Patient-level validation of a foundation-model pipeline for pediatric lung sounds: physician-annotated adventitious events are recognized, disease groups are not reliably predicted

Abstract Background Lung-sound classifiers are usually evaluated on recording- or event-level splits, although each child contributes many recordings. We evaluated a foundation-model pipeline (PulmoVec) with the patient as the unit of partitioning and asked whether disease group can be predicted beyond age and sex. Methods We analyzed 19693 physician-annotated respiratory events from 736 children in the public SPRSound database. A frozen Health Acoustic Representations encoder with low-rank adapters was trained for screening (normal versus adventitious), sound pattern (normal, crackles, wheeze/rhonchi) and disease group (pneumonia, bronchial disease, normal/other); event probabilities were stacked with age, sex and auscultation site. The primary analysis was nested patient-grouped cross-validation in which every child was tested once, with a locked hold-out as secondary; comparators were the majority class, annotated event duration and demographics, and intervals came from bootstrapping patients. Results For physician-segmented events the area under the receiver operating characteristic curve (AUC) was 0.95 (95% CI 0.94 to 0.96) for screening and 0.95 (95% CI 0.94 to 0.96) for sound pattern, against 0.78 and 0.75 for event duration alone; the positive predictive value for adventitious events was 0.72 (95% CI 0.66 to 0.77). Disease-group prediction reached an AUC of 0.58 (95% CI 0.54 to 0.62); its accuracy (0.44) was below the majority-class rate (0.62) at event level and at patient level (0.46 versus 0.60), and age and sex carried 55.5% of the second-stage attribution for this outcome. Conclusions Recognition of physician-segmented adventitious events holds under patient-level validation and its discrimination is not explained by event duration or demographics, although in exploratory analyses its decisions favored short events and its positive predictive value would limit unaided use. Prediction of disease group, a clinical diagnosis, from lung-sound events alone was not demonstrated. Lung-sound studies should partition data by patient, as prior evidence indicates, and report non-acoustic comparators.