Peer-reviewed literature confirms that cough acoustics carry clinically significant information about respiratory disease. Every attempt to deploy this at scale has failed, not because the science is wrong, but because models trained on clean, controlled audio degrade in real-world conditions. Our approach: train where we deploy, from real OPDs, under real acoustic conditions.
End-to-end pipeline: ingestion, quality gating and feature extraction
Recordings — PoC dataset
Target — full study
Every published acoustic respiratory AI study was built on controlled-environment recordings with limited geographic and demographic diversity. None have been validated under real Indian OPD conditions, ambient noise, diverse patient populations, varying disease severity, that define actual clinical deployment environments.
None cover the full spectrum of respiratory disease with molecularly and objectively confirmed diagnosis labels. None have been trained on the real-world compressed audio that a low-bandwidth deployment channel actually produces. This dataset does not exist. We are building it.
We are training on real-world compressed audio from real OPDs, the same acoustic domain as deployment. Previous models trained on clean audio collapsed in production; this is the gap our data collection is specifically designed to close.
Two-phase protocol separates recording from labelling. Ground truth uses GeneXpert for TB, spirometry for COPD and asthma, radiology for ILD. No provisional impressions.
TB, COPD, Asthma, Bronchitis, ILD, Pleural Disease, classified simultaneously. Real OPD clinical reality, not binary TB-vs-healthy models.
Repeat recordings from the same patients across their treatment journey. No published acoustic respiratory AI dataset has this. Enables treatment monitoring, a new research frontier.
Multi-site recruitment across urban, semi-urban and rural India. Systematic demographic characterisation including occupation, city of residence and environmental exposure.
Once validated, this model reaches every Indian with a respiratory concern through a single voice message, no app download, no clinic visit, no new infrastructure.
Recordings linked to clinically diagnosed (non-molecularly confirmed) outcomes are retained in a separate sub-cohort and excluded from primary model training. These are used only for post-hoc validation, ensuring the primary model trains exclusively on confirmed-diagnosis ground truth. No prior acoustic AI study has made this distinction explicitly.
A prospective multi-centre observational study collecting cough recordings linked to confirmed clinical diagnoses across participating chest and pulmonary medicine outpatient departments in India. No intervention. No drugs. No disruption to clinical workflow.
A validated, zero-infrastructure acoustic respiratory AI screening model, deployable immediately to 500 million Indians without app download, without infrastructure, without clinical staff. No such model exists today.
A structured, multi-geography acoustic repository supporting peer-reviewed clinical validation and post-hoc academic analyses across participating institutional sub-cohorts.
A peer-reviewed methodology for low-bandwidth-domain acoustic data collection establishing the standard for future studies globally, including the dual-track longitudinal protocol.
Longitudinal acoustic trajectory data across treatment journeys, enabling a treatment monitoring model for TB and COPD that could supplement expensive spirometry and GeneXpert follow-up in resource-limited settings.
We are actively onboarding clinical research sites across India. If your institution is interested in participating as a co-investigator site, we would welcome a conversation.
