Paper deep dive
RespFocus: Learning Temporal Evidence for Weakly Supervised Respiratory Sound Classification
Yi Tian, Xiaohuan Chen, Yufei tang, Xinyu Lin, Hongliang Li, Chengbin Hou, Degen Huang, Rongshan Yu
Intelligence
Status: not_run | Model: - | Prompt: - | Confidence: 0%
Entities (0)
Relation Signals (0)
No relation signals yet.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Respiratory auscultation provides a low-cost means of assessing pulmonary function, yet transient crackles and sustained wheezes can be difficult to recognize in noisy clinical recordings. Conventional deep learning classifiers often operate as black boxes, offering limited insight into the temporal acoustic evidence supporting their predictions. We present RespFocus, a weakly supervised framework that uses multiple instance learning (MIL) to learn inspectable temporal evidence from cycle-level labels and integrate it into classification. We also introduce PedResp2026, a multicenter pediatric respiratory sound cohort collected in routine clinical environments with adjudicated cycle-level annotations. In classification experiments, RespFocus achieves four-class Scores of 69.55\% on public ICBHI 2017 and 72.91\% on PedResp2026. Evidence diagnostics and matched temporal interventions support the functional contribution of the learned evidence to model predictions. Together, these findings highlight RespFocus's potential as a tool for automated respiratory sound diagnosis.
Tags
Links
- Source: https://arxiv.org/abs/2610.15702v1
- Canonical: https://arxiv.org/abs/2610.15702v1
Trouble viewing inline? Open PDF directly →
Full Text
44,477 characters extracted from source content.
Expand or collapse full text
RESPFOCUS: LEARNING TEMPORAL EVIDENCE FOR WEAKLY SUPERVISED RESPIRATORY SOUND CLASSI- FICATION Yi Tian 1,2 Xiaohuan Chen 1,2 Yufei Tang 1,2 Xinyu Lin 1,2 Hongliang Li 1,2 Chengbin Hou 1,* Degen Huang 1,* Rongshan Yu 2,* 1 Fuyao University of Science and Technology (fyust.edu.cn) 2 Xiamen University (xmu.edu.cn) * Corresponding authors Email: tianyi319@stu.xmu.edu.cn ABSTRACT Respiratory auscultation provides a low-cost means of assessing pulmonary func- tion, yet transient crackles and sustained wheezes can be difficult to recognize in noisy clinical recordings. Conventional deep learning classifiers often operate as black boxes, offering limited insight into the temporal acoustic evidence sup- porting their predictions. We present RespFocus, a weakly supervised framework that uses multiple instance learning (MIL) to learn inspectable temporal evidence from cycle-level labels and integrate it into classification. We also introduce Pe- dResp2026, a multicenter pediatric respiratory sound cohort collected in routine clinical environments with adjudicated cycle-level annotations. In classification experiments, RespFocus achieves four-class Scores of 69.55% on public ICBHI 2017 and 72.91% on PedResp2026. Evidence diagnostics and matched temporal interventions support the functional contribution of the learned evidence to model predictions. Together, these findings highlight RespFocus’s potential as a tool for automated respiratory sound diagnosis. Conventional classification Full respiratory recording 02468 Time (s) RespFocus NCWB P r o b a b i l i t y Prediction: Both Cycle-level prediction NCWB P r o b a b i l i t y Cycle 4: Both + Temporal evidence Crackle Wheeze Cycles NCWB 02468 Time (s) Figure 1: Illustrative comparison of conventional cycle classification (left) and RespFocus (right) with cycle intervals and temporal evidence. Amber/teal denote crackle/wheeze evidence. N/C/W/B denote Normal/Crackles/Wheezes/Both. 1INTRODUCTION Respiratory auscultation provides a low-cost means of assessing respiratory sounds in routine care (Rocha et al., 2019; Bardou et al., 2018; Gairola et al., 2021). Crackles typically appear as brief, discontinuous transients, whereas wheezes exhibit a more sustained tonal structure (Pasterkamp et al., 2016). Automated respiratory sound classification remains challenging because these sounds differ in duration and can be obscured by speech, motion, and sensor-contact noise. A respiratory cycle may contain crackles, wheezes, or both, and the two subtypes need not occur at the same time. Recognizing co-occurrence therefore requires gathering evidence for each subtype 1 arXiv:2610.15702v1 [cs.LG] 11 Oct 2026 across the cycle. Keeping this evidence linked to the recording would also allow a listener to revisit the corresponding portions of the sound when reviewing a prediction. Pretrained audio representations have improved respiratory sound classification (Bae et al., 2023; He et al., 2024; Zhang et al., 2024), providing a starting point for learning such evidence. However, the annotations used here specify cycle boundaries and sound categories without marking individual events. Under this weak supervision, the model must learn which temporal responses are useful by aggregating them into a prediction and comparing it with the cycle label. Averaging across a cycle can dilute a brief crackle among less informative positions, while a single maximum captures only one response from distributed wheeze activity. A common aggregation rule therefore imposes the same temporal preference on acoustically different subtypes. This motivates separate rules that preserve crackle peaks and combine multiple wheeze responses. We present RespFocus, a weakly supervised framework that uses multiple instance learning (MIL) (Dietterich et al., 1997; Ilse et al., 2018) to learn inspectable temporal evidence from cycle-level labels and integrate it into classification. Each annotated cycle forms a bag of temporal instances. Building on interpretable time-series MIL (Early et al., 2024), it combines OPERA-CT and Audio Spectrogram Transformer (AST) representations (Zhang et al., 2024; Gong et al., 2021). A shared abnormal-presence head guides local refinement in separate subtype branches. Crackle max pooling preserves peak responses, while wheeze top-k pooling combines multiple high-scoring positions. The two subtype decisions compose the four cycle categories and represent co-occurrence within a cycle. Each prediction is accompanied by shared and subtype-specific evidence indexed to the supplied interval (Figure 1). To evaluate this approach alongside ICBHI 2017, we introduce PedResp2026, a new multicenter pe- diatric respiratory sound dataset collected during routine care. It contains 493 recordings from 137 children at seven hospitals, with expert-adjudicated cycle labels. Patient-disjoint partitions support evaluation on held-out children from the collection. The new cohort complements the public bench- mark with recordings from pediatric care and provides a resource for learning temporal evidence under weak supervision. RespFocus improves classification on both datasets. Progressive ablations examine how perfor- mance changes as the architecture is assembled, with a controlled comparison isolating the pooling design. Matched feature-space suppression and retention test whether high-ranked evidence con- tributes to predictions. These experiments evaluate the classification outcome, the design choices behind it, and the role of the selected evidence. Our contributions are: 1. A weakly supervised framework that uses MIL to learn inspectable temporal evidence from cycle-level labels and integrate it into respiratory sound classification. 2. A morphology-aware aggregation and decoding design that uses crackle max pooling and wheeze top-k pooling to reflect their temporal differences, then composes the two subtype decisions into four cycle categories. 3. PedResp2026, a new multicenter pediatric respiratory sound dataset with 493 recordings from 137 children at seven hospitals, expert-adjudicated cycle labels, and patient-disjoint evaluation partitions. 4. Evaluation on ICBHI 2017 and PedResp2026, with four-class Scores of 69.55% and 72.91%, respectively. Architectural ablations and matched feature-space interventions on ICBHI 2017 connect the classification gains to the proposed design and the evidence used for prediction. 2RELATED WORK Weakly supervised temporal learning. MIL learns from bag labels without instance annotations (Dietterich et al., 1997). Respiratory cycle annotations follow this structure, with an acoustic label for each cycle and latent locations for the supporting sounds. Attention-based MIL learns instance aggregation (Ilse et al., 2018). MILLET extends this approach to time-series classification and local explanations (Early et al., 2024). Both make aggregation an explicit link between local responses and a sequence-level decision. Weakly supervised action localization separates foreground activity from background (Paul et al., 2018; Nguyen et al., 2019; Ren et al., 2023), while sound event detection 2 Full recording 64-bin log-mel OPERA-CT 128-bin filterbank AST TSCA bridge H O H A H Shared abnormal 퐀 퐀 =퐀(퐀 퐀 /퐀 퐀 ) ℎ 퐀 =ℎ 퐀 ⊙(1+퐀 퐀耀(퐀 퐀 )) Residual guidance ℎ 퐀 Crackle DWConv (k = 3) F C Cycle abnormal support 퐀 퐀 = 1 |ℐ 퐀 | 퐀∈ℐ 퐀 퐀 퐀 Select top 퐀 퐀 (퐀)=퐀(퐀 퐀 퐀 ) 퐀 퐀 퐀 Morphology-aware MIL max (a) (b) mean σ p c C ℓ C ℓ c W σ p c W Crackles 00 Normal 10 01 Wheezes 11 Both 퐀 퐀 =(퐀 퐀 ,퐀 퐀 (퐀),퐀 퐀 (퐀)) 퐀∈ℐ 퐀 퐀 퐀 퐀 퐀 (퐀) 퐀 퐀 (퐀) 퐀 퐀 ↦ℐ 퐀 Select max 퐀 퐀 (퐀)=퐀(퐀 퐀 퐀 ) 퐀 퐀 퐀 (c) (d) Report Findings Crackle and wheeze evidence. Evidence Cycle 2, 2.1–5.1s Impression Both Evidence summary Q-Former Projector MLP Medical LLM LoRA adapter No gradlent flows to (a) or (b). (퐀 퐀耀(퐀 퐀 )) sg denotes stop-gradient Temporal evidence within a cycle Optional text summary ℎ 퐀 퐀 + Linear head Wheeze DWConv (k = 15) F W + Linear head Encoding and evidence guidance MIL and compositional decoding 퐀 퐀 Cycle 2 2.1-5.1s Within each cycle Cycle 1 Normal Cycle 2 Both Cycle 3 Normal 0s2.1s 5.1s 8.0s (퐀∈ℐ 퐀 ) Figure 2: Illustrative RespFocus overview. (a) Dual-stream fusion and guidance from shared abnor- mal evidence. (b) Subtype MIL pooling and compositional decoding. (c) Shared abnormal, crackle, and wheeze evidence. (d) Optional text summaries with temporal references. addresses event duration and acoustic background (Dinkel et al., 2021; Xin et al., 2023). RespFocus uses cycle labels to learn evidence within supplied intervals. Its subtype-specific pooling encodes the different temporal preferences of crackles and wheezes, and its compositional decoder preserves their co-occurrence within a bag. Respiratory sound representations and explanations. Convolutional and transformer models sup- port respiratory sound classification (Bardou et al., 2018; Gairola et al., 2021; Bae et al., 2023; He et al., 2024). Supervised audio pretraining, masked modeling, and acoustic tokenization pro- vide transferable features (Kong et al., 2020; Gong et al., 2022; Huang et al., 2022; Chen et al., 2023). OPERA specializes pretraining to respiratory audio (Zhang et al., 2024). RespFocus builds on OPERA-CT and AST representations by guiding, refining, and pooling their temporal features for the two sound subtypes. Prototype and example-based approaches make predictions interpretable through reference exam- ples (Ren et al., 2022; Chang et al., 2022). RespFocus exposes temporal evidence within the input cycle and tests its contribution through interventions on the model’s representations. The shared evidence describes abnormal support, and the subtype scores expose the responses pooled into the cycle-level logits. This makes the evidence accessible along the same path that produces the predic- tion. 3METHODOLOGY Given a recording x and annotated cycle intervals I c N c=1 , RespFocus returns (I c , ˆy c ,S c ) N c=1 : the supplied intervals, predicted categories, and indexed temporal evidence. Each interval defines an instance bag, with predicted category ˆy c ∈ Normal, Crackles, Wheezes, Both. Both denotes crackles and wheezes within the same cycle, whether or not their events overlap. The model fuses OPERA-aligned temporal features with AST context, uses shared abnormal evi- dence to guide subtype refinement, and pools the subtype logits within each cycle (Figure 2). The shared head modulates temporal features according to abnormal support. The subtype branches re- fine these features into crackle and wheeze responses, which cycle pooling converts into the final decision. The optional text interface verbalizes the acoustic outputs. 3.1DUAL-STREAM TEMPORAL ENCODING OPERA-CT encodes a 64-bin log-mel spectrogram, while AST encodes a normalized 128-bin fil- terbank (Zhang et al., 2024; Gong et al., 2021). Writing the frontends asM 64 andF 128 , we obtain X O =M 64 (x), H O = f O (X O )∈R T×768 , X A =F 128 (x), H A = f A (X A )∈R T A ×768 . (1) 3 Here, T is the length of the OPERA-aligned temporal sequence, and T A is the number of AST tokens. Temporal-spectral cross-attention (TSCA) projects both outputs to d b = 1024 and queries AST features with the OPERA-aligned sequence (Vaswani et al., 2017), Z O = H O W O , Z A = H A W A , ˆ Z O = LN Z O + MHA(Q = Z O ,K = Z A ,V = Z A ) . (2) A residual gate and output projection produce the shared temporal features, g O = σ φ O ( ˆ Z O ) , ̄ Z O = ˆ Z O ⊙ (0.1 + 0.9g O ), H = MLP out ( ̄ Z O ) = [h 1 ,...,h T ]∈R T×D . (3) Here, D = 512. Each supplied interval I c maps to a nonempty token set I c ⊆ 1,...,T using the nearest OPERA-token center times. Features are computed before cycle grouping. AST retains at most the first 512 input frames, so its context covers the initial segment of longer recordings. OPERA supplies the temporal indexing. 3.2SHARED ABNORMAL-PRESENCE EVIDENCE AND SUBTYPE REFINEMENT For each fused token h t , a shared head combines local evidence and an auxiliary abnormal cue into a logit s t . Its score modulates the feature through residual guidance, a t = σ(s t /τ a ), ̃ h t = h t ⊙ 1 + α sg(a t ) .(4) Here, τ a > 0 is the sigmoid temperature, α is the guidance strength, and sg stops gradients through the multiplier, broadcast across channels. The direct feature path remains trainable, and the shared head receives a separate abnormal-presence objective (Section 3.4). Inspired by convolutional local modeling (Gulati et al., 2020; Peng et al., 2022), we refine the guided sequence ̃ H = [ ̃ h 1 ,..., ̃ h T ] with separate subtype blocks, using a shorter kernel for crackles and a broader kernel for wheezes. Each block is followed by a linear subtype head, H C = F C ( ̃ H;k = 3), u C t = w ⊤ C h C t + b C , H W = F W ( ̃ H;k = 15), u W t = w ⊤ W h W t + b W . (5) Kernel sizes are in tokens and express relative temporal preferences. The subtype scores e C (t) = σ(u C t ) and e W (t) = σ(u W t ) form the indexed evidence collection S c = a t ,e C (t),e W (t) t∈I c .(6) Shared evidence a t guides refinement, while e C (t) and e W (t) expose the subtype responses whose logits are pooled for classification. All three scores retain correspondence with the recording. 3.3MORPHOLOGY-AWARE MIL AND COMPOSITIONAL DECODING Crackle pooling selects the largest logit, while wheeze pooling averages multiple high-scoring posi- tions: ℓ C c = max t∈I c u C t , K c = max(1,⌈ρ W |I c |⌉), ρ W = 0.35, ℓ W c = 1 K c X t∈TopK K c (I c ;u W ) u W t . (7) TopK K c selects the K c largest wheeze logits, whose positions can be noncontiguous. The count K c differs from the convolutional kernel size k. Building on MIL aggregation (Ilse et al., 2018; Early et al., 2024), we use asymmetric pooling to preserve peak crackle responses while combining 4 distributed wheeze evidence. Both operations produce cycle-level logits from the same temporal evidence exposed for inspection. The pooled logits yield two subtype probabilities, which are thresholded independently using validation-selected thresholds, p q c = σ(ℓ q c ), ˆ b q c = 1[p q c > θ q ], q ∈C,W.(8) The ordered pair of crackle and wheeze decisions determines the category, ˆy c = Normal,( ˆ b C c , ˆ b W c ) = (0, 0), Crackles, ( ˆ b C c , ˆ b W c ) = (1, 0), Wheezes, ( ˆ b C c , ˆ b W c ) = (0, 1), Both,( ˆ b C c , ˆ b W c ) = (1, 1). (9) The decoder assigns Both when both subtype decisions are positive. Shared abnormal evidence guides feature refinement upstream of this decision. 3.4TRAINING OBJECTIVE Let y C c ,y W c ∈ 0, 1 encode the presence of each subtype, and let y A c = 1[y C c ∨ y W c ] denote ab- normal presence. The primary objective applies asymmetric loss (ASL) to the MIL-pooled subtype logits (Ridnik et al., 2021), L pair = 1 N N X c=1 ASL (ℓ C c ,ℓ W c ), (y C c ,y W c ) .(10) The auxiliary objective trains the shared head to recognize abnormal support across both subtypes. It averages the shared evidence logits within each cycle and applies a binary abnormal-presence loss, ̄s c = 1 |I c | X t∈I c s t , L abn = 1 N N X c=1 BCEWithLogits( ̄s c ,y A c ). (11) The acoustic model is optimized with L =L pair + ω abn L abn , ω abn = 1.(12) Both labels activate both subtype targets, while Normal labels set both to zero. Each subtype branch can therefore learn from cycles containing its subtype alone or together with the other subtype. The losses supervise temporal responses through cycle aggregation, allowing evidence learning without event-level annotations. For evidence analysis and presentation, we separately summarize the shared scores as r c = 1 |I c | X t∈I c a t .(13) Thus, analysis uses the mean probability r c , whereas training uses the mean logit ̄s c . 5 4EXPERIMENTAL SETUP 4.1DATASETS We evaluate respiratory sound classification on ICBHI 2017 (Rocha et al., 2019) and PedResp2026. Our fixed ICBHI partition uses 777/90/53 recordings and 5,815/657/426 cycles for training, valida- tion, and testing. Test recordings follow the OPERA held-out list. PedResp2026 is a newly collected pediatric cohort comprising 493 recordings from 137 children at seven hospitals, with nine device identifiers. Children aged 0.1–13 years were recorded in rou- tine clinical rooms using an eight-site thoracic protocol. Two physicians annotated cycles, with adjudication by an additional physician from an eight-physician panel. The patient-disjoint train- ing, validation, and test sets contain 109/14/14 children, 394/49/50 recordings, and 899/119/117 respiratory cycles, respectively. 4.2BASELINES AND IMPLEMENTATION We compare with respiratory sound classifiers using convolutional, transformer, contrastive, masked-audio, and adaptation-based approaches (Moummad & Farrugia, 2023; Bae et al., 2023; Kim et al., 2023; 2024b;a; Niizumi et al., 2024; He et al., 2024; Dong et al., 2025a;b). We also consider general MIL baselines for time-series classification (Early et al., 2024; Chen et al., 2024), anomaly detection (Jang & Kwon, 2025), and instance aggregation (Tang et al., 2024), alongside an interpretable mixture-of-experts model (Xie et al., 2026), for adaptation to respiratory sound classification. OPERA-CT uses 16-kHz audio, 64 mel bins, a 2,048-sample FFT/window, and a 512-sample hop. AST uses a normalized 128-bin Kaldi filterbank, a 10-ms frame shift, and 512 input frames. We train with AdamW and cosine learning-rate decay over five independent seeds. Checkpoints and subtype thresholds are selected on validation data and fixed for testing. The wheeze pooling ratio is ρ W = 0.35, and the crackle and wheeze kernels are 3 and 15 tokens. 4.3EVALUATION METRICS Classification. We use the ICBHI challenge metric definitions (Rocha et al., 2019) on the study par- titions above. Let M ij count cycles with reference class i and predicted class j, where N,C,W,B denote Normal, Crackles, Wheezes, and Both. Specificity (Sp) measures correct normal classifi- cation. Sensitivity (Se) measures correct abnormal-subtype classification, penalizing both missed abnormalities and subtype confusions: Sp = M N P j M Nj , Se = M C + M W W + M B P i∈C,W,B P j M ij , Score = 1 2 (Sp + Se). (14) We report Sp, Se, and Score as percentages, alongside macro-AUROC and macro-F1. Temporal evidence. We evaluate two properties of the shared evidence. Abnormal-presence AU- ROC uses r c to rank cycles labeled Normal versus Abnormal. CycleHitRate measures the fraction of abnormal cycles containing at least one token above a validation-selected evidence threshold, CycleHitRate = 1 |C abn | X c∈C abn 1 max t∈I c a t ≥ τ hit .(15) Here, C abn contains abnormal cycles. We test functional relevance by suppressing or retaining se- lected tokens under matched feature-space budgets, including random, low-evidence, and energy- matched controls. 6 AnchorMoEMILLETRRT-MILTAIL-MILTimeMILRespFocus S c o r e ( % ) 40 60 80 A U R O C ( % ) 40 60 80 M a c r o - F 1 0.2 0.5 0.8 135135135135135135 ICBHI 2017PedResp2026Right: mean ± SD Figure 3: Five-run results by metric (rows) and method (columns) for ICBHI 2017 (solid black lines, filled circles) and PedResp2026 (colored dashed lines, open diamonds). Right summaries show mean± sample SD. Run indices do not imply paired initialization across methods. 5RESULTS 5.1CLASSIFICATION PERFORMANCE RespFocus achieves the highest mean four-class Score among the compared methods on both datasets (Table 1). Scores of 69.55% on ICBHI and 72.91% on PedResp2026 exceed the strongest baselines, MVST (66.56%) and Patch-Mix CL (68.41%), by 2.99 and 4.50 percentage points. Resp- Focus also achieves the highest sensitivity, improving by 4.92 and 16.06 points over these baselines. On PedResp2026, sensitivity rises from 52.12% to 68.18%, showing stronger abnormal-subtype recognition. Table 1: Four-class classification (%). Sc denotes Score. RespFocus reports five-run means. Bold/underlined values mark best/second-best results. MethodBackboneVenueICBHI 2017PedResp2026 ScSeSpScSeSp CE (Moummad & Farrugia, 2023)CNN6WASPAA’2355.19 40.32 70.07 50.00 20.00 80.00 SCL (Moummad & Farrugia, 2023) CNN6WASPAA’2357.57 39.21 75.93 61.92 46.97 76.86 AFT on Mixed (Kim et al., 2023)ASTNeurIPS W’2361.80 42.86 80.73 65.31 47.88 82.75 Fine-tuning (Bae et al., 2023)ASTINTERSPEECH’23 59.52 41.90 77.13 57.08 27.88 86.27 Patch-Mix CL (Bae et al., 2023)ASTINTERSPEECH’23 62.34 43.02 81.67 68.4152.1284.71 Audio-CLAP (Kim et al., 2024b)CLAPINTERSPEECH’24 62.73 44.60 80.87 65.24 49.70 80.78 BTS (Kim et al., 2024b)CLAPINTERSPEECH’24 63.56 45.71 81.40 64.83 48.48 81.18 DAT (Kim et al., 2024a)ASTICASSP’2459.84 42.54 77.13 64.13 43.94 84.31 SG-SCL (Kim et al., 2024a)ASTICASSP’2461.68 43.49 79.87 64.73 51.82 77.65 M2D-X (Niizumi et al., 2024)M2D ViTTASLP’2463.31 45.08 81.53 62.50 46.97 78.04 MVST (He et al., 2024)ASTICASSP’2466.5651.1182.00 54.44 28.48 80.39 Low-rank (Dong et al., 2025a)LoRAICASSP’2560.68 36.03 85.33 56.76 16.67 96.86 Single-view (Dong et al., 2025b)ASTBSPC’2566.03 49.52 82.53 51.75 38.79 64.71 RespFocusDual-stream This work69.55 56.03 83.0772.91 68.18 77.65 RespFocus also outperforms five adapted MIL and temporal baselines in mean Score, macro- AUROC, and macro-F1 on both datasets. The strongest baselines by Score are AnchorMoE (Xie et al., 2026) on ICBHI and RRT-MIL (Tang et al., 2024) on PedResp2026. RespFocus exceeds their Scores by 11.91 and 19.85 percentage points, respectively. Compared with RRT-MIL, the strongest macro-F1 baseline on both datasets, RespFocus raises macro-F1 from 0.371 to 0.625 on ICBHI and from 0.372 to 0.744 on PedResp2026. These gains show that the advantage extends to perfor- mance averaged equally across the four classes. Figure 3 shows the variation across five runs, with RespFocus Score SDs of 4.92 points on ICBHI and 2.42 on PedResp2026. We next examine the design choices within RespFocus through ablations, complementing these comparisons of complete classification pipelines. 7 Table 2: Progressive ICBHI 2017 four-class ablation (%, five-run mean± SD). VariantTSCAEvidenceBinary pairAsym. poolingSp (%)Se (%)Score (%) Flat AST classifier× 76.40± 6.35 43.17± 4.07 59.79± 4.47 Dual-stream global pooling✓× 79.40± 5.67 45.40± 3.48 62.40± 4.11 Evidence pathway + flat four-way✓× 81.07± 5.91 48.25± 3.52 64.66± 4.28 Mean pooling✓× 82.67± 5.48 50.95± 3.14 66.81± 2.85 Full RespFocus✓ 83.07± 8.50 56.03± 4.07 69.55± 4.92 CrackleWheeze Original waveform Relative evidence High evidence suppression Random suppression 1 0 0.00.51.01.52.02.5 Time (s) 1 0 0.00.51.01.52.02.5 Time (s) DeletionRetention High evidenceRandomEnergy matched Crackle Wheeze 0 0.005 0.010 0255075100 D C 0.590 0.600 0.610 0.620 0255075100 R C 0 0.005 0.010 0255075100 D W Deleted (%) 0.590 0.600 0.610 0.620 0255075100 R W Retained (%) Figure 4: Illustrative waveform interventions with within-cycle min–max-normalized evidence (left). Right panels show deletion D q (b) = p q (x)−p q (x delete,b ) and retention R q (b) = p q (x retain,b ) for subtype q ∈C,W and cycle-duration fraction b. 5.2ARCHITECTURAL ABLATIONS Adding temporal encoding, evidence guidance, and subtype decisions progressively improves ICBHI classification (Table 2). The mean ICBHI Score rises from 59.79% for a flat AST classifier to 62.40% with dual-stream TSCA, 64.66% with the evidence pathway, and 66.81% with binary-pair decoding and mean pooling. The last comparison holds the remaining architecture and training pro- tocol fixed: replacing mean pooling in both subtype branches with crackle max pooling and wheeze top-k pooling raises Score to 69.55%, a 2.74-point gain. The pooling change raises sensitivity from 50.95% to 56.03%, while specificity changes from 82.67% to 83.07%. Most of the Score increase therefore comes from correct abnormal-subtype predictions. Since this sensitivity metric rewards the correct Crackles, Wheezes, or Both category, the result links the aggregation change to recognition of the sounds present within abnormal cycles. 5.3FUNCTIONAL ROLE OF TEMPORAL EVIDENCE The shared evidence distinguishes abnormal cycles and contributes to feature refinement. Its cycle summary r c achieves an AUROC of (83.51± 0.48)% for separating abnormal from normal cycles. To test its role in refinement, we perturb the shared guidance scores a t at a fixed checkpoint. Replac- ing the scores with the recording mean removes temporal variation and reduces Score from 70.02% to 65.78%. Removing guidance lowers Score to 62.63%. Within-cycle shuffling preserves each cycle’s score distribution but changes its temporal assignment, yielding a Score of 68.28%. These decreases show that the classifier uses the shared guidance, including its temporal arrangement. Matched feature-space interventions then assess the contribution of the selected representations. Under equal token budgets, suppressing high-evidence tokens reduces mean ICBHI Score by 11.13 percentage points, compared with 4.81 for random tokens and 6.23 for energy-matched random to- kens. Retaining only high-evidence tokens yields a Score of 65.12%, versus 55.17% for random retention. The larger suppression effect and stronger retention performance jointly support the evi- dence ranking as a way to identify features that contribute to classification. Figure 4 illustrates complementary interventions on waveform windows ranked by subtype evidence. On the left, normalized crackle evidence peaks near a brief transient, while wheeze evidence remains elevated over a longer interval. The highlighted windows show where high-evidence and random suppression modify the waveforms. On the right, deleting high-evidence windows causes larger drops in subtype probability than random or energy-matched deletion at partial budgets. Retaining high-evidence windows also yields higher subtype probabilities at small and intermediate budgets. The curves converge at the full-cycle budget, when all strategies delete or retain the same samples. These examples connect the temporal evidence profiles to the prediction changes caused by selecting different portions of the waveform. 8 Figure 5 illustrates two cycles from a wheeze-dominant case. RespFocus correctly predicts Cycle 2 as Both but also assigns Both to the Wheezes-only Cycle 1. The extra crackle prediction propagates into the illustrative text panel. 6DISCUSSION (a) Recorded audio dB 0 −60 kHz 0.05 0.2 1 4 Cycle 1Cycle 2 1.966.4911.37 Time (s) (b) Temporal evidence C W (c) Decisions(d) Report Cycle 1Cycle 2 Expert WheezesBoth Pred.BothBoth Cycle 1: extra crackle Cycle 2: label matched Findings: Crackle and wheeze evidence in both cycles. Evidence: Cycle 1, 2.07–5.65 s Cycle 2, 6.82–9.93 s Impression:Both Figure 5: Two-cycle illustration of a wheeze- dominant case. RespFocus connects weakly supervised respira- tory sound classification with inspection of the temporal evidence used for each decision. The classification results (Table 1) and pooling abla- tion (Table 2) support using sound morphology to guide how evidence is accumulated. This al- lows brief crackle responses and longer wheeze activity to contribute through different aggrega- tion rules, while compositional decoding repre- sents their co-occurrence within a cycle. The matched interventions show that the evidence ranking identifies representations on which pre- dictions depend. Because these representations retain their temporal indices, the predicted cat- egory can be examined alongside the responses that produced it. This gives a listener a con- crete starting point for reviewing a prediction and investigating an incorrectly assigned sub- type. PedResp2026 provides a new multicenter pediatric setting for studying this connection, with adjudicated recordings from routine care and evaluation on held-out children. The optional text interface conditions generation on acoustic pre- dictions, subtype probabilities, and cycle intervals. This conditioning provides an explicit reference for generation and can reduce the risk of language model hallucinations. The extra crackle statement in Figure 5 reflects an error inherited from the acoustic model. Limitations. PedResp2026 remains limited in size, and future work will expand the cohort to include more patients and a wider range of respiratory sounds. Expert event annotation is also labor- intensive, particularly for brief crackles that require millisecond-scale temporal precision. Learning from cycle-level labels reduces this annotation burden, although evaluating event localization still requires fine-grained reference annotations. The current patient-disjoint split leaves generalization to unseen hospitals and devices open. Future evaluations should test these acquisition-domain shifts (Koh et al., 2021). Evaluation of the text interface should distinguish errors propagated from the acoustic model from unsupported content introduced during text generation. It should also assess how conditioning affects hallucination risk and whether the generated text helps listeners review recordings. 7CONCLUSION We presented RespFocus, a weakly supervised framework that uses MIL to learn temporal evidence from cycle-level labels and integrate it into respiratory sound classification. Shared abnormal evi- dence guides separate crackle and wheeze branches, whose pooling rules preserve brief peaks and combine distributed responses before compositional decoding determines the cycle category. Eval- uation on ICBHI 2017 and PedResp2026, our new multicenter pediatric cohort, shows improved classification, while matched interventions support the contribution of the learned evidence to pre- dictions. These findings support learning temporal evidence as part of the classifier, connecting each predicted sound category to the temporal responses used to produce it. 9 AI USE STATEMENT We used GPT-6 Astra for language polishing, literature search, manuscript-structure review, and checks of numerical consistency and claim–evidence alignment. The authors take full responsibility for the manuscript’s content, including its text, references, and scientific claims. ETHICS STATEMENT The PedResp2026 study received institutional ethics approval. Written informed consent was ob- tained from legal guardians, with child assent where appropriate. The data were de-identified before analysis to protect participant privacy. We plan to release PedResp2026 through controlled access upon acceptance to support reproducibility and further research. REPRODUCIBILITY STATEMENT Our code and dataset access information are available at https://anonymous.4open. science/r/RespFous-31C6/. REFERENCES Sangmin Bae, June-Woo Kim, Won-Yang Cho, Hyerim Baek, Soyoun Son, Byungjo Lee, Changwan Ha, Kyongpil Tae, Sungnyun Kim, and Se-Young Yun. Patch-mix contrastive learning with audio spectrogram transformer on respiratory sound classification. In Proceedings of Interspeech 2023, p. 5436–5440. ISCA, 2023. doi: 10.21437/Interspeech.2023-1426. URL https://doi. org/10.21437/Interspeech.2023-1426. Dalal Bardou, Kun Zhang, and Sayed Mohammad Ahmad. Lung sounds classification using con- volutional neural networks. Artificial Intelligence in Medicine, 88:58–69, 2018. doi: 10.1016/j. artmed.2018.04.008. URL https://doi.org/10.1016/j.artmed.2018.04.008. Yi Chang, Zhao Ren, Thanh Tam Nguyen, Wolfgang Nejdl, and Bj ̈ orn W. Schuller. Example-based explanations with adversarial attacks for respiratory sound analysis. In Proceedings of Interspeech 2022, p. 4003–4007. ISCA, 2022. doi: 10.21437/Interspeech.2022-11355. URL https:// w.isca-archive.org/interspeech_2022/chang22h_interspeech.html. Sanyuan Chen, Yu Wu, Chengyi Wang, Shujie Liu, Daniel Tompkins, Zhuo Chen, Wanxiang Che, Xiangzhan Yu, and Furu Wei. BEATs: Audio pre-training with acoustic tokenizers. In Proceed- ings of the 40th International Conference on Machine Learning, volume 202 of Proceedings of Machine Learning Research, p. 5178–5193. PMLR, 2023. URL https://proceedings. mlr.press/v202/chen23ag.html. Xiwen Chen, Peijie Qiu, Wenhui Zhu, Huayu Li, Hao Wang, Aristeidis Sotiras, Yalin Wang, and Abolfazl Razi. TimeMIL: Advancing multivariate time series classification via a time-aware multiple instance learning. In Proceedings of the 41st International Conference on Machine Learning, volume 235 of Proceedings of Machine Learning Research, p. 7190–7206. PMLR, 2024. URL https://proceedings.mlr.press/v235/chen24af.html. Thomas G. Dietterich, Richard H. Lathrop, and Tomas Lozano-Perez. Solving the multiple instance problem with axis-parallel rectangles. Artificial Intelligence, 89(1–2):31–71, 1997. doi: 10. 1016/S0004-3702(96)00034-3. URL https://doi.org/10.1016/S0004-3702(96) 00034-3. Heinrich Dinkel, Mengyue Wu, and Kai Yu. Towards duration robust weakly supervised sound event detection. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 29:887–900, 2021. doi: 10.1109/TASLP.2021.3054313. URL https://doi.org/10.1109/TASLP. 2021.3054313. Gaoyang Dong, Yufei Shen, Jianhong Wang, Shunwang Xie, Minghui Zhang, and Ping Sun. Low- rank adaptation method for respiratory sound classification: A necessary road towards large mod- els. In ICASSP 2025–2025 IEEE International Conference on Acoustics, Speech and Signal 10 Processing (ICASSP), p. 1–5. IEEE, 2025a. doi: 10.1109/ICASSP49660.2025.10889965. URL https://ieeexplore.ieee.org/document/10889965. Gaoyang Dong, Yufei Shen, Jianhong Wang, Mingli Zhang, Ping Sun, and Minghui Zhang. Res- piratory sounds classification by fusing the time-domain and 2D spectral features. Biomedical Signal Processing and Control, 107:107790, 2025b. doi: 10.1016/j.bspc.2025.107790. URL https://doi.org/10.1016/j.bspc.2025.107790. Joseph Early,Gavin Cheung,Kurt Cutajar,Hanting Xie,Jas Kandola,and Niall Twomey.Inherently interpretable time series classification via multiple instance learning.In International Conference on Learning Representations, 2024.URL https://proceedings.iclr.c/paper_files/paper/2024/hash/ fb9d01fb202e360b2f78510c47e0fa3e-Abstract-Conference.html. Siddhartha Gairola, Francis Tom, Nipun Kwatra, and Mohit Jain.RespireNet: A deep neu- ral network for accurately detecting abnormal lung sounds in limited data setting. In 2021 43rd Annual International Conference of the IEEE Engineering in Medicine and Biology So- ciety (EMBC), p. 527–530. IEEE, 2021. doi: 10.1109/EMBC46164.2021.9630091. URL https://doi.org/10.1109/EMBC46164.2021.9630091. Yuan Gong, Yu-An Chung, and James Glass. AST: Audio spectrogram transformer. In Proceedings of Interspeech 2021, p. 571–575. ISCA, 2021. doi: 10.21437/Interspeech.2021-698. URL https://doi.org/10.21437/Interspeech.2021-698. Yuan Gong, Cheng-I Lai, Yu-An Chung, and James Glass. SSAST: Self-supervised audio spectro- gram transformer. Proceedings of the AAAI Conference on Artificial Intelligence, 36(10):10699– 10709, 06 2022. doi: 10.1609/aaai.v36i10.21315. URL https://ojs.aaai.org/index. php/AAAI/article/view/21315. Anmol Gulati, James Qin, Chung-Cheng Chiu, Niki Parmar, Yu Zhang, Jiahui Yu, Wei Han, Shibo Wang, Zhengdong Zhang, Yonghui Wu, and Ruoming Pang. Conformer: Convolution- augmented transformer for speech recognition. In Proceedings of Interspeech 2020, p. 5036– 5040, 2020. doi: 10.21437/Interspeech.2020-3015. URL https://w.isca-archive. org/interspeech_2020/gulati20_interspeech.html. Wentao He, Yuchen Yan, Jianfeng Ren, Ruibin Bai, and Xudong Jiang. Multi-view spectro- gram transformer for respiratory sound classification. In ICASSP 2024–2024 IEEE Interna- tional Conference on Acoustics, Speech and Signal Processing (ICASSP), p. 8626–8630. IEEE, 2024. doi: 10.1109/ICASSP48485.2024.10445825. URL https://doi.org/10.1109/ ICASSP48485.2024.10445825. Po-Yao Huang, Hu Xu, Juncheng Li, Alexei Baevski, Michael Auli, Wojciech Galuba, Florian Metze, and Christoph Feichtenhofer. Masked autoencoders that listen. In Proceedings of the 36th International Conference on Neural Information Processing Systems, NIPS ’22, Red Hook, NY, USA, 2022. Curran Associates Inc. ISBN 9781713871088. Maximilian Ilse, Jakub M. Tomczak, and Max Welling. Attention-based deep multiple instance learning. In Proceedings of the 35th International Conference on Machine Learning, volume 80 of Proceedings of Machine Learning Research, p. 2127–2136. PMLR, 2018. URL https: //proceedings.mlr.press/v80/ilse18a.html. Jaeseok Jang and Hyuk-Yoon Kwon. TAIL-MIL: Time-aware and instance-learnable multiple in- stance learning for multivariate time series anomaly detection. In Proceedings of the AAAI Con- ference on Artificial Intelligence, volume 39, p. 17582–17589, 2025. doi: 10.1609/aaai.v39i17. 33933. URL https://ojs.aaai.org/index.php/AAAI/article/view/33933. June-Woo Kim, Chihyeon Yoon, Miika Toikkanen, Sangmin Bae, and Ho-Young Jung. Adversar- ial fine-tuning using generated respiratory sound to address class imbalance. In NeurIPS 2023 Workshop on Deep Generative Models for Health, 2023. URL https://openreview.net/ forum?id=z1AVG5LDQ7. 11 June-Woo Kim, Sangmin Bae, Won-Yang Cho, Byungjo Lee, and Ho-Young Jung. Stethoscope- guided supervised contrastive learning for cross-domain adaptation on respiratory sound clas- sification. In ICASSP 2024–2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), p. 1431–1435. IEEE, 2024a. doi: 10.1109/ICASSP48485.2024. 10447734. URL https://doi.org/10.1109/ICASSP48485.2024.10447734. June-Woo Kim, Miika Toikkanen, Yera Choi, Seoung-Eun Moon, and Ho-Young Jung. BTS: Bridging text and sound modalities for metadata-aided respiratory sound classification.In Proceedings of Interspeech 2024, p. 1690–1694. ISCA, 2024b. doi: 10.21437/Interspeech. 2024-492. URL https://w.isca-archive.org/interspeech_2024/kim24f_ interspeech.html. Pang Wei Koh, Shiori Sagawa, Henrik Marklund, Sang Michael Xie, Marvin Zhang, Akshay Balsub- ramani, Weihua Hu, Michihiro Yasunaga, Richard Lanas Phillips, Irena Gao, Tony Lee, Etienne David, Ian Stavness, Wei Guo, Berton Earnshaw, Imran Haque, Sara M Beery, Jure Leskovec, Anshul Kundaje, Emma Pierson, Sergey Levine, Chelsea Finn, and Percy Liang. WILDS: A benchmark of in-the-wild distribution shifts. In Proceedings of the 38th International Conference on Machine Learning, volume 139 of Proceedings of Machine Learning Research, p. 5637– 5664. PMLR, 2021. URL https://proceedings.mlr.press/v139/koh21a.html. Qiuqiang Kong, Yin Cao, Turab Iqbal, Yuxuan Wang, Wenwu Wang, and Mark D. Plumbley. PANNs: Large-scale pretrained audio neural networks for audio pattern recognition. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 28:2880–2894, 2020. doi: 10.1109/ TASLP.2020.3030497. URL https://doi.org/10.1109/TASLP.2020.3030497. Ilyass Moummad and Nicolas Farrugia. Pretraining respiratory sound representations using meta- data and contrastive learning. In 2023 IEEE Workshop on Applications of Signal Processing to Audio and Acoustics (WASPAA), p. 1–5. IEEE, 2023. doi: 10.1109/WASPAA58266.2023. 10248130. URL https://doi.org/10.1109/WASPAA58266.2023.10248130. Phuc X. Nguyen, Deva Ramanan, and Charless C. Fowlkes.Weakly-supervised action lo- calization with background modeling.In Proceedings of the IEEE/CVF International Conference on Computer Vision, p. 5502–5511, 2019.doi: 10.1109/ICCV.2019.00560. URL https://openaccess.thecvf.com/content_ICCV_2019/html/Nguyen_ Weakly-Supervised_Action_Localization_With_Background_Modeling_ ICCV_2019_paper.html. Daisuke Niizumi, Daiki Takeuchi, Yasunori Ohishi, Noboru Harada, and Kunio Kashino. Masked modeling duo: Towards a universal audio pre-training framework. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 32:2391–2406, 2024. doi: 10.1109/TASLP.2024. 3389636. URL https://doi.org/10.1109/TASLP.2024.3389636. Hans Pasterkamp, Paul L. P. Brand, Mark Everard, Luis Garcia-Marcos, Hasse Melbye, and Kostas N. Priftis. Towards the standardisation of lung sound nomenclature. European Res- piratory Journal, 47(3):724–732, 2016. doi: 10.1183/13993003.01132-2015. URL https: //doi.org/10.1183/13993003.01132-2015. Sujoy Paul, Sourya Roy, and Amit K. Roy-Chowdhury. W-TALC: Weakly-supervised temporal activity localization and classification. In Proceedings of the European Conference on Computer Vision, p. 588–607, 2018. doi: 10.1007/978-3-030-01225-0 35. URL https://doi.org/ 10.1007/978-3-030-01225-0_35. Yifan Peng, Siddharth Dalmia, Ian Lane, and Shinji Watanabe. Branchformer: Parallel MLP- attention architectures to capture local and global context for speech recognition and under- standing. In Proceedings of the 39th International Conference on Machine Learning, volume 162 of Proceedings of Machine Learning Research, p. 17627–17643. PMLR, 2022. URL https://proceedings.mlr.press/v162/peng22a.html. Huan Ren, Wenfei Yang, Tianzhu Zhang, and Yongdong Zhang.Proposal-based multiple instance learning for weakly-supervised temporal action localization.In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 2394–2404, 2023.doi: 10.1109/CVPR52729.2023.00237.URL https://openaccess.thecvf. 12 com/content/CVPR2023/html/Ren_Proposal-Based_Multiple_Instance_ Learning_for_Weakly-Supervised_Temporal_Action_Localization_ CVPR_2023_paper.html. Zhao Ren, Thanh Tam Nguyen, and Wolfgang Nejdl. Prototype learning for interpretable respiratory sound analysis. In ICASSP 2022–2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), p. 9087–9091. IEEE, 2022. doi: 10.1109/ICASSP43922.2022. 9747014. URL https://doi.org/10.1109/ICASSP43922.2022.9747014. Tal Ridnik, Emanuel Ben-Baruch, Nadav Zamir, Asaf Noy, Itamar Friedman, Matan Prot- ter, and Lihi Zelnik-Manor.Asymmetric loss for multi-label classification.In Proceed- ings of the IEEE/CVF International Conference on Computer Vision, p. 82–91, 2021. doi:10.1109/ICCV48922.2021.00015.URL https://openaccess.thecvf.com/ content/ICCV2021/html/Ridnik_Asymmetric_Loss_for_Multi-Label_ Classification_ICCV_2021_paper.html. Bruno M. Rocha, Dimitris Filos, Lu ́ ıs Mendes, Gorkem Serbes, Sezer Ulukaya, Yasemin P. Kahya, Nik ˇ sa Jakovljevic, Tatjana L. Turukalo, Ioannis M. Vogiatzis, Eleni Perantoni, Evangelos Kaimakamis, Pantelis Natsiavas, Ana Oliveira, Cristina J ́ acome, Alda Marques, Nicos Maglav- eras, Rui Pedro Paiva, Ioanna Chouvarda, and Paulo de Carvalho. An open access database for the evaluation of respiratory sound classification algorithms. Physiological Measurement, 40(3): 035001, 2019. doi: 10.1088/1361-6579/ab03ea. URL https://pubmed.ncbi.nlm.nih. gov/30708353/. Wenhao Tang, Fengtao Zhou, Sheng Huang, Xiang Zhu, Yi Zhang, and Bo Liu. Feature re- embedding: Towards foundation model-level performance in computational pathology. In Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 11343– 11352, 2024.URL https://openaccess.thecvf.com/content/CVPR2024/ html/Tang_Feature_Re-Embedding_Towards_Foundation_Model-Level_ Performance_in_Computational_Pathology_CVPR_2024_paper.html. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems, volume 30, 2017. URL https://papers.nips.c/paper_files/ paper/2017/hash/3f5e243547dee91fbd053c1c4a845a-Abstract.html. Tao Xie, Zexi Tan, Haoyi Xiao, Mengke Li, Yiqun Zhang, Yang Lu, Cuie Yang, and Yiu-ming Cheung. Anchormoe: Interpretable time series classification via anchor-routed moe. In Pro- ceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.2, KDD ’26, p. 5720–5731, New York, NY, USA, 2026. Association for Computing Machin- ery. ISBN 9798400722592. doi: 10.1145/3770855.3817786. URL https://doi.org/10. 1145/3770855.3817786. Yifei Xin, Dongchao Yang, and Yuexian Zou. Background-aware modeling for weakly super- vised sound event detection.In Proceedings of Interspeech 2023, p. 1199–1203. ISCA, 2023. doi: 10.21437/Interspeech.2023-330. URL https://w.isca-archive.org/ interspeech_2023/xin23_interspeech.html. Yuwei Zhang, Tong Xia, Jing Han, Yu Y. Wu, Georgios Rizos, Yang Liu, Mo- hammed Mosuily, Jagmohan Chauhan, and Cecilia Mascolo.Towards open res- piratory acoustic foundation models:Pretraining and benchmarking.In Advances in Neural Information Processing Systems,volume 37,p. 27024–27055,2024. URL https://proceedings.neurips.c/paper_files/paper/2024/ hash/2f803abdcad9de35b45d5a656dade45c-Abstract-Datasets_and_ Benchmarks_Track.html. 13