In submission

RespiraFrame

Respiratory sensing in an eyeglass frame, by bone conduction and impedance

Glasses drawn as the idea; the prototype appears in the photos

Why breathing

Breathing says a lot.

Respiratory events

Breathing patterns

+ breath holding

Airflow drawn

Head: Lee Perry-Smith scan · CC BY 3.0

Systems today

Unobtrusive, or robust?

Method	Form	Unobtrusive	Bystander	Noise	Wind	Motion
Anaya et al.	Chest belt	−	N/A	N/A	N/A	−
Ida et al.	Finger PPG	−	N/A	N/A	N/A	−
Listen2Cough · Whosecough	Any microphone	✓	−	−	−	−
CoughBuddy	Earphone	✓	✓	−	−	−
BreathPro	Earphone	✓	−	✓	−	N/A
EarSAVAS	Earphone	✓	−	✓	−	−
EchoBreath	Glasses	✓	✓	✓	−	−
?	Glasses

✓ good · − poor, or not reported · N/A not applicable · the paper's Table 1

Why it matters

Noise floods an ordinary microphone.

Measured Conventional microphone · the paper's recordings

Spectrogram of a cough recorded by a conventional microphone, no added noise

A cough · no added noise

Spectrogram of a cough recorded by a conventional microphone under 80 dB noise

A cough · 80 dB noise

80 dB ≈ a busy street

Our goal

RespiraFrame fills the row.

Method	Form	Unobtrusive	Bystander	Noise	Wind	Motion
Anaya et al.	Chest belt	−	N/A	N/A	N/A	−
Ida et al.	Finger PPG	−	N/A	N/A	N/A	−
Listen2Cough · Whosecough	Any microphone	✓	−	−	−	−
CoughBuddy	Earphone	✓	✓	−	−	−
BreathPro	Earphone	✓	−	✓	−	N/A
EarSAVAS	Earphone	✓	−	✓	−	−
EchoBreath	Glasses	✓	✓	✓	−	−
RespiraFrame	Glasses	✓	✓	✓	✓	✓

✓ good · − poor, or not reported · N/A not applicable · the paper's Table 1

Our idea

Two channels, different weaknesses.

Bone conduction · a sensor in the frame picks up vibrations carried through the face and the frame

Impedance · electrodes feel the face change as you breathe

Room noise · the frame's sensor picks up less of it; impedance is not acoustic

VPUbeside the nose pad

Electrodesnose pads, temples

The glasses prototype — Real · prototype

Drawn, slowed down

Head: Lee Perry-Smith scan · CC BY 3.0

The sensors

The sensors sit in the frame.

Glasses drawn as the idea; the photos show the prototype

ListenVPU in a notch by the nose pad · up to 8 kHz

Feel8 dry electrodes · nose pads, temple tips · 50 kHz · 400 µA

The sensing board — Real · sensing board

Acquire · stream · a separate board, wireless

Bone conduction

Bone conduction hears less of the room.

Embedded, the frame is a barrier.

Cough
Cough + 80 dB
80 dB noise
60 dB
40 dB

—

Measured The paper's Fig. 2 · no common colour scale given: compare each with itself

Measured VPU only · one colour scale · the paper's Fig. 6

Speech noise, 80 dBThroat clearing Touching the skin VPU spectrogram, direct skin contact, speech noise at 80 dB

VPU spectrogram, direct skin contact, throat clearing

Hollow frame VPU spectrogram, hollow frame, speech noise at 80 dB

VPU spectrogram, hollow frame, throat clearing

Embedded

Embedded: the paper's best balance · some leakage remains

Rings and ripples drawn

Head: Lee Perry-Smith scan · CC BY 3.0

Impedance

Tiny currents read the face.

Breathing changes the channels.

A neighbouring pair injects 50 kHz · 400 µA; neighbouring pairs measure; the pair rotates.

Injecting: —

Electrode order assumed · paths drawn

Measured One 4 s window per class · one participant
Smoothed and differenced as in the paper · paths drawn

Head: Lee Perry-Smith scan · CC BY 3.0

Fusion

Each task's model listens to both.

Four tasks, four such models

Fused vs the weaker channel: p < 0.001 · the paper's bars carry ±7–10-point error bars, not drawn here

Results · lab

Four tasks, one frame.

92.5% Abnormal events

84.9% Nasal / oral

88.0% Deep / shallow

87.8% Breath holding

19 participants (breath holding: 15) · user-dependent

Real · the live demo of each task

80.6% unseen users 75.4% unseen users 81.6% unseen users 87.8% unseen users

Leave-one-participant-out

Robustness

Hard conditions, one at a time.

Noise

< 2–3 points lost

40 → 80 dB, every task. No significant differences.

Recorded noise added to the VPU test data · EIT unchanged

Wind

Events 85 · 80 · 69%

Fan at full speed · hair dryer low · hair dryer full (the paper's text). Other tasks: no significant change.

Full hair dryer: a stress test

Wind drawn

Real · a hair dryer in the live demo

Bystander

2.7% false alarms

A neighbour has the five events; the wearer talks, drinks, tidies, uses a phone or breathes.

19 participants in pairs

Second head and rings drawn

Real · a bystander in the live demo

Motion

Route 93.2 · 93.9 · 94.1%

Standing · walking · running in place: no significant difference. Events: 87.1% → 70.3% running.

Motion study · fused model

Noise drawn · in the study it came from a laptop ≈ 30 cm away

Head drawn bobbing, as when running in place

Head: Lee Perry-Smith scan · CC BY 3.0

Semi-wild

Office, cafeteria, outdoors.

Task	Office · lab	Cafeteria	Outdoor	Average
Abnormal events	81.6	86.2	95.2	87.6%
Nasal / oral	77.1	79.4	88.0	81.5%
Deep / shallow	76.0	89.0	88.6	84.6%

10 participants · semi-wild · user-dependent · window-level · no breath holding · tested on a different day from training

Longitudinal

No systematic decline over four weeks.

10 participants · days 7, 14, 21, 28 · the paper's Table 2

Questionnaire

Participants rate the form factor highly.

6.03comfort · wearability

6.08social acceptability

6.38usefulness · continued use

Out of 7 · 26 participants · the current prototype

6.62 · glasses suit this better than masks or ear-worn devices

5.65 · the lowest: wearing the current prototype in public

RespiraFrame

01Two channels · bone conduction and impedance, one frame
02Four tasks · 84.9 to 92.5% in the lab, on each user's own data
03Hard conditions · noise, wind, bystanders, motion
04Four weeks · no systematic decline

Thank you.

In submission

RespiraFrame

Respiratory sensing in an eyeglass frame, by bone conduction and impedance

Thank you.

Head model in the talk: Lee Perry-Smith scan · CC BY 3.0