Emotion Recognition from PPG Signals: Methods, Accuracy, and Applications
Technical overview of PPG-based emotion recognition covering feature extraction, machine learning classifiers, accuracy benchmarks, and real-world applications in affective computing.
PPG-based emotion recognition uses cardiovascular features extracted from the optical pulse wave to classify emotional states, achieving 65 to 85 percent accuracy for arousal detection and 55 to 75 percent for valence classification across standard benchmarks like DEAP and WESAD. The approach leverages the tight coupling between the autonomic nervous system and emotional experience to infer affective states from a single wrist-worn or fingertip sensor.
Why PPG for Emotion Recognition
The autonomic nervous system responds to emotional stimuli through measurable cardiovascular changes. Fear increases heart rate and vasoconstriction. Sadness often slows heart rate and reduces peripheral blood flow. Joy activates a mixed autonomic pattern with moderate sympathetic arousal. These responses manifest in the PPG signal as changes in pulse rate, pulse wave shape, and beat-to-beat variability.
PPG offers several advantages over other physiological sensors for emotion recognition:
- Ubiquitous hardware: Every modern smartwatch and fitness band includes a PPG sensor
- Continuous monitoring: PPG can be sampled 24/7 without user effort
- Low obtrusiveness: No electrodes, gel, or special preparation required
- Involuntary response: Unlike facial expressions, cardiovascular responses are difficult to consciously suppress
The primary disadvantage is that PPG provides a single physiological channel. Compared to multimodal systems using EEG, EMG, EDA, and PPG together, single-channel PPG has lower discriminability for distinguishing emotions that share similar arousal levels.
The Arousal-Valence Framework
Most PPG emotion recognition research uses Russell's circumplex model, which maps emotions onto two dimensions:
- Arousal: Low (calm, relaxed) to high (excited, agitated)
- Valence: Negative (sad, angry, fearful) to positive (happy, amused, content)
| Quadrant | Arousal | Valence | Example Emotions |
|---|---|---|---|
| High arousal, positive valence | High | Positive | Excitement, joy, amusement |
| High arousal, negative valence | High | Negative | Anger, fear, anxiety |
| Low arousal, positive valence | Low | Positive | Calm, contentment, relaxation |
| Low arousal, negative valence | Low | Negative | Sadness, boredom, depression |
PPG features correlate more strongly with arousal than valence because sympathetic and parasympathetic activation directly modulate heart rate and vascular tone. Valence discrimination requires detecting subtler patterns in pulse wave morphology.
Feature Extraction Methods
Time-Domain HRV Features
The most commonly used features for PPG-based emotion recognition include:
- RMSSD (root mean square of successive differences): Reflects parasympathetic activity
- SDNN (standard deviation of NN intervals): Captures overall HRV
- pNN50 (percentage of successive intervals differing by more than 50 ms): Parasympathetic marker
- Mean heart rate: Basic arousal indicator
Frequency-Domain HRV Features
Spectral analysis of the inter-beat interval series yields:
- LF power (0.04-0.15 Hz): Mixed sympathetic and parasympathetic influence
- HF power (0.15-0.40 Hz): Primarily parasympathetic (respiratory sinus arrhythmia)
- LF/HF ratio: Sympathovagal balance estimate
Pulse Wave Morphology Features
Beyond HRV, the shape of individual PPG pulses carries emotional information:
- Pulse amplitude: Correlates with stroke volume and vascular tone
- Pulse width: Related to systolic ejection time
- Dicrotic notch position: Reflects arterial compliance and afterload
- Rise time: Time from foot to systolic peak
For more on PPG morphology analysis, see our guide to PPG morphology features.
Deep Learning on Raw PPG
Recent work bypasses manual feature engineering by feeding raw PPG segments directly into convolutional neural networks (CNNs) or recurrent networks (LSTMs). These models learn hierarchical representations that can capture features humans might not identify. Studies using the DEAP dataset report that 1D-CNN models on 10-second PPG windows achieve arousal classification accuracy of 78 to 85 percent (Shu et al., 2018).
Benchmark Results Across Datasets
| Dataset | Subjects | Stimuli | Best PPG Accuracy (Arousal) | Best PPG Accuracy (Valence) |
|---|---|---|---|---|
| DEAP | 32 | Music videos | 82.4% | 73.6% |
| WESAD | 15 | Stress tasks | 87.1% | N/A (stress/non-stress) |
| AMIGOS | 40 | Short videos | 78.5% | 68.2% |
| MAHNOB-HCI | 27 | Film clips | 76.3% | 65.8% |
These numbers represent best-reported results from various studies using subject-dependent models. Cross-subject generalization typically reduces accuracy by 8 to 15 percentage points because individual baseline HRV varies widely.
Challenges in Real-World Deployment
Individual Variability
Resting heart rate ranges from 40 to 100 bpm across healthy adults, and baseline HRV varies by a factor of 5 or more. A model trained on population averages performs poorly on individuals at the extremes. Personalization through transfer learning or calibration sessions improves cross-subject accuracy by 5 to 12 percent.
Motion Artifacts
During daily activities, wrist motion introduces signal corruption that can be misinterpreted as emotional arousal. Accelerometer-based artifact rejection is essential for ambulatory emotion monitoring. For related discussion on signal quality, see our article on PPG signal quality assessment.
Temporal Dynamics
Emotions unfold over seconds to minutes, while HRV features require 30 seconds to 5 minutes of data for stable computation. This temporal mismatch means PPG-based systems detect sustained emotional states more reliably than brief emotional reactions.
Ground Truth Labeling
Self-reported emotion labels are subjective, noisy, and delayed. Participants often struggle to rate their own emotional state accurately, which introduces label noise into training data. This is a fundamental limitation shared by all affective computing approaches.
Applications
Automotive Safety
Driver emotion and drowsiness monitoring systems use steering wheel or seatbelt-integrated PPG sensors to detect stress, fatigue, and emotional distraction. When combined with camera-based gaze tracking, PPG provides an additional physiological channel for driver state estimation.
Mental Health Monitoring
Longitudinal PPG-based affect tracking could enable early detection of mood disorder episodes. Research prototypes track daily emotional patterns and flag sustained negative affect that might indicate a depressive episode. Clinical validation is ongoing.
Adaptive Interfaces
Affective computing researchers are building user interfaces that adapt to emotional state. A stressed user might receive simplified layouts and calming color schemes, while an engaged user gets richer content. PPG from smartwatches provides the physiological input for these adaptive systems.
Gaming and Entertainment
Emotion-responsive games adjust difficulty, narrative, and atmosphere based on the player's physiological arousal. PPG sensors in game controllers or VR headsets enable this without interrupting gameplay. For more on PPG applications in computing, see our article on PPG affective computing applications.
Frequently Asked Questions
Can PPG signals detect emotions?
PPG signals can classify broad emotional states with 65 to 85 percent accuracy for arousal and 55 to 75 percent for valence. The approach works because emotional states modulate autonomic cardiovascular control in measurable ways.
What PPG features are used for emotion recognition?
Common features include HRV metrics (RMSSD, SDNN, LF/HF ratio), pulse wave morphology features (amplitude, width, dicrotic notch position), and statistical signal features. Deep learning models can learn directly from raw PPG waveforms.
How does PPG emotion recognition compare to facial expression analysis?
PPG captures internal physiological arousal while facial analysis captures external behavioral displays. PPG is harder to fake but has lower resolution for distinguishing specific emotions. Combining both modalities outperforms either alone.
Is PPG-based emotion recognition used in real products?
It is primarily in research and prototyping stages. Consumer wearables track stress through HRV but do not label specific emotions. Automotive driver monitoring is the closest to production deployment.
What is the DEAP dataset?
DEAP is a benchmark dataset with PPG and other physiological signals from 32 participants watching 40 music videos rated on arousal, valence, dominance, and liking. It is the most widely cited benchmark in PPG emotion recognition research.
Summary
PPG-based emotion recognition is a maturing field that leverages the autonomic nervous system's response to emotional stimuli. Current systems achieve useful accuracy for arousal classification but struggle with fine-grained emotion labeling and cross-subject generalization. As deep learning models improve and wearable PPG sensor quality increases, the gap between laboratory and real-world performance continues to narrow, opening practical applications in automotive safety, mental health, adaptive interfaces, and entertainment.
Frequently Asked Questions
- Can PPG signals detect emotions?
- PPG signals can classify broad emotional states, particularly arousal level (calm vs. excited), with 65 to 85 percent accuracy depending on the number of emotion categories and the machine learning model used. Valence classification (positive vs. negative emotion) is more challenging, typically achieving 55 to 75 percent accuracy because positive and negative emotions can produce similar cardiovascular responses at the same arousal level.
- What PPG features are used for emotion recognition?
- Common PPG features for emotion recognition include heart rate variability metrics (RMSSD, SDNN, LF/HF ratio), pulse wave morphology features (pulse transit time, systolic peak amplitude, dicrotic notch depth), and statistical features of the raw signal (mean, variance, skewness, kurtosis). Deep learning approaches skip manual feature extraction and learn directly from the raw PPG waveform.
- How does PPG emotion recognition compare to facial expression analysis?
- PPG emotion recognition captures internal physiological arousal while facial expression analysis captures external behavioral displays. PPG is harder to fake because autonomic responses are largely involuntary, but it has lower spatial resolution for distinguishing specific emotions. The two modalities are complementary, and multimodal systems combining both typically outperform either alone.
- Is PPG-based emotion recognition used in real products?
- PPG emotion recognition is primarily used in research settings, automotive driver monitoring prototypes, and specialized wellness applications. Consumer wearables like Garmin and Fitbit track stress (a component of emotional state) through HRV but do not label specific emotions. The technology is maturing but has not yet reached mainstream consumer deployment for full emotion classification.
- What is the DEAP dataset used for in PPG emotion research?
- The DEAP dataset (Database for Emotion Analysis using Physiological Signals) contains peripheral physiological signals including PPG from 32 participants watching 40 music videos rated on arousal, valence, dominance, and liking scales. It is one of the most widely used benchmarks for training and evaluating PPG-based emotion recognition models.