CHOMP: Multimodal Chewing Side Detection with Earphones

TL;DR

CHOMP introduces the first earphone-based system for continuous chewing-side detection by leveraging OpenEarable 2.0’s multimodal sensing and a wavelet-based time-frequency CNN pipeline. Through 20 participants and diverse eating scenarios, including noisy environments and a full meal, it demonstrates that microphones provide the strongest single-sensor performance and that sensor fusion with a bone-conduction mic and IMU yields higher accuracy while remaining robust on device. The approach achieves median $F_1$ scores of $94.5\%$ (LOFO) and $92.6\%$ (LOSO) with microphones alone, rising to $97.7\%$ (LOFO) and $95.4\%$ (LOSO) with fusion, and maintains performance under audio interference. The results suggest CHOMP can enable unobtrusive, longitudinal CSP monitoring in real-world settings, offering clinicians and patients objective insights into jaw function and CSP-related TMD risk, with feasible edge deployment on smartphones. Limitations include seated eating, lack of bilateral chewing, and a dentition-homogeneous sample, pointing to directions like test-time adaptation and personalization for broader applicability.

Abstract

Chewing side preference (CSP) has been identified both as a risk factor for temporomandibular disorders (TMD) and behavioral manifestation. Despite TMDs affecting roughly one third of the global population, assessment mainly relies on clinical examinations and self-reports, offering limited insight into everyday jaw function. Continuous CSP monitoring could provide an objective proxy for functional asymmetries. Prior wearable approaches, however, mostly use specialized form factors and demonstrate limited performance. We therefore present CHOMP, the first system for chewing side detection using earphones. Employing OpenEarable 2.0, we collected data from 20 participants with microphones, a bone-conduction microphone, IMU, PPG, and a pressure sensor across eleven foods, five non-chewing activities, and three noise conditions. We apply the Continuous Wavelet Transform to each sensing modality and use the resulting multi-channel scalograms as inputs to CNN-based classifiers. Microphones achieve the strongest single-sensor unit performance, with median F1 scores of 94.5% in leave-one-food-out (LOFO) and 92.6% in leave-one-subject-out (LOSO) cross-validations. Fusing sensing modalities further improves performance to 97.7% for LOFO and 95.4% for LOSO, with additional evaluations under noise interference indicating robust performance. Our results establish earphones as a practical platform for continuous CSP monitoring, enabling clinicians and patients to assess jaw function in everyday life.

Figures (13)

• Figure 1: A. The anatomical structures relevant to TMD are located in close proximity to the ear (adapted from university_of_leeds_improving_2025); B. Sensing modalities integrated in OpenEarable 2.0 and their rationale for effective chewing side detection, along with supporting literature on earphone-based chewing recognition.
• Figure 2: Experimental overview. Participants completed a pre-questionnaire (see \ref{['app: pre_quest']}), performed four experimental blocks while wearing OpenEarable 2.0 devices, and then filled out a post-questionnaire (see \ref{['app: post_quest']}). Block 1 included five non-chewing activities (see \ref{['sec: selection_non-chew']}) while Blocks 2–4 comprised different chewing tasks on the left and right side: consumption of ten single food items (see \ref{['sec: selection_foods']}), a whole meal (see \ref{['sec: whole_meal']}), and two selected foods under three audio noise conditions (see \ref{['sec: external_validity']}).
• Figure 3: A. Definitions and procedure of determining brittleness and chewiness; B. Brittleness and chewiness of the selected foods. Nodes depict the means while the whiskers represent standard deviations.
• Figure 4: Selection of different single food items (left) and an example of a whole meal (right).
• Figure 5: A. The scalograms across all sensor channels, exemplified here by the inner microphones, reveal inconsistent log-power value distributions between the left and right earables, illustrated here as histograms with sample size $N$. This discrepancy manifests differently for each subject, suggesting that sensor bias is driven by subject-specific physiology and physical placement rather than just hardware offsets.; B. Applying standardization independently per subject, earable, and sensor channel centers these divergent distributions, enhancing robustness to inter-subject variability.