Breath Print

"BreathPrint turns a 30s phone recording of a cough, breath & vowel into instant, explainable respiratory risk triage—running 100% offline to empower frontline health workers."

Comment

Inspiration

Respiratory disease remains one of the largest and most solvable gaps in global health—not because we lack effective treatments, but because we lack accessible, point-of-care screening.

According to the latest World Health Organization (WHO) data, 8.3 million new tuberculosis cases were notified in 2024—the highest number ever recorded—with over 1.23 million deaths. Crucially, roughly 1 in 5 active TB cases (~22%) goes undiagnosed, continuing unchecked transmission in rural and low-infrastructure communities.

Confirmatory diagnostic tools like GeneXpert cartridges, chest X-rays, and sputum microscopy are lab-bound, expensive ($10–$30/test), electricity-dependent, and require days of turnaround. Community Health Workers (CHWs), who serve as the sole point of care for millions, currently rely on subjective symptom checklists. This bottleneck causes early cases to be missed while healthy patients clog referral labs.

I built BreathPrint to answer a fundamental question: What if the smartphone already in a health worker's pocket could serve as an instant, explainable acoustic triage filter—completely offline?

What it does

BreathPrint is an on-device AI acoustic screening and triage app designed for community health workers and rural clinics:

  • 30-Second Multi-Modal Recording: Captures three complementary acoustic biomarkers on any smartphone: (1) a voluntary cough, (2) deep breathing, and (3) a sustained vowel phonation ("say ahhh").
  • On-Device Risk Triage: An embedded TinyML engine instantly classifies respiratory risk into Low / Medium / High Risk tiers for Tuberculosis and Pneumonia (extendable to COPD/Asthma).
  • 100% Offline Edge Intelligence: Operates with zero internet connectivity and zero cloud dependency, preserving patient privacy and functioning in the most remote off-grid clinics.
  • Plain-Language Explainability (XAI): Rather than outputting a black-box number, Grad-CAM spectrogram attention maps are translated into plain-language clinical insights (e.g., "wet cough quality + irregular breath rhythm detected").
  • Closed-Loop Referral Action: Automatically generates a shareable, printable QR referral slip and dispatches an automated SMS alert (via sandbox gateway) to fast-track high-risk patients to confirmatory testing.
  • Longitudinal Recovery Tracking: Allows diagnosed patients to re-record weekly to give health workers an objective signal of treatment adherence and recovery.

Note: BreathPrint is explicitly designed as a frontline triage decision-support tool, not a diagnostic replacement.

How I built it

  1. Acoustic Dataset Grounding: Leveraged curated, ethically collected open biomedical corpora for pretraining and validation:
    • Coswara (Indian Institute of Science - IISc): 2,500+ multi-modal audio recordings with curated clinical metadata.
    • COUGHVID (EPFL Switzerland): 25,000+ cough audio samples with expert pulmonologist annotations.
    • Virufy Open Corpus: Clinical multi-country respiratory sound dataset.
  2. Signal Processing Pipeline: Using Librosa and SciPy, audio is denoised via spectral subtraction, segmented using Voice Activity Detection (VAD), and transformed into 128-band log Mel-spectrograms, 40 MFCCs (vocal tract dynamics), spectral centroids, and zero-crossing rate features.
  3. Multi-Branch CNN & Attention Fusion: Designed a 3-branch lightweight convolutional network (dedicated encoders for cough, breathing, and phonation) coupled with a Cross-Modal Self-Attention Fusion Layer that dynamically weighs acoustic features based on signal saliency.
  4. TinyML Quantization: Model was exported and quantized to INT8 TensorFlow Lite, achieving a footprint under 15 MB and on-device inference latency under 250ms on commodity ARM Cortex mobile processors.
  5. Frontline Mobile UX: Built a low-literacy, multi-lingual Android application interface featuring visual color-coded risk indicators, large touch targets, and local offline encrypted SQLite storage.

Challenges I ran into

  • Acoustic Noise in Real-World Field Settings: Commodity phone microphones in noisy rural environments introduce heavy acoustic artifacts. I implemented adaptive SNR thresholding, spectral gating, and background noise data augmentation during model training.
  • Limitations of Single-Sound Classifiers: Existing research often relies exclusively on cough audio, which lacks specificity. Fusing breathing and vowel phonation solved this by capturing airway resistance and vocal fold harmonics that cough alone misses.
  • Extreme On-Device Resource Constraints: Running a multi-branch deep network on low-cost ($50) Android phones required aggressive post-training INT8 quantization while maintaining sensitivity calibration to minimize false negatives.
  • Frontline AI Trust & Explainability: Health workers often distrust black-box scores. Integrating Grad-CAM spectrogram attribution to highlight frequency anomalies in plain language overcame this barrier.

Accomplishments that I'm proud of

  • Designed and implemented a complete, end-to-end multi-modal acoustic triage pipeline that runs 100% offline on low-end hardware.
  • Achieved a sub-15MB model footprint with near-zero marginal cost per screening, enabling virtually limitless scalability.
  • Bridged the critical gap between raw AI scoring and actionable clinical workflows by wiring predictions directly into printable QR referral slips and automated clinic alerts.
  • Built this entire initiative from concept, data architecture, model engineering, UX design, and clinical framing as a solo participant for Hack2Heal 2.0.

What I learned

  • The deep biophysical richness of human respiratory acoustics—how pulmonary infections distinctly alter vocal tract formants, turbulence, and time-frequency spectrogram energy.
  • The practical engineering rigor required to quantize and deploy TinyML models on resource-constrained edge hardware.
  • The reality of frontline global health: an AI algorithm alone doesn't save lives—it must fit seamlessly into a health worker's field workflow to drive tangible clinical referrals.

What's next for BreathPrint

  • Clinical Prospective Validation: Partner with public health NGOs and teaching hospitals to benchmark BreathPrint against confirmatory GeneXpert and Chest X-Ray ground truths.
  • Privacy-Preserving Federated Learning: Pilot federated model updates across 2–3 partner clinics to continuously adapt to regional accents and background acoustics without centralizing raw patient audio.
  • Multi-Disease Expansion: Extend the acoustic classification architecture to COPD monitoring, asthma exacerbations, and pediatric pneumonia triage.
  • Regulatory Triage Pathway: Pursue digital health screening and triage software pre-qualification with WHO and regional health authorities.

Built With

Share this project:

Updates

Submission history