Empatica EmbracePlus: Medical Sensor Platform

• Empatica EmbracePlus is a medical-grade, wrist-worn sensor platform that records multimodal data (PPG, EDA, temperature, accelerometry) for clinical and neuroscience applications.
• It employs precise synchronization, tailored filtering, and advanced feature extraction to enable real-time emotion recognition, cognitive assessment, and agitation prediction.
• The device integrates robust machine learning architectures and fusion pipelines, delivering actionable biomarkers and reliable sensor streams for diverse research studies.

The Empatica EmbracePlus is a medical-grade, wrist-worn sensor platform designed for continuous acquisition of multimodal physiological and movement signals. Integrating photoplethysmography, electrodermal activity measurement, tri-axial accelerometry, and skin temperature sensing, EmbracePlus is utilized in research for emotion recognition, cognitive assessment, and neuropsychiatric state monitoring via machine learning and advanced signal processing pipelines. Its flexible placement, reliable sensor streams, and integration with annotation and fusion architectures have established it as a cornerstone device for real-world affective and clinical neuroscience applications.

1. Sensor Modalities, Signal Specifications, and Hardware Integration

Empatica EmbracePlus includes four principal sensor modalities:

Modality	Physical Principle	Sampling Rate (Hz)
Photoplethysmography (PPG/BVP)	Microvascular pulse via reflected light	64
Electrodermal Activity (EDA)	Skin conductance via two gel electrodes	4
Skin Temperature (TEMP)	Thermistor (wrist)	4
Accelerometer (ACC_x, ACC_y, ACC_z)	MEMS tri-axial IMU	64

All modalities undergo hardware anti-aliasing, ADC digitization, and timestamping via on-board firmware. Data streaming is secured for cloud upload (HIPAA-compliant), but device resolution and dynamic range remain proprietary and are not specified in extant studies (Indrasiri et al., 3 Dec 2024, Habadi et al., 7 Nov 2025, Badawi et al., 26 Oct 2024). The device is typically worn on the non-dominant wrist, with EDA electrodes attached to the middle and ring fingers of the same hand, and strap tension adjusted to prevent slippage without occluding blood flow.

2. Data Acquisition, Filtering, and Preprocessing Pipelines

For each modality, raw signal acquisition is followed by tailored preprocessing steps:

• Synchronization: All EmbracePlus instances are coordinated to a network time protocol (NTP) server for precise timestamp alignment with external systems (e.g., VR headset, vest sensors) (Indrasiri et al., 3 Dec 2024).
• Filtering:
  • ACC: 0.5–20 Hz band-pass Butterworth filter removes drift/high-frequency noise.
  • BVP: 0.5–4 Hz band-pass isolates cardiac signals.
  • EDA: Upsampled from 4 Hz to 64 Hz; processed with 0.05 Hz high-pass (tonic shift removal) and 0.5 Hz low-pass.
  • TEMP: 1 s moving-average smoothing (channel-dependent).
• Normalization: After filtering, each channel $x(t)$ standardized as:

$x_{\mathrm{norm}} = \frac{x - \mu}{\sigma}$

where $\mu$ and $\sigma$ are calculated over dedicated time segments, yielding z-scored data.

• Segmentation: Data divided into sliding windows, e.g., 2 s windows with 50% overlap for 40 s intervals, yielding $T\times F$ tensors per channel. In cognitive studies, windows are further stratified, and wavelet transforms (e.g., Daubechies db4) decompose signals into localized time-frequency bands (Habadi et al., 7 Nov 2025).

3. Feature Extraction, Biomarker Derivation, and Machine Learning Integration

Signal features extracted from EmbracePlus are leveraged in statistical and deep learning frameworks:

• Statistical Features (per segment/wavelet band): mean, variance, standard deviation, spectral energy, min/max, skewness, kurtosis.
• Digital Biomarkers: heart rate variability indices (RMSSD, SDNN via PPG IBIs), skin conductance level (SCL), phasic peak count, accelerometer vector magnitude ($M(t)$), step count, movement intensity, temperature slope (Badawi et al., 26 Oct 2024).
• Feature Engineering: EmbracePlus signals serve as input to multi-scale attention-based LSTM architectures, where each modality $X^i\in\mathbb{R}^{T\times F}$ passes through two-layer LSTMs, followed by short-/medium-/long-term attention pooling and domain fusion via Squeeze-and-Excitation blocks (Indrasiri et al., 3 Dec 2024).
• Classical ML: Feature vectors enable regression/classification via regularized linear models (Ridge, Lasso, Elastic Net), Random Forests (with blending for subgroups), Extra Trees, XGBoost, and SVM with rigorous cross-validation and bootstrapping (Habadi et al., 7 Nov 2025, Badawi et al., 26 Oct 2024).

4. Representative Applications and Performance Metrics

A. Emotion Recognition in VR (Indrasiri et al., 3 Dec 2024)

• EmbracePlus signals serve as the “peripheral domain” in multi-domain affect recognition, fused with VR headset and trunk biosignals.
• LSTM+Attention+SE architectures integrate $ACC_z$, EDA, TEMP for unimodal classification.
• Performance:
  • Peripheral-only (watch) valence: $G_1$_V: 55.85–61.83%, $G_2$_V (expert): up to 73.58%
  • Full fusion (all domains): $G_2$_V: 86.77%; arousal: up to 65.15%
  • SE-blocks improved valence by 2–3 points and arousal by 1–2 points.

B. Cognitive Assessment Prediction (Habadi et al., 7 Nov 2025)

• EmbracePlus used to predict NIH Toolbox cognitive battery scores under supervised learning:
  • Working Memory: Ridge model on BVP, skin temperature, ACC_x; Spearman’s $\rho=$0.822, MAE=0.143.
  • Processing Speed: Lasso, notable synergy between accelerometry and EDA ($\rho\sim0.73$).
  • Attention: Random Forest + wavelets on BVP/EDA/temperature/ACC_x; $\rho=0.81$, MAE=0.104.
• Device proved robust, yielding strong correlations and MAEs significantly lower than naive mean predictors.

C. Agitation Prediction in Dementia (Badawi et al., 26 Oct 2024)

• EmbracePlus deployed for minute-by-minute agitation monitoring; feature vector $\mathbf{x}\in\mathbb{R}^{150}$ includes HRV, SCL, movement intensity, and more.
• Extra Trees classifier achieved calibrated individual models:
  • Accuracy: 90–99%
  • AUC: 0.980–0.993
  • Sensitivity/Specificity: Approaching 99%
• Pre-agitation windows identified: tonic EDA, accelerometer intensity, and skin temperature slope diverged from baseline $\approx$6 minutes before overt agitation.

5. Practical Implementation Considerations and Limitations

• Data Integrity: Stable skin contact for EDA electrodes is essential; artifacts due to electrode lift are addressed by filtering/upsampling, but no formal artifact rejection was performed.
• Device Metadata: Proprietary limitations on resolution and ADC range; only sampling rates leveraged in published studies.
• Battery/Buffering: Wristband operates for over 24 hours, buffering raw streams, and offloads to smartphone/cloud with $<$1 minute latency.
• Sample Sizes: Studies range from $n=3$ (pilot, agitation) to $n=23$ (emotion/cognition), potentially constraining generalizability and statistical power.
• Edge/Cloud: All classification has run in the cloud; no on-device inference currently available.
• Annotation: Precise event labeling achieved via synchronized, privacy-preserving video for agitation onset/offset.

6. Implications, Future Directions, and Open Challenges

EmbracePlus advances the research frontier for continuous, noninvasive physiologic monitoring with diverse modalities. Studies have demonstrated:

• Feasibility of real-time affect estimation and agitation warning systems, laying the groundwork for timely intervention in dementia care.
• Predictive biomarker identification for cognitive fatigue and domain-specific effort.
• Robust multi-modal architectures (LSTM, SE-blocks) and classical ML pipelines (ensemble trees, wavelet feature engineering), with validated results under small-sample conditions.
• Unresolved technical gaps include full disclosure of sensor resolution, adaptation of preprocessing/filter protocols to naturalistic use, and development of on-device inference for immediate feedback.

A plausible implication is that as larger, naturalistically annotated datasets accrue, model generalizability and decoder specificity will be enhanced, establishing EmbracePlus as both a research and clinical tool for personalized physiologic state monitoring. Further integration with home platforms and extension to diverse populations may reveal circadian, behavioral, and environment-driven signal patterns.