Mode
Listen & Record
Monitor and capture lung sounds only
HEAR-COPD
AI analysis for COPD risk scoring
Input Device
Output Device
Record
Monitor: live listen through headphones.
Record: capture lung sounds.
On iOS, connect AirPods before tapping Monitor.
Analyzing…
0%
About OpenStethoscope
▼
Hardware — OpenSteth 1.0
A DJI Mic Mini inserted into a standard Littmann stethoscope cup. ~$50 + ~$50 = <$100, 30 min to assemble. Works with any recording app — fully independent of the AI software.
Software — HEAR-COPD
A first proof-of-concept using Google's HeAR acoustic foundation model to classify lung sounds for COPD risk scoring. Validated on 280 patients across 4 countries (AUROC 0.939). The pipeline is open and extensible — any researcher can retrain it for other respiratory conditions.
Recorded Audio
HEAR-COPD Risk Score
COPD Risk
—
—
Scientific explanation
▼
- Algorithm
- HEAR-COPD — HeAR acoustic foundation model (Google, frozen, 512-dim embeddings) + logistic regression probe.
- What this measures
- Probability that lung sounds are consistent with COPD vs. healthy controls.
- Training data
- ICBHI 2017 (126 patients, Portugal/Greece) · Fraiwan/KAUH (112 patients, Jordan) · RDTR (42 patients, Turkey). 280 patients total, 4 countries. Patient-level stratified cross-validation.
- Performance
- COPD vs Healthy: OOF AUROC 0.939 [0.900–0.970] · COPD vs All patients: 0.890 [0.848–0.930]
- Signal preprocessing
- 5th-order Butterworth low-pass at 2000 Hz applied before embedding, aligning the broadband consumer microphone (DJI Mic Mini, 20–20,000 Hz) with stethoscope-limited training data (~500 Hz hardware cutoff). Preserves all clinically relevant sounds: normal breath (100–500 Hz), wheeze (100–2400 Hz), crackle (200–2000 Hz).
Additional HEAR-COPD Scores
Research use only — not validated for clinical decisions
COPD (vs all patients)
—
Pneumonia
—
Asthma
—
Heart Failure
—