@article{huang2025reconstructing,
author = {Huang, Yongzhi and Zhao, Jiayi and Wu, Kaishun},
title = {Reconstructing Ear Canal Channels for Fine-Grained Detection of Tympanic Membrane Changes},
year = {2025},
issue_date = {September 2025},
publisher = {Association for Computing Machinery},
address = {New York, NY, USA},
volume = {9},
number = {3},
url = {https://doi.org/10.1145/3749516},
doi = {10.1145/3749516},
abstract = {Recent work has begun to leverage commercial headphones for ear canal health monitoring. However, existing solutions are limited to detecting coarse abnormalities using single-frequency probe tones and specialized hardware. Detecting fine-grained conditions---such as tympanic membrane retraction---remains a significant challenge due to anatomical variability, the limitations of low-sensitivity microphones, and the uncontrolled nature of real-world audio. We present EarCSI, a reconstruction-driven framework that enables precise ear-canal sensing from passive broadband audio using commodity headphones. The core of our system is a lightweight frequency-domain Channel Reconstruction Module, which models the ear canal geometry by analyzing spectral features such as peak spacing and angular propagation behavior. To achieve this, we design a set of novel estimation techniques, including peak-trough-based coarse length inference, spectral angle-based shortest path estimation, and a reflection-aware transfer matrix model that captures cumulative impedance effects. These methods allow the system to reconstruct user-specific ear canal profiles without per-user training or access to invasive scans. Through modeling and experimentation, we uncover a critical constraint: reliable reconstruction requires signal duration, even for short ear canals. To ensure robustness in daily scenarios, we further introduce signal-level and distribution-level interference mitigation strategies that compensate for background noise, headphone misalignment, nonlinearity, and environmental drift. Ultimately, a low-latency classifier extracts health-related features from the reconstructed frequency response and accurately detects tympanic retraction.EarCSI achieves over 95\% classification accuracy and under 5\% reconstruction error across 88 ears using multiple commercial headphones, operating in real-time (168 ms latency). It enables passive and continuous monitoring of tympanic responses during everyday listening, offering a new pathway for personalized auditory health sensing.},
journal = {Proc. ACM Interact. Mob. Wearable Ubiquitous Technol.},
month = sep,
articleno = {89},
numpages = {48},
keywords = {Accelerated Matrix Operations, Acoustic Sensing, Ear Canal Channel Reconstruction, Earable Computing, Hearing Protection, Mobile and Ubiquitous Sensing, Real-Time Signal Processing, Sparse FIR Modelling, Tympanic Reflex Detection, Wearable Health Monitoring}
}