How Can a Radar Detect Teeth Grinding Without Touching You?

Bruxism is that annoying habit of grinding teeth, often while sleeping, which can ruin rest and damage teeth. Until recently, detecting it required contact sensors or cameras, solutions that are uncomfortable and sometimes invasive. Now researchers have tried using a millimeter‑wave radar, similar to those in 5G phones, to "listen" to the jaw's micro‑movements. By analyzing the signal echo, they identified 11 typical features of grinding. With this data they trained a Random Forest model, an AI algorithm, which recognized bruxism with 96% accuracy on a test set. The results show high precision, recall and F1‑score, proving the radar can separate sleep noise from teeth‑grinding sounds. The advantage is clear: no contact, no camera, so more privacy and less hassle. The authors now aim to make the system more robust, able to work in different settings and filter out other facial micro‑movements. If successful, we could soon have bedside or wearable devices that alert us in real time when we start grinding, helping us sleep better and protect our teeth.

Millimeter‑wave radar operates in the 30‑300 GHz band, emitting picosecond pulses. When these pulses hit the mandibular area, their echo undergoes phase and amplitude variations proportional to dental surface movements. Spectral analysis of these changes allows extraction of temporal parameters (time‑of‑flight, jitter) and frequency metrics (vibration spectrum). The 11 selected features include: average echo amplitude, standard deviation, inter‑frame correlation coefficient, energy in the 10‑30 Hz range (typical of bruxism), and spectral entropy metrics. The Random Forest model combines decision trees that, via bagging, reduce overfitting and improve generalization. Five‑fold cross‑validation yielded an ROC curve with AUC = 0.98, indicating excellent discrimination. However, the system is sensitive to unrelated micro‑movements (yawning, eye blinks). To mitigate these effects, adaptive filters based on independent component analysis (ICA) and convolutional neural networks are being tested to recognize noise patterns. Moreover, personalized calibration using a short reference recording can enhance robustness across heterogeneous populations. Integration with IoT devices would enable continuous data collection, paving the way for longitudinal bruxism studies and real‑time feedback‑based therapeutic interventions.