Sounding out Dysphagia: Using a Digital Stethoscope to Assess Audiometric Differences in Swallowing

Abstract

Objectives:

A clinical swallow evaluation (CSE) is a subjective exam with limited diagnostic accuracy. We hypothesize that audiometric data collected superficially can enhance the detection of swallowing dysfunction. This study examines audiometric differences between normal and dysfunctional swallows on Flexible Endoscopic Evaluation of Swallowing (FEES) using a digital stethoscope.

Methods:

In this single-center prospective study, adult patients undergoing FEES for dysphagia had a digital stethoscope placed on the left neck. Audiometric data were analyzed by consistency and compared to Penetration-Aspiration Scale (PAS) and Yale Pharyngeal Residue Severity Rating Scale (YPS) scores.

Results:

Forty-nine patients with a total of 617 individual swallow events were included. Twenty-two patients (45%) had a PAS score of 1–2, 20 (41%) had a PAS score of 3–5, and 7 (14%) had a score of PAS 6–8. Multivariable analysis revealed significant differences in swallow duration (F=4.65;p=0.01), peak amplitude (F=10.6;p<0.0001), area under the curve (F=6.13;p=0.002), and median frequency (F=8.52;p=0.0002). Patients with PAS ≥3 had prolonged swallow (5.4s versus 3.3s, p<0.0001), lower peak amplitudes (0.25dB versus 0.32dB; p<0.001), and a lower median frequency (169.4 Hz versus 214.8 Hz; p<0.0001) compared to those with PAS 1–2. Higher YPS scores (III-V) were associated with a longer swallow duration (5.5s versus 3.5s, p<0.0001) and lower median frequency (158.1 Hz versus 221.1 Hz; P=0.0005).

Conclusion:

Audiometric data reveals significant differences in swallow duration, amplitude, and frequency in patients with swallow dysfunction. These findings suggest that audiometric analysis may serve as a valuable adjunct to the standard CSE.

Level of Evidence:

Level 3

INTRODUCTION

Dysphagia is a significant public health concern, particularly within the swiftly growing aging population.¹ Estimates of its prevalence range from 10–20% in individuals over 50 but increase to as much as two-thirds in patient populations with Parkinson’s disease, stroke, head and neck cancer, and those residing in nursing homes or acute care facilities.² Beyond reducing quality of life,³ dysphagia significantly increases the risk of life-threatening complications⁴ and contributes substantially to morbidity and mortality. Hospital admissions for dysphagia last an average of 3.8 days longer than other inpatient admissions and are associated with a 1.7-fold greater likelihood of in-hospital mortality.⁵ Given its high prevalence and potential for severe complications, dysphagia is a major driver of healthcare costs, contributing an estimated $4.3 to $7.1 billion annually in additional hospitalization expenses.⁵

Patients admitted to the hospital at risk for aspiration are often evaluated using non-instrumented bedside screenings or clinical swallow evaluations (CSE), which rely primarily on subjective visual observation of swallowing, to determine if further evaluation is needed. However, these assessments have repeatedly demonstrated significant limitations, with studies reporting wide variability in sensitivity (43%–100%) and specificity (46%–100%), raising concerns about diagnostic accuracy.^6,7 Such inconsistencies contribute to missed or delayed diagnoses, ultimately increasing healthcare utilization, costs, and morbidity.^8,9 Delays in diagnosis reduce opportunities for timely intervention, including postural adjustments, modifications in food viscosity and volume, sensory enhancements, neuromuscular training, strengthening maneuvers, and surgical management.¹⁰ These challenges highlight the need for a more reliable, objective, and accessible non-invasive screening tool for dysphagia assessment.

Cervical auscultation, a technique historically used to assess aspiration risk, has declined in clinical practice due to poor interrater reliability and low sensitivity.^6,7,11 However, the emergence of digital stethoscopes presents an opportunity to standardize swallowing sound analysis using a familiar clinical tool. Digital auscultation enables the collection of swallow acoustic data, leveraging advances in data synthesis and signal processing to provide a more quantitative assessment of swallowing physiology.

Despite this potential, the ability to use digital stethoscope-acquired acoustic data to accurately predict aspiration risk remains underexplored. To that end, we collected audiometric data of swallowing during a Flexible Endoscopic Evaluation of Swallow (FEES) to investigate associations between penetration, aspiration, and acoustic signal changes. Herein we demonstrate clear associations between temporal, frequency, and energy-based acoustic features in patients with evidence of airway invasion, laying the foundation for future advanced modeling approaches to improve dysphagia risk assessment through acoustic analysis.

METHODS

Study Design and Setting

This single-center prospective study included adult patients evaluated at the University of California, San Diego’s Center for Airway, Voice & Swallowing (UCSD CAVS). Participants were included if they presented with a chief complaint of dysphagia. At the time of recruitment, all participants underwent a Flexible Endoscopic Evaluation of Swallowing (FEES). Patients were excluded if they deferred a FEES or if they were deemed too unsafe to undergo further evaluation of their swallow function. Approval for this study was obtained from the UCSD Institutional Review Board (#810212).

Data collection & FEES Protocol

Participants had a digital stethoscope (3M^™ Littman Core Digital Stethoscope) placed on the left neck, lateral to the thyroid cartilage and above the cricoid cartilage, during the FEES. The stethoscope was manually fixated by a trained researcher throughout the FEES. Care was taken to provide full contact of the stethoscope on the neck without excessive force or distortion of the overlying skin. This area of the neck was selected due to its high signal-to-noise ratio, low variance, and simplicity for an untrained clinician to find.¹² A 3M Littman Core Digital Stethoscope was used due to this model of digital stethoscope being one of the most advanced on the market and studies showing that older models of the Littman digital stethoscope displayed good sensitivity for picking up the sounds of swallowing.¹³

FEES was performed using a laryngeal flexible endoscope (VNL-J10, Pentax Corporation Tokyo, Japan). Video and audio recordings were performed throughout the FEES. A standardized FEES protocol was followed, which included: two trials of 5cc thin liquid, two trials of 10cc thin liquid, two tsp trials of purees (applesauce), two tsp trials of soft & bite-sized solid (diced peach), two bites of a regular solid (cracker), and ended with 2 trials of self-administered cup sips of thin liquid. All items were colored with blue food dye as per routine. The FEES was terminated early if the patient’s pulmonary safety was overtly compromised due to swallow dysfunction. All evaluations were scored by either a fellowship-trained laryngologist or a trained speech pathologist. Quantifiable data collected during each FEES included Murray Secretion Scale (MSS)¹⁴, Yale Pharyngeal Residue Severity Rating Scale (YPS)¹⁵, and a Penetration-Aspiration Scale scores.¹⁶

Data Processing, Segmentation, and Validation

Video and audio files were initially segmented based on trial consistency. We then developed a computational pipeline in Python that processed audio recordings that consisted of three main phases: preprocessing, segmentation, and feature extraction.

The raw audio files (wav files) were first loaded into the pipeline that was created using the Librosa library. These files were converted into numerical waveforms for analysis. To maintain consistency across trials, all files were processed using a uniform sampling rate of 22,050 Hz. We then segmented swallows through a multi-staged approach that identified distinct peaks on the time domain audio signal. First, we removed low-energy silent regions in the audio signal. Next, audio signals that fell outside of an accepted range were filtered out using amplitude as a cutoff filter with a cutoff amplitude 0 and 2.0, and closely adjacent peaks were merged. To eliminate erroneous detections due to brief non-swallow noises, short segments were manually removed by applying hard-coded thresholds (length_cutoff=0.01 s, allowed_min_amplitude=0.2, allowed_max_amplitude=0.85) that, through parameter tuning, produced the best segmentation results. Lastly, detailed visual representations of the audio data, including a waveform graph with segmented swallows and a spectrogram were created and evaluated by researchers to ensure erroneous sounds were not included (Figure 1). The remaining segmented audio files were then passed to the feature extraction phase, where temporal, energy, and frequency-based characteristics were analyzed.

Figure 1:

Figure 1A: example of a normal Swallow segmented by our code showing both duration and amplitude of the swallow. Figure 1B: Swallow segmentations of an abnormal swallow. Figure 1C: Spectrogram of frequencies in a normal swallow

The accuracy of swallow detection was validated through a multistage approach. First, visual confirmation of the waveform was completed for each swallow. Subsequently, we randomly selected 20% of the data and compared the audio files with the corresponding FEES videos. During this process: each swallowing event was manually timestamped in the FEES video. Audio files were then independently timestamped at the start and end of each swallow by an individual blinded to the video timestamps. The two sets of timestamps were then compared to confirm that the audio detected was a visualized swallow on the exam. Ultimately, our methodology achieved an accuracy of detecting 96.8% of real swallows identified on the video.

Feature Extraction

Once the individual swallow events were processed, we then extracted quantitative features to use for analysis. These features were chosen based on clinical judgment by experts who evaluate dysphagia regularly and prior studies that have completed preliminary analyses on audiometric features of swallow.¹⁷

The first features evaluated were temporal features, which included the total duration of the swallow per trial, the number of swallow attempts per trial, and the duration of each individual swallow attempt. The total duration of the swallow per trial was calculated based on the manual segmentation of the audio file. The number of swallow attempts per trial was calculated using $Len=\frac{N}{sr}$ where N is the number of samples in the audio file and sr is the sampling rate (samples per second). The duration of each swallow attempt was calculated using $L_i=\frac{t_{end}-t_{\mathit{\text{start}}}}{sr}$ where t_start is the sample index at the beginning of the swallow segment and t_end is the sample index at the end of the swallow. We attempted to evaluate the presence of cough with each bolus, however, the amplitude of the cough was similar to the amplitude of voicing and thus was eliminated in our preprocessing.

Next, we evaluated amplitude-based features. First, we evaluated peak amplitude to determine the maximum absolute signal intensity within a swallow segment. This was calculated using $A_{\mathit{\text{peak}}}=max\left(\right|a\left[t\right]|\forall t\in [t_{\mathit{\text{starti}}},t_{end}\left]\right)$ where a[t] is the amplitude at sample t, t_start is the sample index of the beginning of the swallow segment, and t_end is the sample index of the end of the swallow segment. Next, we used the average area under the curve (AUC) to evaluate the total signal energy over the swallow duration. This was calculated using the trapezoidal rule for numeric integration with $\mathit{\text{AvgAUC}}=\frac{\mathit{\text{trapz}}\left(\right|y\left|\right)}{N\cdot sr}$ where trapz(|y|) calculates the AUC of the absolute value of the swallow segment.

Lastly, we evaluated frequency-based features. For this analysis, each swallow segment was divided into small overlapping segments using a Short-term Fourier Transform (STFT). We then calculated mean frequencies using the formula $f_{avg}=\frac{1}{N}{\sum }_{i=1}^F{f}_{\mathit{\text{peak}},i}$ is the dominant frequency in the ith frame F is the number of frames in the swallow segment’s STFT. This was similarly completed for the median frequency of the audio file using the formula: $f_{med}=\mathit{\text{median}}(f_{\mathit{\text{peak}},1},f_{\mathit{\text{peak}},2},...,f_{\mathit{\text{peak}},F})$.

Statistical Analysis

Demographics included in our analysis were based on prior studies that showed evidence of dysfunctional swallows based on different demographic variables.^18–22 Our post hoc analysis looking at YPS scores was based on a prior study showing that higher vallecular residue scores were associated with a higher risk of aspiration pneumonia.²³ Fisher exact test was used to compare demographics and comorbidities between the two groups. ANOVA and ANCOVA were used to compare audiometric features between groups. All analyses accounted for consistency, BMI, and age based on prior studies showing that consistency, BMI, and age can affect the audiometric features of the swallow.^24,25 BMI was also included as a covariable due to concerns that BMI and neck thickness may affect the amplitude-based features of our analysis.^24,25 We also included any demographics variable that were significantly different between groups as a covariate.

RESULTS

Demographics

Forty-nine patients were evaluated with a total of 617 swallow events included in our analysis (Table 1). Of those 49 patients, 22 patients (45%) had a PAS score of 1–2, 20 patients (41%) had a PAS score of 3–5, and 7 patients (14%) had a PAS score of 6–8. Twenty-five patients (51%) had no or trace YPS scores (I-II) and 24 (49%) had mild to severe YPS scores (III-V).

Table 1:

Participant Demographics

	Total (n=49)	PAS 1–2 (n=22)	PAS 3–8 (n=27)
Number of Swallows	617	273	344
Age (SD)	66.9 (13.6)	62.4 (13.5)	70.7 (12.6)
Gender (M:F)	24:25	8:14	16:11
Race (%)
- White	34 (69.4)	12 (54.5)	22 (81.5)
- Hispanic	4 (8.2)	2 (9.1)	2 (7.4)
- Asian	3 (6.1)	1 (4.5)	2 (7.4)
- Other	8 (16.3)	6 (27.2)	2 (7.4)
BMI (SD)	26.2 (4.8)	26.7 (3.8)	25.8 (5.5)
Comorbidities (%)
- Prior aspiration pneumonia	10 (20.4)	1 (4.5)	9 (33.3)
- COPD	9 (18.3)	3 (13.6)	6 (22.2)
- Neurodegenerative Disease	5 (10.2)	0 (0)	5 (18.5)
- Prior Stroke	4 (8.2)	2 (9.1)	2 (7.4)
- History of Head and neck cancer	3 (6.1)	0 (0)	3 (11.1)

The cohort had a mean age of 66.9 years (standard deviation 13.6), with a near-equal gender distribution (49% male). The average BMI was 26.2 kg/m². Common comorbidities associated with dysphagia included a history of aspiration pneumonia (20.4%), chronic obstructive pulmonary disease (COPD) (18.3%), neurodegenerative disease (10.2%), prior stroke (8.2%), and head and neck cancer (6.1%). Comparing demographics and comorbidities between patients with normal swallowing function (PAS 1–2) and those with penetration or aspiration (PAS 3–8), significant differences were observed in age (p= 0.032) and the prevalence of aspiration pneumonia (p = 0.015). However, no significant differences were found in gender (p = 0.15), ethnicity (p = 0.15), BMI (p = 0.54), COPD (p = 0.49), history of head and neck cancer (p = 0.24), prior stroke (p = 1.0), or neurodegenerative disease (p = 0.056).

Temporal Features Based on Penetration Aspiration Scores

We first compared acoustic features between three groups: patients with a relatively normal swallow (PAS 1–2), patients with evidence of penetration (PAS 3–5), and patients with evidence of aspiration (PAS 6–8). When comparing swallow acoustics across these three groups while controlling for trial consistency, age, and BMI, significant differences were found between groups in total trial duration (F=4.65; p=0.01) and duration of each individual swallow (F=7.97; p=0.004). However, the number of swallows per trial did not differ between groups (F=0.16; p=0.85).

Further analysis comparing patients with normal (PAS 1–2) and abnormal swallowing (PAS 3–8) revealed that patients with abnormal PAS scores required significantly more time to effectively swallow any solid or liquid trial (5.4s vs. 3.3s, p<0.0001) (Table 2). Similar time differences were seen at each consistency but were not powered enough to achieve statistical significance. Similarly, individual swallow duration was significantly longer in the PAS 3–8 group (0.71s vs. 0.56s, p=0.002). The number of swallows per trial, however, did not differ significantly (1.46 in PAS 1–2 patients versus 1.37 in PAS 3–8 patients; p=0.59).

Table 2:

Acoustic Features by PAS Score

	Total (n=617)	PAS 1–2 (n=273)	PAS 3–8 (n=344)	P value
Total Trial Duration (s)	4.42 (5.5)	3.3 (3.2)	5.4 (6.7)	<0.0001
- Thin Liquid 5cc	3.7 (3.5)	3.3 (2.2)	4.3 (4.3)	0.322
- Thin Liquid 10cc	4.4(4.1)	3.7 (2.9)	4.9 (4.9)	0.664
- Puree	4.9 (5.1)	3.7 (3.4)	6.1 (6.2)	0.107
- Soft & Bite-sized	4.6 (6.8)	3.2 (3.2)	6.0 (9.0)	0.312
- Solid	4.8 (8.3)	2.1 (1.6)	7.3 (11.0)	0.077
- Self-administered cup sip thin liquid	4.3 (5.6)	3.3 (5.0)	5.0 (5.9)	0.411
Individual Swallow Duration (s)	0.64 (0.6)	0.56 (0.4)	0.71 (0.8)	0.002
Number of Swallows/Trial	1.41 (2.52)	1.46 (0.76)	1.36 (3.37)	0.588
Median Frequency (Hz)	189.5 (169.0)	214.8 (192.3)	169.4 (145.0)	<0.0001
Mean Frequency (Hz)	230.1 (136.2)	253.9 (158.7)	211.1 (111.8)	<0.0001
Total Area Under the Curve	0.014 (0.03)	0.021 (0.04)	0.0083 (0.01)	0.0005
Peak Amplitude	0.28 (0.30)	0.32 (0.34)	0.25 (0.26)	0.0024

Frequency and Amplitude Features Based on Penetration Aspiration Scores

When comparing again across 3 groups based on PAS scores, multivariable analysis showed significant differences in median frequency (F=8.52, p=0.0002), mean frequency (F=8.47, p=0.0002), area under the curve (F=6.13, p=0.002), and peak amplitude (F=10.66, p<0.0001).

Patients with abnormal PAS scores (PAS 3–8) exhibited a significantly lower median frequency (169.4 Hz versus 214.8 Hz, p < 0.0001) and mean frequency (211.2 Hz versus 253.9 Hz, p<0.0001). Similarly, patients with abnormal PAS scores had significantly lower acoustic energy with lower peak amplitudes (0.25 versus 0.32, p=0.002) and area under the curve (0.008 versus 0.02, p=0.0005).

Acoustic Features Between Patients Based on Yale Vallecular Residue Scores

Swallow acoustics were further analyzed based on YPS Scores, comparing patients with no to trace residue (scores I-II) to those with mild to severe residue (scores III-V) (Table 3). Patients with higher residue scores took significantly longer to effectively swallow any solid or liquid trial (5.5 seconds versus 3.5 seconds, p<0.0001). Similarly, individual swallow duration was prolonged in patients with abnormal residue scores (0.69 seconds versus 0.60 seconds, p=0.02). However, the number of swallows per trial did not differ significantly (1.45 in scores of I-II versus 1.36 in scores of III-V; p=0.20).

Table 3:

Acoustic Features by Yale Pharyngeal Severity Rating Scale Residue Score

	Yale 1–2 (n=307)	Yale 3–5 (n=310)	P value
Total Trial Duration (s)	3.5 (3.3)	5.5 (7.0)	<0.0001
- Thin Liquids 5cc	2.9 (2.0)	4.7 (4.5)
- Thin Liquids 10cc	3.9 (2.9)	4.9 (5.2)
- Purees	3.7 (3.4)	6.4 (6.5)
- Mixed Consistency	3.2 (3.2)	6.3 (9.4)
- Solids	3.0 (3.3)	7.0 (11.6)
- Self-selected liquids	3.6 (5.1)	4.8 (6.0)
Individual Swallow Duration (s)	0.60 (0.5)	0.69 (0.8)	0.02
Number of Swallows/Trial	1.45 (0.74)	1.36 (3.6)	0.20
Median Frequency (Hz)	221.1 (196.2)	158.1 (129.6)	0.0005
Mean Frequency (Hz)	254.7 (156.6)	205.7 (107.2)	0.004
Total Area Under the Curve	0.015 (0.03)	0.013 (0.03)	0.99
Peak Amplitude	0.29 (0.34)	0.27 (0.26)	0.011

Multivariable analysis revealed that patients with higher residue scores exhibited significantly lower median frequency (158.1 Hz versus 221.1 Hz, p=0.0005), mean frequency (205.7 Hz versus 254.7 Hz, p=0.004), and peak amplitude (0.27 versus 0.29, p=0.011). No significant difference was found in area under the curve (0.013 versus 0.015, p=0.99).

DISCUSSION

Despite its high prevalence, accurately predicting penetration and aspiration without endoscopic or fluoroscopic testing remains limited. Current bedside screening tools need improvement, emphasizing the need for additional research to elucidate objective factors that can predict swallow dysfunction. The imperative lies in developing novel tools to obtain objective data and redefining standardized bedside protocols given the cost and expertise required to perform instrumented examination and the sheer number of patients that need screening. Given this, we used acoustic data as an easily accessible tool for evaluating swallow dysfunction and determining associations between acoustic features and FEES outcomes.

In this cross-sectional study of over 600 swallows, we identified significant differences in temporal, frequency, and energy-related variables. Patients with penetration or aspiration on FEES exhibited longer swallow durations, lower frequencies, and reduced acoustic energy compared to those with relatively normal swallowing. Similarly, higher YPS Scores correlated with prolonged swallow times, lower frequencies, and lower peak amplitudes. The impact of the present study is two-fold: (1) this study highlights that patients with dysfunctional swallows produce distinct acoustic signatures potentially reflecting inefficient or weak swallowing mechanisms; (2) this study lays the groundwork for future machine learning models that utilize acoustic data to predict penetration or aspiration risk.

The concept of using a microphone to detect swallow sounds has been explored in previous studies. However, these studies have been limited by either being proof of concept designs with small sample sizes^26–29 or a lack of detailed acoustic analyses of each swallow.³⁰ For instance, Fukuike et al. developed a microphone capable of detecting swallow sounds and generating a visual analysis, but their study was small and relied on a significant manual component.³¹ Similarly, Khalifa et al. used an accelerometer and microphone to detect swallowing sounds with 95% accuracy through rigorous coding.³⁰ These studies served as the groundwork for this current study signifying that automatic detection of swallow was possible.

Our findings regarding prolonged swallow duration are novel within the context of using a digital stethoscope to obtain audiometric data in patients undergoing FEES. However, similar observations have been reported in other studies using different methodologies. For example, Morton et al. and Power et al. examined pharyngeal dwell time (PDT) in modified barium swallow studies (MBS) and found that longer PDTs correlated with an increased risk of aspiration.^{32 33} While it is well established that prolonged pharyngeal swallowing is linked to higher aspiration risk and reduced swallowing efficiency, few studies have described a method for measuring this without MBS. Our study is one of the first to use acoustic data alone to quantify these temporal differences. Although a few previous studies have identified prolonged swallow durations using microphones during MBS, their methodologies were limited, as they did not account for comorbidities such as age and obesity, which can influence acoustic signals.^17,34 By utilizing an accessible tool to identify temporal markers of swallow dysfunction, our study provides valuable objective data that can assist bedside clinicians in assessing aspiration risk and determining the need for further evaluation.

Similarly, our analysis of energy and frequency features provides novel insights into the objective acoustic characteristics of dysfunctional swallows. The few studies that have evaluated energy or frequency-based features have the same limitations as noted above with a lack of adjusting for comorbidities or other factors that can influence acoustic features. However, despite the lack of robust analysis in these prior studies, our results align with previous findings, demonstrating decreased amplitude and lower frequency in patients with penetration or aspiration.¹⁷ Given that amplitude is a marker for overall energy of the swallow, one possible explanation to explain this is that a reduced amplitude may indicate a weaker pharyngeal muscle contraction, leading to less forceful bolus propulsion and reduced muscular effort, though direct measurement of muscle strength was not conducted in this study. Additionally, patients with dysphagia often experience prolonged swallowing phases, which could hypothetically result in slower, more effortful swallows that generate less abrupt pressure changes, leading to lower peak amplitudes and frequencies. Lastly, effective airway protection during swallowing requires quick and forceful closure of the vocal folds and epiglottic inversion. In dysphagic patients, incomplete or sluggish airway closure may result in dampened vibrations, reducing both amplitude and frequency. However, all these potential physiological factors that may explain our findings are currently theoretical and unable to be clearly proven in the current study.

Understanding these acoustic differences in patients with penetration or aspiration is important for several reasons. First, acoustic data is readily accessible, acquired non-invasively, and can complement existing clinical screening tools. Second, there are currently few objective measures available for the evaluation of dysphagia. Even gold-standard assessments such as FEES and MBS have been questioned regarding their interrater reliability.^35,36 A standardized tool that provides objective numerical data could be particularly useful for training clinicians who do not regularly evaluate swallowing. Finally, this study serves as a proof of concept for future machine learning models that can leverage acoustic data to identify patients at risk of aspiration, facilitating earlier intervention and reducing the time to further evaluation.

The main limitation in this study is based on the number of participants and heterogeneity of the study population. While analyzing over 600 swallow events allowed us to conduct a robust analysis accounting for comorbidities, we were unable to further delineate differences between patients with PAS scores of 3–5 and those with PAS scores of 6–8 given the lower frequency of observed swallow pathology. However, the clear differences in acoustic features between normal and dysfunctional swallows suggest that future machine learning models may be able to refine these distinctions further and future studies will be aimed at delineating acoustic characteristics between patients with evidence of penetration and aspiration.

CONCLUSION

Audiometric analysis of swallowing reveals significant differences in temporal, frequency, and energy variables in patients with penetration or aspiration. These findings support the potential utility of acoustic data as an adjunct to standard CSE and provide a foundation for future machine learning models aimed at improving aspiration risk prediction.