Machine learning has real value in respiratory care when it adds validated insight, but "AI for respiratory medicine" now describes very different capabilities. Vendor websites, conference abstracts, and funding announcements use the term for work ranging from advanced acoustic breath signal processing to rules-based symptom questionnaires with chatbot interfaces. Clinicians can reasonably struggle to tell new methods from existing approaches given new marketing language.
This article aims to make that distinction clearer. We work in acoustic monitoring and therefore have a view on what machine learning applied to breath sounds can add. We also want to describe the wider field fairly, including areas where respiratory AI is moving faster in presentation than in supporting evidence.
Where Evidence Is Strongest
Radiology has the strongest evidence base for respiratory AI. Convolutional neural networks applied to chest X-rays and CT scans have been validated across several independent datasets for nodule detection, pneumonia identification, and measurement of emphysema extent. Some tools have received regulatory clearance in the UK and EU. Their performance is documented, their failure modes are understood, and workflow integration is relatively clear. The model serves as a second reader or triage filter, while the radiologist reviews the case and remains responsible for the final report.
Spirometry interpretation is another area with meaningful evidence. Models analysing flow-volume loops can identify technically inadequate manoeuvres, propose diagnostic categories, and recognise patterns linked with particular obstructive and restrictive disorders. The evidence is reasonably solid, and the clinical role is defined: decision support within an established assessment, rather than autonomous diagnosis.
Promising Areas with Immature Evidence
Our field is acoustic analysis of respiratory sounds, where the fairest description is encouraging but incomplete. The biological rationale is sound. Airway changes during an exacerbation alter breath sound characteristics, and controlled studies show these changes can be captured using consumer-grade microphones. Research groups have reported that acoustic models can classify wheeze severity and identify early exacerbation signals in retrospective datasets.
A large prospective trial has not yet shown that home acoustic monitoring, incorporated into standard COPD care, reduces exacerbations or hospitalisations at a statistically meaningful level in a representative patient population. Our own pilot data is encouraging on a small scale. We state that plainly, because overstating the evidence would not help clinicians or patients.
Continuous wearable measurement of respiratory rate and effort occupies a similar position. The technology exists and works reasonably well in compliant patients, but evidence of meaningful outcome improvement remains at the pilot stage in most respiratory settings. It is promising and merits attention, but it is not definitive.
Where Caution Is Needed
Some respiratory AI applications have a wide enough gap between their marketing and their actual capability to warrant caution.
Conversational tools labelled "virtual respiratory nurses" or "symptom AI triage" are one example. Most rely on a form of large language model or a decision-tree chatbot applied to symptom questions that could also be collected on paper. Calling it "AI" does not make the result more clinically reliable than the questionnaire beneath it. In some cases, the conversational format may create extra risk if patients mistake it for a clinical interaction with greater validity than it has.
Exhaled breath analysis presented as a diagnostic AI tool is another area for caution. Research into volatile organic compounds in breath is legitimate and interesting. However, moving from research-grade mass spectrometry findings to a consumer or clinical device that reliably stratifies COPD severity or predicts exacerbations is a substantial step. Several commercial products have taken that step ahead of their evidence.
MHRA or UKCA Status
In the UK, an AI tool intended to affect clinical decisions for individual patients will likely need MHRA registration as a medical device, or at least a clear regulatory assessment explaining why it does not. This is not administrative detail. It is part of the clinical governance framework protecting patients from devices that produce clinical-sounding outputs without sufficient validation.
Clinicians assessing a respiratory AI tool should ask: what is the product's regulatory status, what clinical claims does it make, and what evidence supports them? A vendor unable to answer clearly, or one that portrays regulatory requirements as a bureaucratic barrier to adoption, is not offering claims that should simply be accepted.
RespiroScreen is not currently a registered medical device. We are a monitoring and alert tool at the pilot stage, with appropriate clinical governance agreements in place with the practices we work with. We are open about the current evidence and are working toward regulatory submission as that evidence develops.
A Framework for Assessment
Four questions can quickly test the substance of a respiratory AI claim. First: what data trained the model, and how closely does it reflect your patient population? Differences between training demographics and the deployment setting can reduce performance without being obvious in headline accuracy figures. Second: what does the model produce, and which clinical decision is it meant to support? An output further removed from direct clinical action is easier to validate but less directly useful. Third: what is its regulatory status, and what validation exists in prospective settings rather than retrospective datasets? Fourth: what happens when the model is wrong? Is there a graceful failure mode, and who retains clinical responsibility for the result?
A credible vendor should be able to answer all of these questions. If it cannot do so directly, it is not ready for clinical deployment, however impressive the demonstration may appear.