layer with 2 neurons and Softmax activation. Training uses Adam optimizer, learning rate = 0.01, and weighted categorical crossentropy with dynamic class weights (Alghoul et al., 10 Jul 2025).

Under LOSOCV, that model is evaluated with AUC, Accuracy, and F1-score, with AUC emphasized as the primary metric. For valence classification, the reported AUC values are 0.64 for CNN, 0.60 for CNN-LSTM, and 0.66 for the proposed model; the corresponding best average F1-score for valence is 0.65 for the proposed model, compared with 0.62 for CNN and 0.60 for CNN-LSTM. For arousal classification, the AUC values are 0.61 for CNN, 0.68 for CNN-LSTM, and 0.69 for the proposed model. Averaging valence and arousal, the AUC values are 0.63 for CNN, 0.64 for CNN-LSTM, and 0.68 for the proposed model (Alghoul et al., 10 Jul 2025).

The consumer-wearable IBI study uses a different end-to-end Bayesian deep neural network adapted from prior ECG work. Its input is the wearable IBI time series. The architecture contains two parallel temporal streams: a convolutional stream with four stacked 1D convolutional layers, each using 128 filters, with receptive-field sizes decreasing from 8 to 2 time steps as depth increases, with ReLU and Monte Carlo dropout after each convolutional block; and a recurrent stream consisting of a bidirectional LSTM with 32 hidden units, also followed by Monte Carlo dropout. The two streams produce a 192-dimensional vector, combining 128 features from the convolutional branch and 64 from the bidirectional LSTM, which feeds a final dense layer that outputs a regression estimate of valence, $\hat{y}$. Training uses Adam for 1000 epochs, with the learning rate decreasing from $N$0 to $N$1, dropout rates of 50% after convolutional blocks and 80% after the BiLSTM, and $N$2 stochastic forward passes at test time. When the confidence threshold is set to $N$3, the model achieves a peak F1 score of 0.70 on the consumer-wearable PPG dataset; as $N$4 increases, F1 improves while coverage falls (Harper et al., 2019).

Taken together, these results establish two complementary benchmark narratives. On raw PPGE, the principal issue is cross-subject generalization, assessed primarily with AUC. On consumer-wearable IBI data, the principal issue is uncertainty-aware deployment, assessed through the trade-off between confidence and coverage. Both are technically central to practical PPG-based emotion recognition.

6. Comparative position, non-equivalence with ECG, and neighboring datasets

An important misconception addressed directly in this literature is that ECG-derived and PPG-derived heartbeat signals are interchangeable for emotion recognition. In the consumer-wearable study, training on $N$5 plus $N$6 from the AMIGOS dataset did not significantly improve performance over PPG alone, with $N$7 from a Mann–Whitney test on the 10 F1 scores at $N$8. Training on $N$9 alone and testing on the wearable PPG dataset performed no better than chance. Using frequency-domain and time-domain HRV features—HF, LF, VLF, LF/HF, mean, median, SDSD, NN20, pNN20, rMSSD, and multiscale entropy—the authors found statistically significant differences for VLF power, SDSD, NN20, pNN20, rMSSD, and multiscale entropy. A simple SVM trained to distinguish ECG-derived from PPG-derived IBI sequences achieved 70% accuracy under 10-fold cross-validation (Harper et al., 2019).

This non-equivalence positions PPGE and related PPG datasets as more than downstream substitutes for laboratory ECG corpora. The literature argues that consumer PPG signals have their own statistical properties and require their own datasets (Harper et al., 2019). That argument is consistent with the 2025 PPGE study, which treats LOSOCV and AUC-driven subject-independent analysis as primary design constraints rather than secondary evaluation details (Alghoul et al., 10 Jul 2025).

Within the broader dataset landscape, PPGE is complemented by multimodal and contactless resources rather than superseded by them. MDEAW is a multimodal database for emotion analysis through EDA and PPG collected during real classroom exams from 10 male students aged 23 to 57, using wearable, wireless, low-cost, off-the-shelf equipment and six basic emotion labels: sadness, happiness, fear, anger, surprise, and disgust (Nandi et al., 2022). CAST-Phys is a public multimodal dataset explicitly designed for remote physiological emotion recognition, comprising 61 participants, synchronized BVP/PPG, EDA, respiration, and uncompressed facial video, with analyses showing that physiological signals—especially PPG-derived HR/HRV—are highly informative for affect recognition and for remote signal recovery (Comas et al., 8 Jul 2025).

In that comparative setting, PPGE remains distinctive because it is public, uses longer emotion-inducing stimuli, provides valence-arousal labels, and is used explicitly for cross-subject evaluation (Alghoul et al., 10 Jul 2025). Its main limitation, also stated in the literature, is its relatively small size. A plausible implication is that PPGE functions most effectively as a high-value benchmark for methodological comparison—especially for generalization, uncertainty estimation, and signal-representation choices—rather than as a large-scale pretraining corpus.