The Swaasa artificial intelligence platform achieved 90 percent accuracy in distinguishing healthy individuals from patients with respiratory conditions, according to prospective clinical validation results published in Frontiers in Medicine — a result that could reshape how lung disease is screened outside traditional clinics.
The findings, reported this week by Bioengineer.org, come from one of the larger prospective evaluations of cough-based AI screening to date, and they arrive as healthcare systems worldwide search for cheaper ways to detect chronic respiratory disease earlier. It is the kind of incremental but rigorously tested progress in latest AI developments that often says more about where medical AI is actually heading than any product launch.
How Swaasa Works
Swaasa is an investigational platform that functions as a Software as a Medical Device, applying artificial intelligence algorithms to cough recordings captured on a mobile device. Developed to generate automated clinical indications from acoustic signals, the platform evaluates cough recordings to determine overall risk as normal or abnormal, classify spirometric physiological patterns as obstructive, restrictive, or normal, and estimate pattern severity as mild, moderate, or severe.
Beyond physiology, the system generates diagnostic indications regarding the presence of specific respiratory conditions, including asthma, chronic obstructive pulmonary disease (COPD), interstitial lung disease, and bronchiectasis. The promise is straightforward: a five-second cough into a phone microphone could flag signs of serious lung disease before a patient ever reaches a spirometry lab.
Inside the Trial
The prospective clinical trial was registered under CTRI/2021/04/032742 on April 12, 2021, enrolling adult participants aged 18 years and older who presented to hospital departments with respiratory symptoms. First enrollment commenced on April 19, 2021, and the trial concluded on September 14, 2023.
The final cohort included 2,179 patients with respiratory diseases alongside 827 normal individuals, with male participants comprising 62 percent of the evaluated study population. During the protocol, participant cough sounds were recorded on a mobile device running the Swaasa AI platform, and spirometry was performed for comparison.
The trial established spirometry as the reference standard for physiological pattern and pattern severity, while physician-established clinical diagnoses served as the reference standard for disease categories. Polychotomous variables were dichotomized into normal versus abnormal classifications, and multi-class outputs were established for disease condition, pattern severity, and individual disease categories.
The Performance Numbers
In binary classifications, the platform demonstrated 90 percent accuracy in differentiating normal individuals from those with abnormal respiratory conditions, matching the 90 percent accuracy recorded for identifying normal versus abnormal spirometry. Descriptive subgroup analyses showed screening performance was broadly comparable across evaluated patient age and gender groups.
The multi-class results were respectable, if less striking. The platform achieved accuracy ranges between 79 percent and 84 percent for classifying spirometric patterns, and between 76 percent and 84.5 percent for classifying pattern severity. When differentiating specific disease types among asthma, COPD, interstitial lung disease, and bronchiectasis, accuracy ranged from 83 percent to 93 percent.
Who Built It
The study authors are affiliated with Christian Medical College Vellore, Salcit Technologies Private Limited — the developer of Swaasa — and The Tamil Nadu Dr MGR Medical University. The authors concluded that Swaasa demonstrates promising screening performance across demographic subgroups in the studied population.
The collaboration structure is notable: an academic medical center providing clinical rigor, a medical university, and a startup supplying the acoustic AI. That combination — clinicians validating a startup's algorithm against established diagnostics — is increasingly the template for credible medical AI research, in contrast to the vendor-run benchmarks that have drawn justified skepticism.
Why Cough Acoustics Matter
Spirometry, the current gold standard for lung function, requires equipment, trained technicians, and a cooperative patient performing forced breathing maneuvers. Cough recording requires a phone. If AI screening can reliably triage who needs confirmatory testing, the economics of respiratory screening change fundamentally, particularly for conditions like COPD that are routinely diagnosed late.
The caveats are real. This was hospital-based enrollment of symptomatic adults, not community screening of the general population, and accuracy in a controlled prospective protocol may differ from messy real-world deployment. The platform remains investigational, and 90 percent accuracy in distinguishing normal from abnormal does not mean it can reliably tell four different diseases apart in every case — the multi-class ranges make that clear.
The regulatory path is the next bottleneck. Because Swaasa is positioned as a Software as a Medical Device, its diagnostic indications will need to clear the oversight processes that apply to clinical decision tools before it can move from investigational use into routine care. Prospective validation published in a peer-reviewed journal is a prerequisite for that conversation, not a substitute for it.
There is also a practical asymmetry working in the platform's favor. Spirometry requires equipment, calibration, and a trained technician in the room; a cough recording requires a device most patients already own and a maneuver every patient already knows. In screening, where the goal is to decide who needs confirmatory testing rather than to deliver a final diagnosis, a slightly less precise test that reaches ten times more people can save more lives than a precise test that sits in a hospital basement. The trial's finding that performance held steady across age and gender groups is exactly the kind of consistency a screening tool needs to scale.
Still, the trajectory matters. A phone-based screen validated against spirometry in more than 3,000 participants is no longer a parlor trick. The next test is deployment: whether cough-based AI screening can hold its accuracy in clinics, pharmacies, and homes, and whether regulators and physicians embrace a diagnostic signal that comes from a sound most people barely think about.
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
