2.1 Study design and reporting framework

This work was designed as a comparative prediction-model study for mixed cardiopulmonary sound classification. The target task was dual binary classification from a single mixed recording: one label for the cardiac component and one for the pulmonary component. The analysis followed the logic of a benchmark-style methods comparison, with explicit separation between ranking performance and thresholded decision behavior.

The principal target-domain analysis used nested grouped five-fold cross-validation over 145 paired triplets. Each triplet served as one grouping unit and contained one mixed recording together with its matched isolated heart and lung recordings. The original outer folds were preserved at the triplet level so that matched recordings could not be split across folds. Within each outer training fold, a group-disjoint inner validation split was used for checkpoint selection, calibration, and threshold selection. The outer fold was reserved for final evaluation only.

2.2 Datasets

Three public datasets were used. HLS-CMDS served as the target-domain dataset [1]. After metadata harmonization and manual audit, the processed HLS index contained 536 rows, of which 535 analyzable recordings formed the study corpus and one additional example was retained for audit only; the final comparative analysis used 145 complete mixed-recording triplets. No new human-subject data were collected for the present study.

PhysioNet/CinC 2016 was used as the external cardiac teacher corpus [2]. After preprocessing, 3240 heart-sound recordings were indexed for teacher training. ICBHI 2017 was used as the external pulmonary teacher corpus [3], yielding 920 indexed respiratory recordings after preprocessing. Healthy recordings were treated as the normal class and all diagnosed respiratory conditions were grouped as abnormal for pulmonary teacher pretraining. The purpose of using these corpora was not to claim direct clinical equivalence across datasets, but to inject organ-specific prior knowledge into the target-domain learner.

2.3 Signal processing and input representation

All audio pipelines were standardized to mono 4-kHz processing in the final configuration. Neural models received 64-bin log-Mel spectrograms generated from 15-s windows using a 512-point short-time Fourier transform and a 160-sample hop. This representation was chosen to harmonize the target-domain and teacher-domain pipelines while keeping the computational burden modest relative to the small target corpus. For hand-crafted baselines, mixed-recording descriptors included MFCC summary statistics, delta and delta-delta terms, spectral centroid, bandwidth, roll-off, zero-crossing rate, and energy-related descriptors.

For the feature-based reference path, mixed-recording descriptors were derived from MFCC statistics, delta and delta-delta terms, spectral centroid, bandwidth, roll-off, zero-crossing rate, and energy summaries. These hand-crafted features were retained to verify that the target task contained learnable structure even under simple representations. However, the principal manuscript conclusions were based on the neural models because the complete five-fold campaign was performed for those systems.

2.4 Target tasks and labels

Each mixed recording was associated with two binary labels. The heart label denoted whether the cardiac source was abnormal, and the lung label denoted whether the pulmonary source was abnormal. This dual-task formulation was selected for two reasons. First, it respects the fact that mixed recordings may simultaneously contain a normal signal from one organ and an abnormal signal from the other. Second, it creates a natural test bed for shared-representation learning, because both outputs are derived from the same acoustic input.

The main evaluation set therefore consisted of 145 paired triplets. Each outer test fold contained approximately 29 mixed recordings, while the corresponding outer training portion was further divided into inner training and inner validation subsets. With partitions of that size, noticeable fold-to-fold fluctuation was expected. Accordingly, the manuscript treats the five-fold means, standard deviations, and fold-level supplementary results as descriptive comparative evidence, whereas single-fold examples are used only to illustrate specific behaviors.

2.5 Model families

Four model families were studied. The first was a shared-backbone CNN baseline. It used one acoustic trunk followed by two task-specific output heads, thereby testing whether straightforward multitask learning improved over separate single-task CNNs. The second was the student-only mixed-sound model. This model also operated directly on the mixed recording, but it was kept free of external teacher constraints and source-aware regularization. In effect, it represented the strongest pure target-domain learner. The CNN baselines and heart-sound modeling choices follow common log-Mel/CNN and benchmark practices [17–21].

The third family introduced dual teachers. A heart teacher was pretrained on PhysioNet/CinC 2016 and a lung teacher on ICBHI 2017. Their outputs regularized the student through auxiliary losses. Two source-aware variants additionally used paired isolated heart and lung recordings from HLS-CMDS to constrain the mixed-recording representation. In practice, the most informative comparison was not simply teacher versus no teacher, but heavy versus light auxiliary weighting. The final candidate that emerged from this progression was a lightly regularized source-aware model.

Teacher checkpoints were trained independently before target-domain comparison. On the fold-0 teacher validation sets, the PhysioNet teacher reached AUROC 0.9572, whereas the ICBHI teacher reached AUROC 0.9855. The lung teacher, however, showed pronounced threshold bias at the default operating point. These fold-0 teacher summaries were used as adequacy checks before target-domain comparison rather than as stand-alone claims about transfer performance.

2.6 Calibration, threshold selection, and metrics

Ranking metrics alone were considered insufficient because early validation experiments revealed marked threshold sensitivity. The revised evaluation therefore separated discrimination, calibration, and threshold-dependent behavior more explicitly. Macro AUROC was used as the primary discrimination endpoint, with macro AUPRC as a complementary ranking metric. Because both target labels were highly prevalent in the mixed subset, AUPRC values were interpreted alongside prevalence-defined no-skill AUPRC baselines.

Temperature scaling and threshold selection were performed within each outer training fold only. Specifically, each outer training portion was split into inner training and inner validation subsets. Models were trained on the inner training data, and the inner validation data were used for checkpoint selection, temperature scaling, and separate heart and lung threshold selection. Temperature parameters were estimated by minimizing negative log likelihood on the inner validation predictions. Heart and lung thresholds were then selected on the calibrated inner validation probabilities to maximize balanced accuracy. The selected temperatures and thresholds were applied once to the untouched outer fold for final evaluation. This same calibration and threshold-selection pipeline was applied to all principal neural models. Temperature scaling and calibration summaries followed established neural-network and clinical-prediction calibration practice [22,23].

2.7 Statistical analysis, reproducibility, and final comparison set

The principal comparisons focused on four calibrated final candidates: the student-only model, the shared-dual CNN, the light dual-teacher model, and the light source-aware model. All four models were evaluated under the same nested grouped five-fold procedure. Because only five grouped outer folds were available and fold-wise estimates are not statistically independent replicates, the cross-validation summaries were treated as descriptive comparative evidence rather than as a basis for formal significance testing. We therefore report fold-wise means, standard deviations, and model contrasts, and we judge consistency across endpoints together with absolute performance.

All analyses were implemented in Python using openly available scientific-computing libraries. The archived supporting package includes the finalized parameter settings, fold-definition tables needed to reconstruct the evaluation, per-fold summary metrics, out-of-fold predictions, and the scripts used to generate the reported analyses. The revised reporting package was also checked against TRIPOD+AI and STARD-AI guidance, and a completed STARD-AI checklist is supplied as Supporting Information [24,25].

3.1 Target-dataset profile

The curated HLS-CMDS metadata table confirmed that the supervised mixed-task subset was structurally clean and suitable for paired evaluation. All 145 identifiers corresponded to complete triplets composed of one mixed recording together with its matched isolated heart and isolated lung source recordings. The mixed subset was strongly imbalanced: 91.0% of recordings were heart-abnormal and 80.7% were lung-abnormal. Such imbalance made it unsurprising that some early models achieved high sensitivity while collapsing specificity at the default operating point.

The location distribution was relatively broad rather than dominated by a single site. Right costal margin (RC), right upper sternal border (RUSB), left costal margin (LC), and left mid anterior (LMA) each contributed at least 13 mixed recordings, while the least frequent location, right upper anterior (RUA), still contributed 9. This spread is modest, but it reduces the risk that the target task merely reflects a single-location artifact. Additional dataset-flow and auscultation-location summaries are provided in the supplementary material.

3.2 Teacher models and preliminary sanity checks

The external teacher models were strong enough to justify testing them as auxiliary guides, but they also suggested that transfer into the mixed target domain might be limited. The PhysioNet teacher showed balanced discrimination and acceptable calibration, whereas the ICBHI teacher achieved high ranking performance but a more asymmetric operating profile at the default threshold. These results support using the teachers as candidate sources of guidance, not as evidence that such guidance would necessarily transfer cleanly to the mixed task.

Preliminary fold-0 baselines further indicated that the target problem was learnable. A hand-crafted SVM reference achieved heart-task AUROC 0.8333, and a random-forest lung reference achieved AUROC 1.000 on fold 0. These exploratory models were not treated as formal competitors in the main comparison because a matched five-fold campaign was not completed for all feature-based baselines. They are therefore reported only as contextual reference points showing that the processed inputs retained usable task information.

3.3 Main nested five-fold comparison

The main nested cross-validation comparison is summarized in Table 1, while the overall discrimination-versus-decision pattern is shown in Fig 2. After applying the same inner-validation calibration and threshold-selection procedure to all principal models, the light source-aware model achieved the highest mean discrimination, with macro AUROC 0.7107 ± 0.1659 and macro AUPRC 0.9318 ± 0.0423. The light dual-teacher model followed closely on ranking metrics, whereas the shared-dual CNN and student-only model remained lower on mean macro AUROC. Because the no-skill macro AUPRC baseline was already high (0.8586 ± 0.0225), the absolute AUPRC values should be read together with their modest margins above this baseline.

Download:

• PNG
  larger image
• TIFF
  original image

Table 1. Nested grouped five-fold results after inner-validation calibration and threshold selection (mean ± standard deviation).

https://doi.org/10.1371/journal.pone.0352180.t001

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. Cross-validated discrimination-versus-decision summary of the four principal calibrated models under the nested evaluation.

https://doi.org/10.1371/journal.pone.0352180.g002

Once decision thresholds were considered, the ordering changed. After inner-validation temperature scaling and threshold selection, the calibrated student-only model achieved the highest mean macro balanced accuracy, 0.6894 ± 0.0548. The corresponding values were 0.6373 ± 0.1474 for the light source-aware model, 0.6217 ± 0.1101 for the shared-dual CNN, and 0.6092 ± 0.1157 for the light dual-teacher model. Thus, better ranking performance did not automatically translate into the most favorable threshold-dependent behavior.

The main comparison is therefore best read as a separation between discrimination and operating-point behavior rather than as a single model ranking. On average, the light source-aware model separated recordings most effectively, whereas the calibrated student-only model produced the most balanced thresholded decisions under the revised nested procedure. The standard deviations across folds were large relative to the mean differences between several models, so these differences should be interpreted as tendencies rather than definitive superiority.

3.4 Effect of calibration and threshold optimization

The main effect of calibration was not a change in AUROC, which would not be expected, but a change in how models behaved once decisions had to be thresholded. In the revised analysis, temperature scaling and threshold selection were performed only on the inner validation split of each outer fold and then applied to the outer test fold. This design gives a more conservative estimate of operating-point behavior than the previous within-fold thresholding summary.

The calibrated student-only model provides the clearest example of this separation. It did not have the highest mean macro AUROC, but it achieved the highest mean macro balanced accuracy after inner-validation threshold selection. Conversely, the light source-aware model produced the strongest ranking tendency but a lower threshold-dependent balanced accuracy. For that reason, calibration is treated here as part of the evaluated decision pipeline, and the resulting operating-point summaries are interpreted alongside fold-level variability rather than as fixed-threshold clinical performance.

Fig 3 summarizes the calibration curves and pooled probability distributions used to examine post-calibration operating behavior.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 3. Reliability diagrams and pooled probability distributions before and after calibration.

https://doi.org/10.1371/journal.pone.0352180.g003

3.5 Effect of teacher guidance and source-aware weighting

The ablation results should be read cautiously because they come from fold 0 rather than the full five-fold campaign. Their role is explanatory rather than confirmatory. On fold 0, the student-only model reached macro AUROC 0.9619. The heavier dual-teacher and source-aware variants lowered this value to 0.9426 and 0.9301, respectively, even though their auxiliary objectives remained active throughout training.

Reducing the auxiliary weight improved that fold-level behavior. The light dual-teacher variant produced more balanced decisions in the exploratory fold-0 analysis, and the light source-aware variant later achieved the strongest mean discrimination in the nested five-fold comparison. That advantage was not present for every endpoint, so the safest reading is that weak external guidance may help ranking performance on average when it remains secondary to target-domain evidence, not that it provides a uniform decision-level gain.

3.6 Task-level interpretation

Task-level results showed that the two branches behaved differently. Lung AUROC tended to be higher than heart AUROC across the principal models, whereas the heart task remained more variable and often limited the macro discrimination summary. This pattern is consistent with the high heart-abnormal prevalence and the small number of heart-normal examples in each outer fold.

Calibration told a different part of the story. Lung ECE was higher than heart ECE for all principal models in Table 1, even though the most visible thresholding failures often appeared in the heart branch. This suggests that calibration difficulty and thresholded specificity collapse were not identical phenomena: the heart branch drove many operating-point instabilities, while the lung branch still showed less well-calibrated probability estimates under the mixed-source setting.

3.7 Operating-point stability across folds

The fold-wise threshold plots add information that the ranking metrics do not provide. Thresholds selected on the inner validation splits varied substantially across outer folds, particularly for the heart branch. This variation is consistent with the small validation partitions and helps explain why decision-level performance could shift even when AUROC changed less dramatically.

These threshold trajectories argue against reporting a single default cutoff without also stating how it was obtained. In the present data, workable operating points existed, but they were model- and fold-dependent rather than fixed properties of the task. The fold-by-fold supplementary table was therefore retained to show whether model ordering was stable or driven by a small number of favorable folds.