Empatica E4: Multimodal Physiological Sensor

• Empatica E4 is a wrist-worn multimodal physiological sensor that captures raw signals including BVP, EDA, skin temperature, and accelerometry for diverse applications.
• It supports both classical feature-based analysis and deep learning, enabling robust assessments in stress, emotion recognition, and cognitive performance studies.
• The device is deployed in controlled and free-living environments, requiring thorough preprocessing and subject-specific calibration to address motion artifacts and signal quality.

Empatica E4 is a wrist-worn multimodal physiological sensing device used extensively in psychophysiology, affective computing, wearable health monitoring, and digital phenotyping. Across the literature, it is described as providing raw or device-derived streams for blood volume pulse (BVP/PPG), electrodermal activity (EDA), skin temperature, accelerometry, heart rate (HR), interbeat interval (IBI), and time-stamped tagging, and it has been deployed in both controlled experiments and free-living observation for tasks including stress assessment, emotion recognition, classroom engagement analysis, mood episode detection, and driver-state modeling (Saganowski et al., 2020, Nikseresht et al., 2021, Azghan et al., 24 Mar 2025).

1. Sensor platform and reported device characteristics

In a comparative heart-rate study, the E4 is described as a medical-grade (class 2a) wristband with an optical PPG sensor, 2 green and 2 red LEDs, 2 photodiodes, a 3-axis accelerometer, EDA, and infrared skin temperature, and no gyroscope (Choksatchawathi et al., 2019). Other studies emphasize the device’s practical value for research not primarily because of embedded analytics, but because it exposes raw physiological channels central to autonomic inference: BVP, EDA, skin temperature, and acceleration, alongside derived HR and IBI streams and, in some workflows, event tags (Saganowski et al., 2020, Nikseresht et al., 2021).

Most studies report the E4 with BVP at 64 Hz, EDA at 4 Hz, acceleration at 32 Hz, and HR at 1 Hz, with IBI derived from BVP; several papers also report skin temperature at 4 Hz as a standard device setting, while one clinical mood-monitoring study reported TEMP at 1 Hz in its exported data representation (Azghan et al., 24 Mar 2025, Puertas-Ramirez et al., 13 Apr 2026, Shahriar, 5 Dec 2025, Corponi et al., 2023). This mixture of raw and derived channels has shaped downstream methodology: some pipelines work directly on BVP, EDA, temperature, and accelerometry, whereas others intentionally exclude HR and IBI because they are proprietary derivatives of BVP (Corponi et al., 2023).

A notable device-specific behavior is the E4 HR algorithm’s conservative handling of low-quality PPG. One study reports that the device “restrains the output of HRPPG if the quality of the observed R-R interval is not detectable,” which reduces spurious output during heavy motion but also produces missing HR values (Choksatchawathi et al., 2019). This property is consequential: it can improve robustness for low-motion data while complicating continuous monitoring under vigorous activity.

2. Deployment regimes and experimental use

The E4 has been used across markedly different acquisition regimes. In laboratory settings, it has been worn during a race-focused Implicit Association Test by 46 undergraduate students, during the Trier Social Stress Test in older adults, during standardized stress-induction protocols involving mental arithmetic, startle response, and cold pressor tasks, and as the wrist reference sensor in WESAD-derived PPG stress studies (Nikseresht et al., 2021, Onim et al., 10 Jul 2025, Amin et al., 9 May 2025, Hasanpoor et al., 2024). In field or semi-naturalistic settings, it has been worn during ordinary classroom lessons over four weeks, during one full day of free-living behavior in cannabis-use research, across many weeks in ecological affect logging, for approximately 48 hours in clinical mood-disorder monitoring, and throughout real-world SAE Level 2 automated driving sessions (Ananthan et al., 2024, Azghan et al., 24 Mar 2025, Saganowski et al., 2020, Corponi et al., 2023, Puertas-Ramirez et al., 13 Apr 2026).

Wearing protocol is generally wrist-based, often on the non-dominant wrist when continuous daily monitoring is intended (Azghan et al., 24 Mar 2025, Nešković et al., 2 May 2025, Corponi et al., 2023). In classroom and wellbeing studies, the E4’s unobtrusive form factor supported repeated or longitudinal capture under ordinary activity, although comfort and user convenience were reported as weaker than for mainstream smartwatches (Ananthan et al., 2024, Saganowski et al., 2020).

Synchronization strategies vary substantially by study. Some designs rely on the E4’s shared device time axis across channels and align modalities by timestamps or identical filenames after export (Puertas-Ramirez et al., 13 Apr 2026). Others anchor physiology to external events via self-reported timestamps for sleep, cannabis consumption, exercise, and perceived stress (Azghan et al., 24 Mar 2025). By contrast, the racial-bias study did not use explicit hardware or software markers to align physiology to IAT events and instead segmented the entire recording into fixed 5-second windows with participant-level label inference performed after sequence smoothing (Nikseresht et al., 2021). This diversity of alignment practice is methodologically important because the E4 itself supplies continuous streams, but task relevance depends on how those streams are temporally linked to stimuli, events, or labels.

3. Preprocessing, quality control, and feature construction

E4-based pipelines span simple filtering through modality-specific decomposition and representation learning. For EDA, multiple studies perform tonic-phasic separation: one uses cvxEDA to decompose raw EDA into tonic and phasic components before extracting AUC, RMS, peak, and variability features, while another applies NeuroKit filtering, artifact removal, baseline correction, and tonic-phasic decomposition with phasic EDA retained for synchrony analysis (Nikseresht et al., 2021, Ananthan et al., 2024). In stress reproducibility work, EDA is range-filtered to remove values below 0.01 $\mu S$ or above 100 $\mu S$, smoothed with a 5-second median filter, normalized to $[0,1]$, and decomposed with cvxEDA (Amin et al., 9 May 2025).

For cardiovascular processing, the literature alternates between explicit HRV extraction and direct raw-signal learning. Hand-crafted pipelines compute SDNN, RMSSD, pNN50, pNN20, SDSD, HR median absolute deviation, and frequency-domain measures such as LF/HF, often from IBI or BVP-derived beat intervals (Nikseresht et al., 2021, Ananthan et al., 2024). By contrast, a PPG-only stress study on WESAD does not compute HRV features at all; it feeds normalized, filtered 64 Hz PPG frames directly into an adaptive 1D CNN+MLP architecture (Hasanpoor et al., 2024).

Segmentation is similarly heterogeneous. Reported windows include 5-second non-overlapping windows for IAT physiology, 60-second windows for stress events and ecological affect labeling, 30-second windows with overlaps or strides for driver-state and stress/exercise benchmarks, 10-second windows for cognitive-load benchmarks, and 512-second windows with 128-second stride for mood-disorder detection (Nikseresht et al., 2021, Azghan et al., 24 Mar 2025, Saganowski et al., 2020, Puertas-Ramirez et al., 13 Apr 2026, Shahriar, 5 Dec 2025, Corponi et al., 2023). These choices reflect different assumptions about the temporal scale of the target phenomenon: phasic arousal, task-level engagement, sustained driver awareness, or mood state.

Accelerometry is commonly used both as signal and nuisance variable. One study summarizes micro-body movement by the magnitude

$\sqrt{x^2 + y^2 + z^2},$

while others extract per-axis means and standard deviations or use acceleration-derived physical activity levels for downstream modeling (Nikseresht et al., 2021, Shahriar, 5 Dec 2025, Choksatchawathi et al., 2019). Several papers explicitly note that accelerometry was available but not always used for artifact gating, even when movement likely affected EDA or PPG (Ananthan et al., 2024, Azghan et al., 24 Mar 2025).

Recent work also replaces classical features with structured representations. Real-world driver-state modeling resampled all E4 channels to 4 Hz, scaled them to image intensities, and converted each 30-second window into recurrence plots, Gramian Angular Fields, or Markov Transition Fields before late fusion through pre-trained ResNet50 backbones (Puertas-Ramirez et al., 13 Apr 2026). Mood-disorder detection instead used an E4-specific Transformer, “E4mer,” whose channel-specific convolution and pooling layers match native E4 sampling frequencies, allowing ACL, BVP, EDA, and TEMP to be fused without explicit resampling (Corponi et al., 2023).

4. Research applications and empirical findings

The E4 has been used as a core sensor platform in bias detection, where multimodal physiology collected during the IAT predicted implicit racial bias with 76.1% accuracy and an F1 of 75.8% under leave-one-participant-out cross-validation; EDA features dominated model importance, and EDA standard deviation was reported as the most differentiating feature between biased and unbiased participants (Nikseresht et al., 2021). The same study found that, among accurately labeled biased participants, 66% showed the largest block of biased windows at the end of the IAT, which the authors interpreted as suggesting increasing sympathetic arousal with task progression (Nikseresht et al., 2021).

In educational monitoring, the E4 has supported both feature-level and synchrony-level analyses. A classroom study using EDA and IBI from 23 students found significant course-dependent differences in intra-student synchrony: one-way ANOVA on FastDTW distances yielded $F = 3.711$ with $p = 3.163 \times 10^{-7}$, Assembly showed the lowest synchrony distance at 3.3, and Physical Education the highest at 195.3 (Ananthan et al., 2024). The same study reported an Assembly-specific positive correlation between RMSSD and SCR peak amplitude mean of $r = 0.54$, emphasizing that E4-derived ANS signals can reflect context-dependent combinations of sympathetic and parasympathetic activity rather than a single monotonic arousal axis (Ananthan et al., 2024).

Stress and affect recognition constitute a major application cluster. In the WellAff system, E4-derived BVP, EDA, SKT, and ACC features supported a binary strong-affect versus neutral classifier with 91% F1 in everyday life (Saganowski et al., 2020). In a controlled undergraduate stress study, Empatica E4 achieved leave-one-subject-out AUROC values of 0.884 for rest versus all stressors with HRV+EDA and 0.953 for rest versus mental arithmetic with HRV+EDA (Amin et al., 9 May 2025). A PPG-only WESAD study reported 96.7% subject-dependent testing accuracy and 92.1% subject-independent testing accuracy for binary stress versus no-stress classification using only the E4’s 64 Hz PPG stream (Hasanpoor et al., 2024).

The device has also been central in emotion and mental-health studies. In older adults undergoing a structured stressor, physiological signals from E4 and Shimmer3, with iMotions facial expression intensities as ground truth, supported regression of Negative, Neutral, and Positive emotion intensities; Random Forest achieved $R^2 = 0.782$, 0.7636, and 0.8033, with MSE values of 0.0006, 0.0014, and 0.0019, respectively (Onim et al., 10 Jul 2025). In clinical mood-disorder monitoring, the SSL-pretrained E4mer achieved 81.23% segment accuracy and 90.63% subject accuracy for acute episode versus euthymia, outperforming both a fully supervised E4mer and an XGBoost baseline on FLIRT features (Corponi et al., 2023).

Free-living descriptive physiology has also been documented. The CAN-STRESS dataset recorded E4 signals from 82 adults over one day and reported higher mean EDA, higher EDA peaks per minute, and higher mean HR in chronic cannabis users than in non-users; stress-labeled 60-second windows showed EDA increasing from 0.85 $\pm$ 0.11 in no stress to 1.79 $\pm$ 0.08 in high stress, and HR increasing from 85.23 $\mu S$0 2.79 in low stress to 89.55 $\mu S$1 4.63 in high stress (Azghan et al., 24 Mar 2025).

In automated driving, E4-based personalized driver-state models substantially outperformed generalized models. A late-fusion image-based pipeline built from E4 BVP, EDA, TEMP, HR, and ACC streams achieved an average personalized accuracy of 92.68%, whereas combined-user generalized models averaged 54.00% accuracy (Puertas-Ramirez et al., 13 Apr 2026). This finding is especially notable because the same paper reports cross-user test accuracies as low as 40.62–63.52%, indicating that physiological signatures acquired by the E4 can be strongly subject-specific in safety-critical settings (Puertas-Ramirez et al., 13 Apr 2026).

5. Modeling paradigms and validation practice

E4 research spans three broad modeling regimes: classical feature-based learning, raw-signal deep learning, and self-supervised or representation-learning pipelines. Classical feature-based work frequently uses XGBoost, Random Forest, SVM with RBF kernel, AdaBoost-wrapped tree models, and kNN or regression baselines on windowed physiological descriptors (Nikseresht et al., 2021, Amin et al., 9 May 2025, Saganowski et al., 2020, Onim et al., 10 Jul 2025). These pipelines typically combine EDA features with HRV, temperature, and accelerometry, and they often rely on subject-separated validation such as LOPO-CV or LOSO to reduce leakage (Nikseresht et al., 2021, Amin et al., 9 May 2025).

Raw-signal deep learning has been pursued both unimodally and multimodally. The PPG-only stress paper used an adaptive 1D CNN+MLP on filtered 64 Hz PPG frames from the E4, avoiding hand-crafted HRV altogether (Hasanpoor et al., 2024). Driver-state modeling transformed each modality into a two-dimensional image and fused seven ResNet50 feature extractors through PCA and an MLP head (Puertas-Ramirez et al., 13 Apr 2026). Mood-disorder work used a Transformer encoder specialized to the E4’s heterogeneous sampling rates, then strengthened it with self-supervised masked prediction or transformation prediction on a large unlabelled E4 corpus before supervised fine-tuning (Corponi et al., 2023).

A recurrent finding is that nonlinear models dominate linear baselines on E4 data. A unified benchmark over three public Empatica E4 datasets reported nonlinear-model accuracy of 0.89–0.98 and ROC-AUC of 0.96–0.99 across stress, exam performance, and cognitive-load tasks, whereas logistic regression remained at AUC $\mu S$2–0.73 (Shahriar, 5 Dec 2025). The same study found multimodal fusion consistently best and showed that removing EDA or TEMP caused large statistically significant drops across datasets, while ACC removal particularly damaged exercise classification (Shahriar, 5 Dec 2025). This suggests that the E4’s value lies less in any single channel than in the joint structure of autonomic, thermal, and movement signals.

Validation design remains a major differentiator of reported performance. Subject-dependent training can yield very high scores, as in the 96.7% single-subject stress result from PPG alone, but subject-independent or cross-user evaluation usually produces lower and more realistic estimates (Hasanpoor et al., 2024). The driver-monitoring and unified-benchmark studies both show substantial inter-individual variability under LOSO-style testing, and the clinical mood-disorder paper similarly emphasizes that self-supervised pretraining on heterogeneous unlabelled E4 data can partly compensate for limited labeled cohorts (Puertas-Ramirez et al., 13 Apr 2026, Shahriar, 5 Dec 2025, Corponi et al., 2023).

6. Measurement limitations, bias, and methodological directions

The E4’s most persistent technical limitation is sensitivity to motion and contact quality. Comparative heart-rate evaluation against ECG reported E4 mean absolute errors of 4.05 $\mu S$3 0.34 bpm in resting, 3.55 $\mu S$4 0.34 bpm in laying down, 5.45 $\mu S$5 0.48 bpm in intense treadmill, and 3.90 $\mu S$6 0.31 bpm overall, with missing HR output under heavy motion because the device suppresses low-quality estimates (Choksatchawathi et al., 2019). A later reproducibility study likewise notes that E4 EDA can suffer frequent artifacts and data loss from loss of skin contact due to body motion, and this study found strong within-study performance but weaker transfer of pre-trained HRV+EDA models to a new E4 deployment (Amin et al., 9 May 2025). These results argue against treating E4-derived HR or EDA as uniformly interchangeable across studies, devices, or use conditions.

Pulse-rate reference use has also been scrutinized. In a driving-simulator comparison between camera-based rPPG and the E4, missing BVP/PR/IBI segments from poor sensor-skin contact and subject movement led to the exclusion of 17 videos, E4 BVP SNR was 7.67 $\mu S$7 2.18 dB, and the absolute difference between rPPG-derived PR and Empatica E4 PR increased approximately linearly with Empatica PR (Nešković et al., 2 May 2025). The authors modeled this with linear fits and explicitly cautioned that the E4 cannot be considered a laboratory-grade ground truth under such dynamics (Nešković et al., 2 May 2025). This suggests that calibration against ECG or other high-quality references is advisable when E4 pulse estimates are used for benchmarking rather than merely for relative monitoring.

Methodological limitations extend beyond sensor physics. Some studies lack explicit event markers or baseline calibration, use very short windows for HRV, omit formal artifact rejection, or collapse nuanced labels into binary outcomes because of sample-size constraints (Nikseresht et al., 2021, Azghan et al., 24 Mar 2025). Classroom work did not use ACC for artifact gating despite high-motion contexts such as Physical Education, and several free-living studies relied on self-reported timestamps for aligning events to physiology (Ananthan et al., 2024, Azghan et al., 24 Mar 2025). In clinical and affective computing, privacy, stigmatization risk, and responsible data use are explicit concerns, especially when physiology is linked to bias, emotion, stress, or psychiatric state (Nikseresht et al., 2021, Azghan et al., 24 Mar 2025, Corponi et al., 2023).

The methodological direction emerging from the literature is relatively consistent. Recommended practices include adding explicit event markers when possible, collecting baseline physiology, using robust EDA decomposition and motion-aware quality control, favoring subject-level normalization, and evaluating with subject-separated protocols rather than pooled windows (Ananthan et al., 2024, Amin et al., 9 May 2025, Shahriar, 5 Dec 2025). For tasks with marked individual differences, personalized or adaptive modeling appears particularly important: the driver-state results, the LOSO variability in unified benchmarks, and the gains from self-supervised E4-specific pretraining all point in that direction (Puertas-Ramirez et al., 13 Apr 2026, Shahriar, 5 Dec 2025, Corponi et al., 2023).