Radar-Based Contactless Bruxism Recognition

• Contactless bruxism recognition is a non-invasive method that uses FMCW radar and phase-difference processing to detect sleep-related teeth grinding.
• It employs spatial filtering, feature extraction, and a Random Forest classifier to achieve robust classification with 96.1% accuracy.
• The approach outperforms traditional EMG and audio-based systems, ensuring privacy and enabling unobtrusive long-term monitoring.

A contactless bruxism recognition system leverages millimeter-wave radar to detect sleep-related teeth grinding (bruxism) with high accuracy and privacy preservation. Bruxism, an oromandibular movement disorder characterized by teeth grinding and clenching, detrimentally impacts sleep quality and dental health. Traditional diagnostic techniques, such as electromyography (EMG) and intraoral sensors, pose drawbacks regarding comfort and privacy. The radar-based approach utilizes the phase sensitivity of frequency-modulated continuous-wave (FMCW) signals to capture sub-millimeter jaw movements, hostile to invasive and audio-based alternatives. The latest system demonstrates that phase-difference signal processing and feature engineering enable robust classification, achieving 96.1% accuracy and outperforming established methods on key recognition metrics (Shen et al., 7 Dec 2025).

1. System Architecture and Hardware Configuration

The system architecture centers on the Texas Instruments IWR6843 FMCW radar platform operating in the 60–64 GHz spectrum. Typical parameters include a carrier frequency $f_c \approx 62$ GHz, chirp bandwidth $B = 4$ GHz, and range resolution $\Delta R = c/(2B) \approx 3.75$ cm. The radar employs a MIMO array with 3 transmit and 4 receive antennas, though a single Tx–Rx channel suffices, balancing complexity against data volume.

Installation involves positioning the radar 55 cm from the subject’s face, slightly above head height and perpendicular to the facial plane. This configuration resolves trade-offs between near-field distortion and signal attenuation. Physiologically, bruxism events—manifest as masseter bulging (0.5–1.5 Hz) and rapid jaw stick-slip oscillations (5–10 Hz)—induce phase shifts in the returned radar signal. Tracking these temporal phase shifts yields a sensitive, noncontact measure of bruxism-specific jaw movements.

2. Signal Processing Pipeline

The signal processing pipeline is structured into three sequential stages:

1. Range-Domain Reorganization: Raw ADC samples are fragmented into an IQ matrix $X \in \mathbb{C}^{N_c \times N_s}$, with $N_c$ chirps in slow time and $N_s$ samples per chirp (fast time).
2. Spatial Filtering via Range FFT and Energy Localization: For each chirp $n$, a 1D fast-time FFT forms the range profile:

$$X_n(k) = \sum_{m=0}^{N_s-1} x_n[m] \cdot e^{-j2\pi (km/N_s)}, \quad k = 0, ..., N_s-1$$

The incoherent range-bin power is:

$P(k) = \sum_{n=1}^{N_c}|X_n(k)|^2$

The facial range is determined as

$k^* = \arg\max_k P(k)$

1. Phase-Difference Extraction: At the selected bin $B = 4$0, the wrapped phase is:

$B = 4$1

Phase unwrapping and differencing produce the dynamic jaw motion signal:

$B = 4$2

This differenced sequence filters out respiratory and head-motion drift, isolating bruxism-specific microvibrations.

3. Feature Engineering

Eleven features are extracted from the phase-difference sequence for classification, grouped by domain:

Time-Domain Features

Index	Description	Formula
F₁	Absolute mean	$B = 4$3
F₂	Variance	$B = 4$4
F₃	Kurtosis	$B = 4$5
F₄	Shannon entropy of histogram	$B = 4$6

Spectral Features

Index	Description	Formula
F₅	Spectral entropy	$B = 4$7
F₆	Spectral variance	$B = 4$8
F₇	Band energy (5–10 Hz)	$B = 4$9

Peak-Structure Descriptors

Index	Description
F₈	Number of local maxima
F₉	Number of local minima
F₁₀	Count of $\Delta R = c/(2B) \approx 3.75$0 rad
F₁₁	Count of $\Delta R = c/(2B) \approx 3.75$1 rad

High-frequency energy (F₇) maps stick-slip grinding, absolute mean (F₁) and peak counts (F₈–F₁₁) quantify masseter bulging amplitude, while entropy and kurtosis (F₃–F₅) measure the stochasticity of grinding micro-motions.

4. Dataset Construction and Protocol

The dataset comprises 180 five-second recordings from three healthy volunteers, equally split into grinding and non-grinding sessions. The radar was deployed at a 55 cm range, with subjects maintaining a relaxed posture, and only anterior–posterior jaw movements were enacted in grinding trials. Ground-truth labeling was conducted in real time by subjects under supervision. All samples underwent a 10-fold cross-validation protocol, allocating 162 for training and 18 for testing per fold.

5. Classification Strategy

Feature vectors $\Delta R = c/(2B) \approx 3.75$2 populate a Random Forest classifier configured as follows:

• 90 trees (estimators)
• Node split criterion: Gini impurity
• Minimum samples per split: 2
• Bootstrap sampling with random feature subset selection at each node
• 10-fold cross-validation to mitigate overfitting

At each node, information gain is maximized over a random feature subset $\Delta R = c/(2B) \approx 3.75$3:

$\Delta R = c/(2B) \approx 3.75$4

Predictions aggregate via majority voting across trees:

$\Delta R = c/(2B) \approx 3.75$5

6. Performance Assessment and Comparative Results

Evaluation utilizes standard metrics:

• Accuracy: $\Delta R = c/(2B) \approx 3.75$6
• Precision: $\Delta R = c/(2B) \approx 3.75$7
• Recall: $\Delta R = c/(2B) \approx 3.75$8
• F1-score: $\Delta R = c/(2B) \approx 3.75$9

The contactless system achieved:

Metric	Grinding	No-grinding
Precision	0.966	0.956
Recall	0.967	0.956
F1-score	0.961	0.961

Overall accuracy reached 96.1%. The normalized confusion matrix is:

Predicted / Actual	No-grinding	Grinding
No-grinding	0.956	0.044
Grinding	0.034	0.966

Comparative experiments determined that the radar system surpassed contemporary EMG-based and audio-based recognition systems across these metrics.

7. Discussion, Limitations, and Prospects

Signal isolation via spatial filtering and phase differencing confers resilience against typical indoor interference and subject variability. By eschewing electrodes and microphones, this approach preserves user privacy and facilitates unobtrusive long-term monitoring. Notably, radar does not record personally identifiable audio or require physical contact.

Limitations emerge in potential false positives due to overlapping spectral components from other facial micro-motions (e.g., chewing, talking, tremors) and limited cohort diversity (three subjects).

Planned advancements comprise expanding cohorts, developing adaptive clutter suppression and multi-antenna beamforming, evaluating deep-learning architectures (e.g., CNNs on time-frequency maps), and deploying in real sleep environments with multimodal fusion (posture sensors, etc.). This suggests ongoing research will focus on robustness, generalizability, and real-time deployment.

In summary, millimeter-wave FMCW radar integrated with phase-difference analysis and Random Forest classification substantiates a high-fidelity, privacy-conscious bruxism monitor, achieving over 96% accuracy in indoor environments and presenting a credible alternative to traditional contact-based techniques (Shen et al., 7 Dec 2025).