A digital stethoscope programme in rural Bangladesh solved one practical problem and exposed another. After three days of training, nine non-physician health workers obtained usable recordings from most young children. But the AI that classified the sounds as normal or abnormal agreed with a paediatrician panel only moderately well.[2][1]
The evaluation enrolled 990 children aged 2–59 months with suspected respiratory illness. The listening panel judged 867 children—87.6%—to have at least three interpretable chest positions. Among quality recordings with timing data, 89.8% were completed within five minutes. Those results show that the collection workflow can function in first-level clinics.[2]
The classifier result
| Measure | Value | Context |
|---|---|---|
| Analysed children | 867 of 990 | Children with at least three chest positions judged interpretable by the paediatrician panel.Source 1Source 2 |
| Sensitivity | 61.8% | Share of panel-defined abnormal cases that the AI classified as abnormal among all analysed children.Source 1 |
| Specificity | 60.7% | Share of panel-defined normal cases that the AI classified as normal among all analysed children.Source 1 |
| IMCI pneumonia subgroup | 63.5% sensitivity; 66.8% specificity | Performance among children classified as pneumonia cases under Integrated Management of Childhood Illness criteria.Source 1 |
A recording can be good even when the interpretation is wrong
The feasibility study asks whether a worker can position the device, keep the child sufficiently still and capture interpretable audio. The performance study asks whether an algorithm can label those sounds in agreement with specialists. Combining the two outcomes into one “AI stethoscope accuracy” figure would hide the central result: collection was strong; classification remained limited.[2][1]
Where the system succeeded and where it struggled
| Stage | Reported result | What still needs improvement |
|---|---|---|
| Capture | 87.6% of children had quality recordings. | Noise, movement and difficult chest positions still made some recordings uninterpretable.Source 2 |
| Speed | 89.8% of timed quality recordings were completed within five minutes. | Clinic workflow must also include consent, cleaning, upload and interpretation time.Source 2 |
| Classification | Sensitivity 61.8%; specificity 60.7%. | The algorithm needs better recognition of abnormal sounds and an uninterpretable category.Source 1 |
| Clinical decision | Not directly tested as an antibiotic or referral intervention. | A prospective study must show safer decisions and patient outcomes, not only agreement with listeners.Source 1 |
Why the missing “uninterpretable” class matters
The paediatrician panel could label a chest position uninterpretable. The AI system classified recordings as normal or abnormal without the same reject option. Forcing uncertain audio into a clinical label can create false reassurance or false alarm. A safe model should know when noise, crying or poor contact makes the input unsuitable for interpretation.[1]
Moderate sensitivity and specificity create two queues of error
With 61.8% sensitivity, a substantial share of panel-defined abnormal sounds would be missed. With 60.7% specificity, many children with panel-defined normal sounds would be flagged. Whether those errors are tolerable depends on how the tool is used. A screening aid can accept more false positives than an automated decision to prescribe or withhold treatment.
The clinical endpoint must come next
A stronger trial would compare care with and without the tool: antibiotic prescribing, referral, missed severe illness, follow-up recovery and clinician override. It should also test the algorithm prospectively on new clinics and seasons and report performance by age, noise level and recording quality.
This is still useful research. It separates the human–device interface from the model and shows that infrastructure can be ready before the classifier is. In clinical AI, that honest split is more valuable than a single optimistic accuracy number.[1][2]
Sources & further reading
- Identification of adventitious lung sounds in children with respiratory illnesses in Bangladesh using artificial intelligence and digital auscultation Scientific Reports · 16 Oct 2026 · Accessed 16 Oct 2026
- Digital stethoscope use by non-physician primary care health workers on children under five years of age in rural Bangladesh: a feasibility study Scientific Reports · 12 May 2026 · Accessed 16 Oct 2026