Automated Dysphagia Screening Using Noninvasive Neck Acoustic Sensing

Abstract: Pharyngeal health plays a vital role in essential human functions such as breathing, swallowing, and vocalization. Early detection of swallowing abnormalities, also known as dysphagia, is crucial for timely intervention. However, current diagnostic methods often rely on radiographic imaging or invasive procedures. In this study, we propose an automated framework for detecting dysphagia using portable and noninvasive acoustic sensing coupled with applied machine learning. By capturing subtle acoustic signals from the neck during swallowing tasks, we aim to identify patterns associated with abnormal physiological conditions. Our approach achieves promising test-time abnormality detection performance, with an AUC-ROC of 0.904 under 5 independent train-test splits. This work demonstrates the feasibility of using noninvasive acoustic sensing as a practical and scalable tool for pharyngeal health monitoring.

• Automated dysphagia screening detects abnormal swallowing with AUC-ROC of 0.904, utilizing domain-specific features on neck acoustics.
• A novel approach combines digital stethoscope recordings with machine-learning models offering swift, non-invasive, and low-cost diagnostic potential compared to current instruments.
• Machine learning identifies key factors such as age and gender (higher risk for males and older patients), alongside specific acoustic features like average frequency intensity, indicating practical directions for enhancement.

Motivation and clinical context

Dysphagia affects 10–20% of adults over age 50 and up to two-thirds of patients with Parkinson's disease, stroke, or head and neck cancer, with annual U.S. hospitalization costs estimated at $4.3–7.1 billion. Current gold-standard evaluations—FEES, VFSS, and high-resolution manometry—are invasive, radiation-exposing, practitioner-dependent, or costly. Uninstrumented clinical swallow evaluations used for bedside screening exhibit limited diagnostic sensitivity and specificity. The paper addresses this gap by pairing a digital stethoscope placed lateral to the thyroid cartilage with machine learning, collecting acoustic data concurrently with FEES so that every swallow is labeled against an instrumented ground truth.

Data collection and annotation

The cohort comprises 49 participants recruited from the UC San Diego Center for Airway, Voice and Swallowing who self-reported dysphagia symptoms. Each underwent standard FEES (8–10 trials across consistencies and bolus sizes) while a 3M Littmann Core Digital Stethoscope recorded neck acoustics in real time. Speech-language pathologists and fellowship-trained laryngologists rated each swallow on the penetration-aspiration scale (PAS), with PAS 1–2 treated as normal, 3–5 as penetration, and ≥6 as aspiration. The cohort is reasonably balanced by gender (25 female / 24 male), with 55.1% of participants exhibiting abnormal PAS scores; comorbidities include COPD (16.3%) and aspiration pneumonia history (22.5%).

After removing corrupt recordings and segmenting swallows via Librosa-based amplitude thresholding (tuned per file against visual FEES data), the dataset yields 617 individual swallow events from 392 audio recordings, with mean swallow duration of 0.64 s. This scale exceeds most prior proof-of-concept studies, which the authors note were typically restricted to single bolus consistencies and small samples.

Feature extraction and modeling

The authors compare three feature families: (i) domain-informed features motivated by prior acoustic analyses of swallowing—top-five FFT/STFT frequencies, average and median frequency over time, peak and average amplitude, area under the absolute waveform (composite trapezoidal integration), and swallow count; (ii) OpenSMILE feature sets; and (iii) embeddings from pretrained audio foundation models (AST, CLAP, OPERA). Participant age and gender are appended to all representations. Preliminary experiments identified OPERA as the strongest embedding baseline, and random forest classifiers (RFC) were selected over SVMs for consistency.

Main results

All evaluation uses patient-level train–test splits (five independent stratified splits), ensuring no patient appears in both training and test sets—an appropriate protocol for the intended clinical deployment on unseen patients.

Task	Method	AUC-ROC	AUC-PRC	Balanced Acc.
Severity (3-class)	OPERA	0.557 ± 0.159	0.434 ± 0.130	0.542 ± 0.047
Severity	OpenSMILE	0.583 ± 0.120	0.503 ± 0.145	0.606 ± 0.079
Severity	Domain-informed	0.611 ± 0.055	0.519 ± 0.061	0.659 ± 0.028
Abnormality (2-class)	OPERA	0.651 ± 0.176	0.718 ± 0.140	0.579 ± 0.080
Abnormality	OpenSMILE	0.778 ± 0.144	0.850 ± 0.094	0.665 ± 0.152
Abnormality	Domain-informed	0.904 ± 0.015	0.913 ± 0.075	0.755 ± 0.061
Abnormality	Domain-informed + OpenSMILE	0.804 ± 0.183	0.862 ± 0.081	0.710 ± 0.159

Two findings stand out. First, the domain-informed features achieve an AUC-ROC of 0.904 for binary abnormality detection, substantially outperforming both generic audio embeddings and OpenSMILE features; notably, concatenating OpenSMILE features degrades performance to 0.804, which the authors attribute to noisy, irrelevant acoustic characteristics. Second, three-class severity classification remains weak (best AUC-ROC 0.611), which the authors attribute to small per-class training samples rather than to a fundamental ceiling—though this remains unverified. A critical difference test confirms the domain-informed approach significantly outperforms all alternatives.

Segmentation and patient-level aggregation

To simulate deployment on continuous recordings containing multiple swallows, the authors evaluate automatic segmentation followed by per-patient aggregation (mean-risk, max-risk, mode-risk):

Segmentation	Mean-risk	Max-risk	Mode-risk
Sliding window (1 s, 50% overlap)	0.893 ± 0.103	0.856 ± 0.106	0.884 ± 0.104
Fixed parameters (IoU = 0.4775)	0.868 ± 0.142	0.942 ± 0.051	0.842 ± 0.141
Human-segmented (benchmark)	0.967 ± 0.054	0.918 ± 0.079	0.971 ± 0.041

Fixed-parameter segmentation achieves 65.8% sensitivity and 87.6% specificity at the segment level. Its max-risk result of 0.942 approaches the human-segmented benchmark, but the authors candidly flag that sliding-window segmentation outperforms it under mean- and mode-risk aggregation, suggesting the max-risk figure may be dataset-specific rather than a consistently superior configuration. Confusion matrices under human segmentation show minimal false negatives—a clinically consequential property, since false negatives risk clearing patients with unsafe swallows for oral diet and aspiration pneumonia.

Interpretability

SHAP analysis identifies age and gender as influential predictors (older age and male sex associated with higher risk, consistent with prior epidemiology), alongside signal features including average amplitude, frequency, area under the curve, and swallow count, with weaker swallows linked to dysphagia. This supports the design choice of fusing demographics with acoustic characteristics despite their marginal standalone contribution in preliminary experiments.

Limitations and open questions

The authors acknowledge that the dataset size remains moderate, limiting generalization across diverse populations and clinical conditions. Several specific questions remain open: whether the fixed-parameter segmentation's strong max-risk result generalizes beyond this cohort; whether multiclass severity classification can be improved with larger, more balanced per-class samples; and whether performance holds in uncontrolled settings such as the home, where background noise and unsupervised bolus administration would differ substantially from the FEES suite. The per-file manual tuning of segmentation parameters also raises a practical question about fully automated preprocessing robustness.

Conclusion

This work demonstrates that surface digital auscultation during FEES captures acoustic signatures with diagnostic value for oropharyngeal dysfunction, achieving an AUC-ROC of 0.904 for binary dysphagia detection on unseen patients using interpretable, domain-informed features. The concurrent FEES labeling protocol and multi-consistency bolus design strengthen external validity relative to prior single-bolus studies, while the modest severity-classification performance and moderate cohort size delineate the boundaries of the current evidence.