Zuchen Li

All projects

CHI 2027 under review Ubiquitous sensing

RespiraFrame

On-Frame Respiratory Monitoring via Electrical Impedance Tomography and Bone-Conduction Acoustics

Experimental design & evaluation, Human Centered Computing Lab, University of Michigan, with Prof. Junyi Zhu

An overview of RespiraFrame: the system recognizes (a) discrete abnormal respiratory events and (b) continuous breathing behaviors, and (c) remains robust across diverse, challenging everyday conditions.

Our breathing says a lot about our health. Coughs, sneezes and sniffs, and whether we breathe through the nose or the mouth, deeply or shallowly, all reflect the state of the respiratory system. Following these behaviors through the day could support self-monitoring and care — but chest belts and masks are obtrusive, and earbuds become uncomfortable and pick up the sounds around us.

RespiraFrame is an eyeglass frame that senses respiration through two complementary channels:

  • Bone-conduction acoustics. A bone-conduction microphone embedded beside the nose captures breathing vibrations as they travel through the face. The sensor sits inside the frame, which shields it from airborne noise; its placement was chosen from four candidates as the one closest to the nasal airway.
  • Electrical impedance tomography (EIT). Eight soft dry electrodes sit where glasses already touch the skin — the nose pads and temples. A small, safe current injected between neighboring electrodes yields 64 measurements per cycle that follow how facial tissue changes as we breathe.

Every four seconds, both signals are encoded by a two-branch network whose features are fused with multi-head attention, recognizing five abnormal respiratory events (coughing, sneezing, sniffing, nose blowing, throat clearing) and distinguishing nasal vs. oral, deep vs. shallow breathing, and breathing vs. breath holding. It runs in real time.

Results

  • Lab study, 19 participants: 92.5% accuracy for abnormal events; 84.8–88.0% for the three breathing tasks.
  • Under background noise up to 80 dB, accuracy changed by less than 3 points; only 2.7% of a bystander’s coughs or sneezes were attributed to the wearer.
  • In the wild — 10 participants, 30 hours in an office, a cafeteria, and outdoors: 87.6% for abnormal events, with no systematic decline over a four-week follow-up.
  • Participants rated the glasses comfortable and socially acceptable — a more suitable everyday form factor than masks or earbuds.

Video

Figures

Sense 1: a bone-conduction microphone embedded beside the nose pad.
Figure 1. Sense 1: a bone-conduction microphone embedded beside the nose pad.
Under 40–80 dB airborne noise, the bone-conduction channel stays clean while a conventional microphone does not.
Figure 2. Under 40–80 dB airborne noise, the bone-conduction channel stays clean while a conventional microphone does not.
Sense 2: electrical impedance tomography with eight soft dry electrodes.
Figure 3. Sense 2: electrical impedance tomography with eight soft dry electrodes.
EIT channels during deep breathing (real data).
Figure 4. EIT channels during deep breathing (real data).
Five abnormal respiratory events and three breathing-pattern tasks.
Figure 5. Five abnormal respiratory events and three breathing-pattern tasks.
Lab study: 19 participants, 10,480 samples.
Figure 6. Lab study: 19 participants, 10,480 samples.
Only 2.7% of a nearby person's events were attributed to the wearer.
Figure 7. Only 2.7% of a nearby person's events were attributed to the wearer.
In the wild: 10 participants, 30 hours across office, cafeteria, and outdoors.
Figure 8. In the wild: 10 participants, 30 hours across office, cafeteria, and outdoors.
Placement study: of four candidate positions, the sensor beside the nose pad gave the highest recognition accuracy, closest to the nasal airway.
Figure 9. Placement study: of four candidate positions, the sensor beside the nose pad gave the highest recognition accuracy, closest to the nasal airway.