{"id":"e59b7c68-852c-403a-98fd-3b715ce915ab","arxiv_id":"2606.02998","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"CoughSense reports 82.3% balanced accuracy on 5-class cough classification from 18k aggregated recordings using Whisper fine-tuning, active-frame pooling, and dual-encoder fusion, outperforming image-model baselines.","lead":"CoughSense classifies cough recordings into five respiratory conditions using a fine-tuned Whisper encoder with active-frame attention pooling to focus on short cough segments. A generalist might read it to see how AI techniques could support low-cost smartphone screening for lung issues.","discovery_kind":"new_application","skeptic_critique":{"model":"grok-4.3","headline":"Label quality and domain-shift compensation in aggregated public datasets remain the least-secured premise for the 82.3% balanced-accuracy claim.","rationale":"The reader's weakest_assumption directly names the same point. Because the full manuscript is now available, the concern can be stated more precisely (label audit + leakage check), but the core risk remains unchanged; hence CONDITIONAL rather than UNVERDICTED.","tokens_in":1987,"tokens_out":349,"duration_ms":12438,"concrete_test":"Randomly sample 200 recordings (50 per source), obtain independent clinician re-labeling, compute label-flip rate; if >8%, retrain the Whisper-tiny model with the corrected labels under identical five-fold protocol and report change in balanced accuracy and per-class recall.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The headline result (82.3% balanced accuracy on five-fold CV, 5.1-point gain from active-frame pooling) is obtained on the union of Coswara, CoughVID, Virufy and West China Hospital data. The paper lists WeightedRandomSampler, Balanced Mixup, supervised contrastive loss and gradient-reversal adaptation as countermeasures, yet provides no quantitative audit of label noise rates or patient-level leakage across the four sources. If inter-rater or self-report error exceeds ~10% (common in crowd-sourced cough corpora), the macro-F1 of 0.817 and the ablation deltas become difficult to interpret as evidence of architectural superiority rather than dataset artifact. This assumption is load-bearing because every other component (Whisper encoder, cross-attention fusion, FiLM conditioning) is evaluated on the same potentially noisy aggregate.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The paper presents CoughSense, a five-class respiratory disease classifier (healthy, COVID-19, asthma/respiratory condition, bronchitis, pneumonia) built on the Whisper-tiny encoder (8.6M parameters). It aggregates 18,301 recordings from Coswara, CoughVID, Virufy and West China Hospital datasets and introduces active-frame QKV attention pooling that restricts attention to the first 200 of 1500 encoder tokens to mitigate silence dilution in short cough audio. Additional components include WeightedRandomSampler, Balanced Mixup, supervised contrastive loss, FiLM symptom conditioning and gradient-reversal domain adaptation to handle 19:1 imbalance and four-source domain shift. A dual-encoder variant fuses Whisper with OPERA-CT via cross-attention. On five-fold cross-validation the single-encoder model reports 82.3% balanced accuracy (macro-F1 0.817, AUC 0.941), outperforming an ImageNet-pretrained EfficientNet-B2 by 11.1 points and a ViT trained from scratch by 29.6 points; the dual-encoder reaches 85.4%. Ablations identify active-frame pooling as the largest single contributor (+5.1 points).","tokens_in":2165,"tokens_out":812,"duration_ms":23586,"significance":"If validated, the work shows that a lightweight pretrained audio encoder plus targeted pooling can deliver practical multi-class cough screening on consumer devices, with the active-frame mechanism offering a reusable solution for short-audio tasks that exceed Whisper's native 30-second window. The empirical gains over strong baselines and the dual-encoder fusion provide concrete evidence of utility for heterogeneous medical audio. The paper supplies reproducible cross-validation metrics and an ablation isolating one architectural change, which strengthens its contribution relative to purely empirical claims.","major_comments":[{"comment":"§3 (Dataset Aggregation): the four public datasets are combined without any reported quantitative audit of label noise rates, inter-rater agreement, or patient-level leakage checks. Because crowd-sourced corpora commonly exhibit self-report error rates above 10%, the 82.3% balanced accuracy, 0.817 macro-F1 and the 5.1-point active-frame ablation delta cannot be confidently attributed to architectural superiority rather than dataset artifacts; this assumption is load-bearing for all performance claims.","section":"§3"},{"comment":"§5 (Results and Ablations): the five-fold CV metrics are reported without error bars, confidence intervals or statistical significance tests against the EfficientNet-B2 and ViT baselines. In addition, the ablation table combines WeightedRandomSampler, Balanced Mixup, contrastive loss and domain adaptation without exhaustive isolation controls, so the isolated 5.1-point contribution of active-frame pooling cannot be verified as independent of the other regularizers.","section":"§5"},{"comment":"§4 (Training Protocol): exact patient-level train-test splits, full hyperparameter schedules and the precise implementation of gradient-reversal domain adaptation are not supplied. Without these details the claim that the listed techniques fully compensate for domain shift and imbalance remains unverifiable and limits assessment of whether the reported gains generalize beyond the aggregated collection.","section":"§4"}],"minor_comments":[{"comment":"The abstract states that 'all five classes passed 74 percent recall' yet the main results table does not list per-class recalls; adding these values would improve clarity.","section":"Abstract / §5"},{"comment":"Notation for the active-frame token limit (200 tokens) and the loss-balancing weights is introduced without an explicit equation or hyperparameter table entry.","section":"§4"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a solid applied ML submission for cs.LG; the main concern is whether the journal expects stronger clinical or statistical validation of label quality before acceptance."},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the constructive and detailed feedback. We address each major comment point-by-point below, providing clarifications and indicating revisions where feasible without misrepresenting the work.","responses":[{"response":"We agree that label noise is an inherent risk in crowd-sourced datasets such as Coswara and CoughVID. These are the standard public corpora used across the cough classification literature, and all baselines (EfficientNet-B2, ViT) were evaluated on the identical aggregated collection, allowing relative gains to be attributed to the proposed methods. Patient-level splits were enforced across the five folds to prevent leakage, though we did not report quantitative noise audits because the source datasets provide no inter-rater or self-report validation labels. We will add an explicit limitations paragraph discussing this issue and its potential effect on absolute performance numbers.","revision_made":"partial","referee_comment":"[§3] §3 (Dataset Aggregation): the four public datasets are combined without any reported quantitative audit of label noise rates, inter-rater agreement, or patient-level leakage checks. Because crowd-sourced corpora commonly exhibit self-report error rates above 10%, the 82.3% balanced accuracy, 0.817 macro-F1 and the 5.1-point active-frame ablation delta cannot be confidently attributed to architectural superiority rather than dataset artifacts; this assumption is load-bearing for all performance claims."},{"response":"We accept that error bars, confidence intervals, and formal significance tests were omitted. In the revision we will recompute all five-fold results with per-fold standard deviations and apply paired statistical tests (e.g., McNemar or Wilcoxon) against the baselines. The ablation table will be expanded to include more isolated controls (e.g., active-frame pooling added to the base model before versus after the other regularizers) so that the +5.1 point delta can be verified independently.","revision_made":"yes","referee_comment":"[§5] §5 (Results and Ablations): the five-fold CV metrics are reported without error bars, confidence intervals or statistical significance tests against the EfficientNet-B2 and ViT baselines. In addition, the ablation table combines WeightedRandomSampler, Balanced Mixup, contrastive loss and domain adaptation without exhaustive isolation controls, so the isolated 5.1-point contribution of active-frame pooling cannot be verified as independent of the other regularizers."},{"response":"We will supply the exact patient-ID lists for each of the five folds in a supplementary table and link to the full training configuration (learning-rate schedule, batch size, epochs, augmentation parameters). For gradient-reversal domain adaptation we followed the standard formulation of Ganin et al. with a fixed lambda of 1.0; the precise PyTorch implementation and hyper-parameter values will be documented in an expanded appendix or code release.","revision_made":"yes","referee_comment":"[§4] §4 (Training Protocol): exact patient-level train-test splits, full hyperparameter schedules and the precise implementation of gradient-reversal domain adaptation are not supplied. Without these details the claim that the listed techniques fully compensate for domain shift and imbalance remains unverifiable and limits assessment of whether the reported gains generalize beyond the aggregated collection."}],"tokens_in":1834,"tokens_out":721,"duration_ms":22207,"standing_objections":["Quantitative audit of label noise rates or inter-rater agreement for the four source datasets, as no such annotations exist in the original public releases."]},"desk_editor":{"model":"grok-4.3","letter":"The main takeaway is that fine-tuning Whisper-tiny with active-frame QKV attention pooling reaches 82.3% balanced accuracy and 0.817 macro-F1 on five respiratory classes from aggregated cough recordings, with the dual-encoder fusion version at 85.4%. The pooling limits attention to the first 200 tokens to keep the model from diluting short coughs across Whisper's long 30-second window, and the ablation credits it with the biggest single gain of 5.1 points.\n\nThe work handles the practical side reasonably. It combines four public datasets, applies WeightedRandomSampler, balanced mixup, supervised contrastive loss, FiLM conditioning, and gradient-reversal adaptation to manage imbalance and domain shift, then shows the model beats an ImageNet EfficientNet-B2 by 11 points and a scratch ViT by nearly 30. All classes clear 74% recall. These are concrete numbers on real data rather than toy setups.\n\nThe soft spot sits in the data foundation. Every result rests on the union of Coswara, CoughVID, Virufy, and West China Hospital recordings. The paper lists the usual imbalance and shift fixes but supplies no patient-level split details, no measured label noise rates, and no cross-source leakage audit. If crowd-sourced self-reports carry the typical 10%+ error, the reported deltas become harder to read as architectural wins. That assumption is load-bearing for the whole claim.\n\nThe paper is for applied researchers building smartphone respiratory tools or adapting Whisper to short audio events. The pooling method and the cross-attention fusion with OPERA-CT are worth testing. It has enough empirical grounding and an explicit ablation to deserve referee time, though the data validation section will need tightening.","headline":"The paper delivers a usable 82.3% balanced accuracy on five-class cough classification with a practical active-frame pooling trick for Whisper, but the aggregated public datasets leave label noise and leakage unaddressed.","tokens_in":2664,"tokens_out":440,"would_cite":false,"duration_ms":17623,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"A fine-tuned Whisper encoder with active-frame pooling classifies coughs into five respiratory conditions at 82.3 percent balanced accuracy.","keywords":["cough classification","respiratory disease","whisper encoder","active-frame pooling","multi-class classification","audio classification","domain adaptation","contrastive learning"],"falsifier":"A test on a new independent cough dataset with expert-verified labels would show balanced accuracy below 70 percent if the claim does not hold.","tokens_in":2862,"feed_emoji":"🩺","tokens_out":528,"duration_ms":15752,"temperature":0.7,"pith_summary":"The paper shows that cough recordings from consumer phones can be sorted into healthy, COVID-19, asthma, bronchitis, or pneumonia using the Whisper speech model as a backbone. It aggregates recordings from four public datasets and adapts Whisper by restricting attention to the active frames within each cough, which normally occupy only a small part of the model's longer input window. Balancing methods address the extreme class imbalance and differences across data sources. A reader would care if this points to a low-cost way to screen multiple respiratory diseases from a single smartphone recording without new hardware.","feed_headline":"Whisper encoder classifies coughs into five diseases at 82% accuracy","feed_subtitle":"Active-frame pooling prevents silence from diluting short cough signals in the model's input window.","key_machinery":"Active-frame QKV attention pooling, which restricts attention to the first 200 of 1500 encoder tokens to avoid silence dilution in short coughs.","core_discovery":"CoughSense fine-tunes the Whisper-tiny encoder and introduces active-frame QKV attention pooling to focus computation on the first 200 tokens out of 1500, avoiding dilution from silence in the 30-second input window. Combined with WeightedRandomSampler, Balanced Mixup, supervised contrastive loss, FiLM conditioning, and gradient-reversal domain adaptation, the model reaches 82.3 percent balanced accuracy on five-fold cross-validation with macro-F1 of 0.817 and AUC of 0.941. All classes exceed 74 percent recall. A dual-encoder version that fuses Whisper with OPERA-CT via cross-attention reaches 85.4 percent balanced accuracy. Ablations identify active-frame pooling as the largest single gain","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["Whisper-tiny achieves 82.3% balanced accuracy on five cough classes","Active-frame QKV attention pooling focuses on cough audio tokens","Dual-encoder model reaches 85.4% accuracy via Whisper and OPERA-CT","Five respiratory conditions identified from coughs using Whisper fine-tuning"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"Label quality is consistent across the four public datasets and the listed training techniques fully compensate for class imbalance and domain shift without introducing new biases or overfitting.","fun_headline_variants_meta":{"raw":{"variants":["Whisper-tiny achieves 82.3% balanced accuracy on five cough classes","Active-frame QKV attention pooling focuses on cough audio tokens","Dual-encoder model reaches 85.4% accuracy via Whisper and OPERA-CT","Five respiratory conditions identified from coughs using Whisper fine-tuning"]},"model":"grok-4.3","cost_usd":0.008441,"raw_usage":{"total_tokens":3936,"prompt_tokens":904,"num_sources_used":0,"completion_tokens":76,"cost_in_usd_ticks":84412000,"prompt_tokens_details":{"text_tokens":904,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":2956,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":904,"tokens_out":76,"duration_ms":21654,"temperature":1.0,"reasoning_tokens":2956,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-06-28T11:42:15.549351+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"A test on a new independent cough dataset with expert-verified labels would show balanced accuracy below 70 percent if the claim does not hold.","supporting_citations":[],"review_version":1}