An AI stethoscope missed too many abnormal lung sounds for stand-alone use
In 867 Bangladeshi children, the algorithm reached 61.8% sensitivity and 60.7% specificity against a paediatrician listening panel. Clinic noise and crying exposed the gap between promise and dependable use.
The short version
3 key points
- 1.Non-physician health workers recorded four chest positions from 990 children; 867 with at least three panel-interpretable positions entered the main analysis.
- 2.Against a paediatrician listening panel, child-level sensitivity was 61.8% and specificity 60.7%.
- 3.The comparator was expert interpretation of recorded sounds, not a definitive disease test, and the algorithm could not label a recording uninterpretable.
The field test was more demanding than a quiet laboratory
An AI-enabled stethoscope was not reliable enough to decide on its own whether a young child had abnormal lung sounds. In a peer-reviewed study from rural Bangladesh, the algorithm detected 61.8% of children whom a paediatrician listening panel classified as having crackles or wheeze. It correctly identified 60.7% of panel-negative children. In this demanding field dataset, those figures describe moderate performance, not a safe replacement for clinical assessment.
Non-physician health workers recorded four chest positions from 990 children under five with cough or difficulty breathing at community clinics. The main analysis included 867 children with at least three recordings the panel considered interpretable. That is a larger, less controlled paediatric field test than many earlier studies, and it exposed noises a real service must handle: crying, vocalising, movement and a crowded clinic.[1]
The model often heard abnormality where experts did not
The panel classified adventitious sounds in 36.7% of children meeting the study's pneumonia definition and 27.9% of those without it. The algorithm marked 44.3% and 47.7% respectively. That high AI-positive rate outside the pneumonia group suggests artefact was often mistaken for wheeze or crackles. Among pneumonia cases, sensitivity rose only to 63.5% and specificity to 66.8%.
There was also a crucial mismatch in what each side could say. Paediatricians could reject an uninterpretable recording; the algorithm could only call it normal or abnormal. Performance improved when the panel was confident, but that does not remove the real-world problem: the noisy recording is exactly the one an automated tool must recognise rather than force into a diagnosis-shaped category.
The result also helps explain why impressive figures from smaller hospital studies may not travel cleanly into primary care. A specialist can often capture cleaner sounds from a child who is already admitted, while a community worker has less control over noise, cooperation and timing. External validity here comes from the messiness, even though that same messiness depressed performance.[1]
The distinction that matters
What the study actually compared
The test measured agreement about recorded sounds, not whether a child truly had pneumonia.
Record
Capture four chest positions
Noise, crying and movement remain in the test
Compare
AI and paediatricians label the sounds
Sensitivity and specificity use the panel as reference
Decide
Combine findings with clinical assessment
This step was not evaluated
A useful assistant must know when not to answer
For a community health worker, automated auscultation could still become useful alongside breathing rate, chest indrawing, oxygen saturation and clinical judgement. But a moderate match with expert listeners cannot establish pneumonia, decide on antibiotics or safely reassure a family. The authors call for training that can recognise uninterpretable and low-confidence recordings.
The evidence check is stricter than the headline numbers. Expert listening is imperfect and is not a gold-standard diagnosis; the study evaluated sound classification, not treatment decisions, hospitalisation or child outcomes. The main analysis also excluded 123 enrolled children without enough interpretable positions, potentially removing hard cases. Confidence would rise if a prospective trial showed that a revised system rejects poor recordings, improves decisions beyond standard IMCI assessment and does not widen errors across clinics or ages. Until then, this is a warning against turning a digital stethoscope into an automated verdict.[1]
Evidence check
The comparator was expert interpretation of stored sounds; 123 enrolled children were excluded from the main analysis for insufficient interpretable recordings.
Reporting basis: Source analysis. Original source date: 16 October 2026.
Explore the full source and evidence record
What this means for people
- Families and frontline workers need a system that flags poor recordings and supports, rather than replaces, escalation when a child is struggling to breathe.
Global context
Pneumonia remains a major cause of death in children under five, particularly where imaging and specialist review are limited. Any useful tool must survive first-contact clinical noise.
What the evidence does not yet show
- The comparator was expert interpretation of stored sounds; 123 enrolled children were excluded from the main analysis for insufficient interpretable recordings.
What to watch next
- Prospective decision-impact trials, independent validation, explicit rejection of poor recordings and evidence that the tool improves care without increasing missed illness or unnecessary treatment.
Living evidence record
Impact record IAI-0FHLS15
Evidence stage
Studied
Confidence
Supported
Reporting basis
Source analysis
Independent or research support
Present
Record status
Monitoring
Last checked
16 October 2026
Source trail
1 direct source across 1 source type.
People impact
Documented in this record.
Uncertainty
Limits and next checks are explicit.
Stages describe the evidence available—not whether a technology is good or bad. See the public method.
Single-source reporting disclosure
This record analyses one direct source. It can establish what Scientific Reports published or reported, but it is not independent corroboration of every performance claim or predicted outcome. The confidence label will change only when broader evidence is added.
Evidence trail
Sources used for this report
Links checked 16 October 2026
Research question, topics and regions
AI classification of paediatric lung sounds in community clinics
Topics: Child health · Pneumonia · Digital stethoscopes · Community health workers · Diagnostic AI
Regions: Bangladesh · South Asia · Global
This report is labelled source analysis. We summarise and analyse source material in our own words; company statements remain attributed claims until independently supported. Translated summaries preserve the meaning of the original source and link back to it. Read our editorial standards.
Watch the background, then check the evidence
The Royal Society
AI-generated illustration · not a video frame
YouTube loads only when you choose. Press play in the player to start. If playback is unavailable, use “Watch on YouTube” below.
Healthcare, culture & human impact · Background explainer
Published 2026-07-01 · 44 min 28 sec
Is this the AI future we want? | The Royal Society
A Royal Society discussion with Madhumita Murgia and biomedical engineer Alison Noble on AI’s potential in healthcare, science and culture, recorded at the 30 June 2026 event.
Evidence check: A discussion of opportunities and risks, including future possibilities. It is not evidence that every application described is available or clinically proven.
Source checked 14 October 2026. Check available captions on the publisher’s video page.
The Impact Brief · Free
Keep following the stories that matter to you.
Get our weekday briefing on what changed and why it matters. Choose your topics below.
Continue the story
There’s more to this story
Health & Life Sciences
How accurate must diagnostic AI be before people accept it?
A Korean survey of 1,328 people found mean minimum acceptable sensitivity and specificity of 90% for AI disease detection. That is a demanding stated threshold—not a clinical performance standard, a treatment decision or proof of what people would accept when facing a real diagnosis.
9 min · 1 source
Health & Life Sciences
Can a high-AUC diabetes model still be unsafe?
New analysis today of a peer-reviewed 2 October audit of 12 machine-learning approaches on two public diabetes datasets. Similarly ranked models can differ materially in calibration, uncertainty and safe deferral, but this is a benchmark study—not a clinical trial, diagnostic approval or evidence of improved patient outcomes.
7 min · 3 sources
Health & Life Sciences
Can an AI medical council give a safer second opinion?
A 300-case preprint found structured multi-stage reports scored better on management planning, but an automated judge, unusual published cases and unequal compute prevent any claim of clinical benefit.
3 min · 1 source
Reader commentary
What does this mean where you are?
Tell us what you have seen at work, in your community or in everyday life. A question is welcome too. No account is needed; we review notes before publishing them.
Explore commentary across the portal →Published reader notes
0No published reader notes yet. You can start the evidence-led discussion above.
Prefer a private correction or response? Contact the newsroom.