BCough: Design and Evaluation of a Bone-Conduction-Embedded AI Platform for On-Device Cough Detection

Continuous cough monitoring provides valuable insights into respiratory health; however, conventional air-conduction microphones are highly susceptible to ambient noise and raise privacy concerns. This work presents BCough, a bone-conduction-based embedded AI platform for on-device cough detection, designed and evaluated on the MAX78000 neural accelerator. The system employs a skin-contact bone-conduction sensor worn on the chest to capture body vibrations transmitted through bone and tissue, detecting cough events while minimizing environmental interference. The custom prototype integrates a bone-conduction microphone, a synchronized 6-axis IMU, power management circuitry, and an embedded neural accelerator to support real-time inference and future multimodal extensions. A compact 8-bit quantized convolutional neural network was optimized for deployment on the MAX78000 and evaluated using leave-one-subject-out cross-validation on one-second cough and non-cough segments derived from a corpus of 19,955 labeled events collected from five participants under controlled conditions. The deployed model achieved 0.80 Macro-F1, 0.81 balanced accuracy, 0.74 cough F1, and 0.89 AUC, with 15–16 ms inference latency and approximately 20 μJ energy per inference on chip. These results demonstrate the feasibility of low-power, privacy-preserving, bone-conduction cough detection on embedded AI hardware within an initial five-participant study. The current design is a benchtop prototype; the findings should therefore be interpreted as an initial feasibility assessment rather than evidence of robust performance across diverse users and real-world conditions. Future work will extend this platform toward miniaturized wearable implementations combining bone-conduction and inertial sensing for continuous multimodal respiratory monitoring.