CoughSense: Cough-Centered Sensing Framework

Updated 5 July 2026

CoughSense is a framework that acquires and analyzes cough signals using smartphones and wearables to detect, count, and cluster cough events.
It integrates engineered acoustic features, deep encoder representations, and active-frame pooling to map cough signals to respiratory disease labels with high accuracy.
The system addresses challenges such as sensor domain shifts, data quality issues, and privacy constraints through adaptive preprocessing and multimodal fusion.

CoughSense denotes a cough-centered sensing framework in which cough signals are acquired, segmented, represented, and mapped to clinically relevant outputs such as cough-event detections, cough-type groupings, respiratory abnormality scores, or disease labels. In recent arXiv literature, the term is used for a smartphone-centered point-of-care system that combines cough with temperature, airflow, and optional ECG/EMG (Belkacem et al., 2020), a smartwatch-based cough detection and cough-type clustering system (Jaiswal et al., 2024), and a five-class smartphone-oriented respiratory disease classifier built on a Whisper encoder backbone (Vincent, 2 Jun 2026). This suggests a recurring architecture: capture a cough-related signal under uncontrolled conditions, isolate cough-active regions, extract robust representations, and infer health status under class imbalance, domain shift, and deployment constraints.

1. Definition and task space

Within the literature, CoughSense spans several task formulations rather than a single fixed benchmark. At the lowest level, it appears as automatic cough detection and counting, where the output is the time and number of cough events (Drugman et al., 2019). At an intermediate level, it appears as cough-type analysis, including smartwatch-based clustering into four cough patterns interpreted as single, double, triple, and quadruple cough sequences (Jaiswal et al., 2024). At the highest level, it appears as cough-based disease screening or differential classification, including COVID-19 detection from cough-only audio (Chung et al., 2023), pediatric healthy-versus-pathology classification from multiple cough epochs (T et al., 2021), and five-class respiratory disease discrimination among healthy, COVID-19, asthma or respiratory condition, bronchitis, and pneumonia (Vincent, 2 Jun 2026).

The breadth of outputs is summarized below.

Instantiation	Inputs	Reported outputs
Portable point-of-care CoughSense (Belkacem et al., 2020)	Cough, breathing, speech, thermal camera, spirometer, optional ECG/EMG	Multi-class respiratory illness, wet vs dry characterization, cough intensity
Smartwatch CoughSense (Jaiswal et al., 2024)	On-wrist microphone	Cough vs no-cough, four cough clusters
C2C cough diagnosis system (Chung et al., 2023)	Cough recordings only	COVID-positive vs COVID-negative
Five-class CoughSense (Vincent, 2 Jun 2026)	Smartphone cough recordings, optional symptom vector	Healthy, COVID-19, asthma/respiratory condition, bronchitis, pneumonia

A frequent misconception is that cough sensing is synonymous with binary COVID-19 screening. That view is inconsistent with the cited work. The 2020 point-of-care proposal explicitly targets respiratory illnesses including COVID-19, bronchitis, tuberculosis, asthma, and pneumonia (Belkacem et al., 2020), while the 2026 CoughSense system is framed as moving “beyond binary COVID-19 detection” to five-class respiratory screening (Vincent, 2 Jun 2026).

2. Sensing modalities and acquisition settings

CoughSense systems are not limited to a single sensor class. The acoustic path remains dominant, but the literature includes air microphones, contact acoustic sensors, smartwatch microphones, earbud IMUs, and multimodal clinical peripherals. In a comparative study of six synchronous sensors—ECG, nasal thermistor, chest belt, accelerometer, contact microphone, and lapel audio microphone—the lapel audio microphone delivered the highest two-feature joint mutual information, the best frame-level Revised Classification Rate, and average event-level sensitivity and specificity of about 94.5% in audio-only cough detection (Drugman et al., 2019). A subsequent two-microphone study similarly found that the contact microphone conveyed little new relevant information compared to audio and that audio-only detection reached average sensitivity 94.7% and specificity 95.0% (Drugman et al., 2020).

Wearable acoustic sensing has also been implemented with body-coupled hardware. “DeepCough” used a chest-adhered piezoelectric contact acoustic sensor sampled at 44.1 kHz and then down-sampled to 16 kHz; its CNN-based detector achieved event-level sensitivity 95.1% and specificity 99.5% on 14 healthy volunteers (Amoh et al., 2015). Smartwatch CoughSense instead used a Samsung Galaxy Watch 4 microphone at 16 kHz, with the watch worn on the non-dominant wrist during controlled seated, standing, and walking blocks (Jaiswal et al., 2024). Earbud-based activation replaces continuous microphone processing with a lower-power IMU trigger: CoughTrigger used a 3-axis accelerometer at 50 Hz and activated a 16 kHz audio pipeline only after an IMU-positive event, achieving 0.77 AUC under leave-one-subject-out evaluation while preserving near-baseline battery life on Galaxy Buds2 (Zhang et al., 2021).

Body-coupled wearables introduce a separate domain-shift problem. BCoughBench evaluated five respiratory acoustic foundation models under five EBEN-simulated body-coupled sensor conditions and showed that mean AUROC declined from 0.785 on smartphone recordings to 0.689–0.723 under body-coupled capture, with the largest degradation under temple vibration pickup and the smallest under the soft in-ear condition (Sanap et al., 23 Jun 2026). This directly qualifies any assumption that smartphone-trained cough models transfer unchanged to wearable body-coupled sensors.

3. Preprocessing and acoustic representation

Preprocessing is central because cough recordings are typically short, sparse, and contaminated by silence, speech, motion, or background noise. C2C uses max-amplitude normalization followed by short-time-energy segmentation with a 22.5 ms window and 11.25 ms hop; frames exceeding 14.5 are flagged as cough starts, the nearest subsequent frame exceeding 0.1 is used for region delineation, and only cough-active frames are retained (Chung et al., 2023). In the technical summary, short-time energy is represented as

$E[n] = \sum_{m=0}^{M-1} (x[n+m] w[m])^2$

and serves as a cough-activity detector rather than a generic voice activity detector (Chung et al., 2023).

Feature representations fall into three broad categories. The first is engineered time–frequency features. Smartwatch CoughSense transforms each audio file into 20 MFCCs plus 20 Mel-spectrogram features, yielding 40 retained features per file (Jaiswal et al., 2024). The 2024 machine-hearing system for robust cough segmentation computes short-term descriptors separately in five fixed frequency bands— $[0,0.5)$ , $[0.5,1)$ , $[1,1.5)$ , $[1.5,2)$ , and $[2,5.5125]$ kHz—then reduces them to a 29-feature short-term subset and finally to a 58-dimensional long-term representation by taking the mean and standard deviation over 299 ms windows (Monge-Alvarez et al., 2024). Sound-Dr instead uses pretrained TRILL or FRILL embeddings, producing 2048-dimensional embeddings per second and then concatenating the temporal mean and standard deviation into a 4096-dimensional recording-level vector (Hoang et al., 2022).

The second category is event-adaptive signal decomposition. DeepCough uses Empirical Mode Decomposition, identifies cough-related energy in the 5th and 9th intrinsic mode functions, derives instantaneous amplitudes via the Hilbert transform, and then forms a three-channel CoughTensor from 33 MFCCs, 33 Mel bands, and 33 LPCS coefficients over 100 frames (Andreu-Perez et al., 2021). Independent-subspace-analysis-based cough counting likewise begins with a 44.1 kHz spectrogram, applies truncated SVD and ICA to produce independent time-activation functions, and detects cough-like peaks as high-variance events (Leamy et al., 2021).

The third category is learned encoder representations. C2C fine-tunes Wav2vec 2.0 on cough recordings for 2000 epochs and then freezes it before ECAPA-TDNN aggregation (Chung et al., 2023). The 2026 CoughSense system maps 80-band Whisper-format log-Mel spectrograms to 1500 encoder tokens, then restricts attention to the first 200 tokens to avoid “silence-dilution” because a typical 3-second cough occupies only about 150 of Whisper’s 30-second input tokens (Vincent, 2 Jun 2026).

Explainable spectral characterization remains relevant even in deep settings. A statistical analysis of cough sounds showed that spectral roll-off, spectral centroid, and spectral bandwidth are higher for cough than for speech; spectral flatness in cough sounds can rise to 0.22, spectral flux typically lies between 0.3 and 0.6, and the ZCR of most cough frames is between 0.05 and 0.4 (Vodnala et al., 2023). Those descriptors recur across later feature-engineered systems.

4. Model architectures and decision mechanisms

CoughSense architectures range from classical temporal models to modern foundation-model pipelines. Early wearable acoustic detection used a compact 2D CNN operating on $64 \times 16$ spectrogram patches; DeepCough employed two convolutional layers, two fully connected layers, and a two-way softmax (Amoh et al., 2015). A distinct classical line used Hidden Markov Models with five physiologically motivated states $A,B,C,D,E$ representing intra-cough stages and silence regimes; the multivariate three-band HMM reached AUC 0.920 for coughing versus non-coughing, outperforming a univariate energy HMM (Teyhouee et al., 2019).

Sequence modeling on handcrafted features remains competitive in some diagnostic settings. A pediatric pathology classifier used 42-dimensional MFCC+ $\Delta$ + $\Delta\Delta$ sequences, two stacked BiLSTM layers with 50 hidden units in each direction, and subject-level majority voting across multiple cough epochs; this design exceeded 84% event-level accuracy for healthy-versus-pathology screening and exceeded 91% subject-level accuracy after fusion (T et al., 2021). Tabular models also remain practical: in MultiSense-Pneumo, cough audio is transformed into MFCC and low-level descriptor summaries and classified by a LightGBM ensemble, producing a normalized cough risk signal for late fusion (Jayakody et al., 4 May 2026).

Representation-learning systems dominate current disease-screening work. C2C combines a Wav2vec 2.0 feature extractor, an ECAPA-TDNN backbone with hidden channels reduced to one-eighth of the vanilla architecture, and a two-layer sigmoid classifier trained with BCE (Chung et al., 2023). DeepCough uses four convolutional blocks over $[0,0.5)$ 0 tensors, followed by global average pooling and a softmax classifier (Andreu-Perez et al., 2021). The 2026 five-class CoughSense system uses Whisper-tiny as a pretrained backbone, active-frame QKV attention pooling over the first $[0,0.5)$ 1 tokens, FiLM symptom conditioning from a seven-dimensional symptom vector, Balanced Mixup with forced minority pairing, a supervised contrastive auxiliary loss, and gradient-reversal domain adaptation; an optional dual-encoder model further fuses Whisper with OPERA-CT through cross-attention (Vincent, 2 Jun 2026).

The attention operator used in the active-frame and cross-attention components is the standard scaled dot-product form,

$[0,0.5)$ 2

with four heads in the active-frame pooling block (Vincent, 2 Jun 2026). The technical significance is not merely architectural complexity: the ablation study shows that restricting pooling to the cough-active token region is the single largest contributor among the reported components.

5. Reported performance across tasks

Reported performance varies substantially because the literature evaluates different endpoints: event detection, cough counting, abnormality screening, COVID-19 screening, pathology classification, and five-class respiratory disease classification. Direct numerical comparison is therefore limited to systems addressing the same task.

System	Task	Reported result
Audio-only sensor relevance study (Drugman et al., 2019)	Event-level cough detection	Average sensitivity 94.39%, specificity 94.45%
Wearable DeepCough (Amoh et al., 2015)	Event-level cough detection	Sensitivity 95.1%, specificity 99.5%
Smartwatch CoughSense (Jaiswal et al., 2024)	Cough detection	Accuracy 0.9849 non-walking; 0.9823 walking
C2C (Chung et al., 2023)	COVID-19 detection from cough	ROC-AUC 0.7810
DeepCough (Andreu-Perez et al., 2021)	COVID-19 screening	AUC 98.80% ± 0.83%, sensitivity 96.43% ± 1.85%, specificity 96.20% ± 1.74%
Five-class CoughSense (Vincent, 2 Jun 2026)	Healthy/COVID-19/respiratory condition/bronchitis/pneumonia	Balanced accuracy 82.3% ± 1.8, macro-F1 0.817 ± 0.02, AUC 0.941 ± 0.01
Dual-encoder CoughSense (Vincent, 2 Jun 2026)	Same five-class task	Balanced accuracy 85.4% ± 1.3, AUC 0.958
Band-specific machine hearing system (Monge-Alvarez et al., 2024)	Robust cough segmentation	Sensitivity 92.71%, specificity 88.58%, AUC 90.69%

Several ablation results are especially diagnostic. In C2C, removing pre-processing collapses performance to ROC-AUC 0.5000, removing Wav2vec 2.0 yields ROC-AUC 0.6277, and removing augmentation yields ROC-AUC 0.6729, against 0.7810 for the full pipeline (Chung et al., 2023). In the same study, cough-only input outperforms deep-breath-only input (0.7810 versus 0.6749), and a cough-plus-breath fusion model converges to a learnable weight $[0,0.5)$ 3, indicating that cough dominates in that dataset (Chung et al., 2023). In the five-class CoughSense system, mean pooling over all 1500 Whisper tokens yields 73.1% balanced accuracy, active-frame cropping raises it to 78.2%, QKV attention pooling to 80.4%, and the full system to 82.3%; the active-frame pool alone contributes +5.1 points (Vincent, 2 Jun 2026).

Robustness patterns are equally informative. Smartwatch detection remains high under walking but drops in precision, recall, and F1 relative to non-walking, consistent with motion and friction noise (Jaiswal et al., 2024). BCoughBench shows that no evaluated foundation model meets the clinical sensitivity threshold Se@Sp95 $[0,0.5)$ 4 on most disease tasks under any body-coupled sensor, even when AUROC degradation is modest (Sanap et al., 23 Jun 2026). MultiSense-Pneumo’s cough pathway reaches overall accuracy 0.78 but only recall 0.39 on the abnormal class, making it a weak standalone screening branch and a better candidate for late fusion (Jayakody et al., 4 May 2026).

6. Deployment constraints, limitations, and future directions

CoughSense systems are usually framed as screening or monitoring tools rather than definitive diagnostic instruments. The 2020 portable point-of-care proposal explicitly emphasizes lab-free support rather than established clinical validation (Belkacem et al., 2020). DeepCough likewise describes its web tool as a primary screening instrument, not a definitive diagnosis system, despite strong cross-validated performance on 8,380 clinically validated samples (Andreu-Perez et al., 2021). Sound-Dr uses self-reported COVID-19 status and identifies the lack of physician diagnosis and PCR-based confirmation as a limitation requiring future clinical validation (Hoang et al., 2022).

Three limitations recur across the literature. The first is data quality and label quality. C2C does not report the number of cough recordings, class balance, devices, or sampling rate, and all reported metrics are from an 8% validation split because the test set was blind (Chung et al., 2023). The five-class CoughSense model harmonizes four public datasets, but bronchitis and pneumonia come only from a pediatric dataset while the majority of recordings are adult, and crowdsourced self-reports likely introduce label noise, especially for COVID-19 (Vincent, 2 Jun 2026). MultiSense-Pneumo uses “normal” versus “abnormal” cough proxy labels rather than pneumonia-confirmed cough labels, which directly limits its pneumonia specificity (Jayakody et al., 4 May 2026).

The second is domain shift across sensors and environments. Sound-Dr reports less train–test distribution shift than Coswara and COUGHVID, yet still treats robust real-world deployment as an open problem (Hoang et al., 2022). BCoughBench shows that body-coupled wearable conditions systematically reduce respiratory foundation-model reliability relative to smartphone recordings (Sanap et al., 23 Jun 2026). This suggests that on-device deployment is not merely a compression problem; it is also a sensor-domain adaptation problem.

The third is computational and privacy trade-off. Heavy encoders such as Whisper, Wav2vec 2.0, ECAPA-TDNN, or OPERA-CT are feasible on modern devices only with careful engineering, whereas compact CNNs, SVMs, LightGBM models, and IMU-triggered cascades are substantially lighter (Chung et al., 2023, Monge-Alvarez et al., 2024, Zhang et al., 2021). Several papers therefore recommend local preprocessing, on-device inference, or selective triggering to reduce exposure of personally sensitive cough audio (Belkacem et al., 2020, Jaiswal et al., 2024).

Future work in the literature is relatively consistent. The five-class CoughSense paper identifies adaptive active-frame selection, richer metadata, addition of tuberculosis, on-device quantization, and test-time adaptation as next steps (Vincent, 2 Jun 2026). BCoughBench calls for real body-coupled cough datasets and BC-aware model adaptation rather than zero-shot transfer from smartphone audio (Sanap et al., 23 Jun 2026). Smartwatch CoughSense proposes larger, more diverse cohorts, semi/self-supervised representation learning, domain adaptation, multi-sensor fusion, and on-device continual learning (Jaiswal et al., 2024). The 2020 point-of-care proposal and related multimodal work further imply that cough will remain most useful when integrated with structured symptoms, fever signals, airflow, speech, or imaging rather than treated as an isolated biomarker (Belkacem et al., 2020, Jayakody et al., 4 May 2026).

Taken together, the literature presents CoughSense not as a single algorithm but as a technically coherent field: cough-centered acquisition; event isolation by energy, EMD, ISA, or learned attention; representation learning through MFCCs, Mel features, pretrained audio encoders, or multimodal embeddings; and deployment-aware inference under real constraints of privacy, battery, calibration, and dataset shift. The strongest current results arise when the cough signal is treated as structured, sparse, and context-dependent rather than as an ordinary speech clip.