Post-swallowing voice-based aspiration screening in dysphagia using a deep learning approach: insights from audio segmentation Authors: Jung-Min Kim, Min-Seop Kim, Sun-Young Choi, Hyun-Jin Kim, Ju Seok Ryu Published: 2026 ABSTRACT Dysphagia presents a serious risk of aspiration that requires continuous monitoring. This study introduces standardized 2 s voice segments for aspiration detection, derived from physiological constraints observed during clinical video-fluoroscopic swallowing study (VFSS) examinations. This prospective cohort study collected voice data from 198 participants aged ≥ 40 years. Of the 128 normal participants (healthy controls and VFSS-normal, penetration-aspiration scale (PAS) 1) and 70 with aspiration-risk (PAS ≥ 5) participants, the duration analysis revealed that the VFSS participants (overall mean: 2.22–2.48 s) produced significantly shorter phonations than healthy controls (6.19 ± 1.70 s) (p < 0.001), which justified the 2 s standardization. The aspiration-risk threshold was set at PAS ≥ 5 as material on the vocal cords significantly alters voice quality. Post-swallowing recordings were preprocessed into mel-spectrograms via short-time Fourier transform to extract time-frequency features. We developed three models based on MobileNetV3 (male, female, integrated) using the EfficientAT framework with 10-fold cross-validation. The integrated model demonstrated optimal performance with an area under the curve (AUC) of 0.8090 and 82.77% sensitivity. The male-specific model achieved an AUC of 0.7586 and 91.88% sensitivity, whereas the female model reached an AUC of 0.7376 and 61.11% sensitivity. This physiologically grounded approach shows promise for telemedicine-based aspiration screening.Trial registration: This study was approved by the institutional review board of Seoul National University Bundang Hospital (approval number B-2109-707-303) and registered in the ClinicalTrials.gov database (NCT05149976). Supplementary InformationThe online version contains supplementary material available at 10.1038/s41598-026-53618-w. Introduction Dysphagia is a disorder characterized by disturbances in the movement of food and liquids from the mouth to the stomach1. It can be caused by various conditions including stroke, neurological diseases, and head and neck cancer1. Aspiration pneumonia, the most serious complication of dysphagia, is a major health risk factor directly linked to high mortality rates owing to impaired safety of food intake2. In particular, this complication is a leading cause of death in the older population, and the resulting increase in medical costs and decrease in the quality of life are emerging as significant social issues in the context of globally increasing aging populations3. Therefore, early detection and continuous monitoring of dysphagia are essential for preventing aspiration pneumonia. In current clinical practice, the degree of aspiration is evaluated using the penetration-aspiration scale (PAS) through a video-fluoroscopic swallowing study (VFSS). The PAS score ranges from 1 to 8, with scores of 2–5 representing penetration, where food material enters the larynx but does not pass below the vocal cords, and scores of 6–8 indicating aspiration, where food material passes below the vocal cords, which is a direct risk factor for the development of aspiration pneumonia4. The most severe case is silent aspiration, corresponding to a PAS score of 8, in which patients continuously experience aspiration without a cough reflex, significantly increasing their risk of developing pneumonia5,6. More broadly, at PAS levels of 5 and above, food material either remains on the vocal cords without being ejected (PAS 5) or passes below them (PAS 6–8)4.The presence of a bolus is presumed to affect phonatory function and acoustic characteristics by altering vocal cord vibration patterns7–9. From the perspective of acoustic changes, however, PAS 3–4 involves material that either resides above the vocal cords or is ejected after contact, leaving no residual material on the vocal cords and thus has a limited impact on voice quality4,7–9. Based on this rationale, including PAS 5 — the level at which residues first begin to persist on the vocal cords — in the aspiration-risk group may enhance the sensitivity of voice-based dysphagia assessment for detecting aspiration-risk at its earliest manifestation4,7–10. Voice production and swallowing share common neuromuscular control mechanisms involving the larynx and oropharynx11. During the pharyngeal phase of swallowing, the larynx elevates vertically and moves anteriorly to facilitate safe bolus transit while protecting the airway12. This laryngeal elevation movement involves the intrinsic and extrinsic laryngeal muscles that are activated during voice production, particularly in pitch modulation11,13,14. Several previous studies have reported that vocal tasks requiring laryngeal elevation, such as maximum vocal pitch elevation, show significant correlations between hyoid and laryngeal excursions during these tasks and during swallowing, suggesting that such vocal tasks may serve as sensitive probes of swallowing-related biomechanics15,16. Sustained vowel phonation, such as/a:/, has clinical utility in detecting aspiration-related voice changes, as vowel production involves the laryngeal musculature associated with both voice production and airway protection during swallowing17–19. Furthermore, maximum phonation time (MPT) of sustained vowels has been used to indirectly evaluate laryngeal function and vocal cord vibration, with numerous studies demonstrating its relevance to swallowing assessment18–20. This suggests that when aspiration occurs, the mechanisms protecting the compromised larynx may be reflected as changes in voice characteristics, indicating that the patient’s voice can be utilized as a non-invasive indicator for dysphagia assessment13. Research on voice analysis for dysphagia assessment has been performed along three main paths, building on the concept of “wet voice,” a vocal characteristic indicating the presence of food or liquid near the vocal cords17,21,22. First, early studies attempted to assess wet voice through perceptual evaluation; however, several studies have reported that perceptually assessed wet voice has limited reliability, with low inter-rater agreement in clinical settings21,22. Second, researchers have attempted to conduct more objective and quantitative studies by applying statistical analysis and classical machine learning approaches to voice parameters23–31. Initial quantitative studies have focused on the statistical analysis of voice parameters, examining indicators such as frequency and amplitude perturbation and the noise-to-harmonics ratio (NHR), with one study achieving 100% sensitivity and 77.1% specificity in distinguishing dysphagia risk groups using combinations of relative average perturbation (RAP) and NHR23–26. Perceptual voice evaluation using the Grade, Roughness, Breathiness, Asthenia, Strain (GRBAS) scale showed low sensitivity (8–29%) but relatively high specificity (59–94%) for detecting dysphagia and penetration-aspiration, whereas acoustic measures including jitter, shimmer, and the signal-to-noise ratio (SNR) demonstrated higher sensitivity (8–76%) but lower specificity (22–77%)24.With advancing technology, classical machine learning algorithms such as Random Forest and eXtreme Gradient Boosting (XGBoost) have been applied to voice analyses for dysphagia assessment, achieving high diagnostic accuracy for dysphagia screening and aspiration risk classification27,28. Recently, in the field of voice disorders, cepstral peak prominence (CPP), derived through cepstral analysis, has been reported to demonstrate superior performance and reliability in voice quality assessment compared with traditional domain-based perturbation measures such as jitter and shimmer29. CPP quantifies the degree of prominence of the dominant peak in the cepstrum of the voice signal and serves as an indicator reflecting the clarity of the harmonic structure of the voice29,30. The American Speech-Language-Hearing Association (ASHA) has recommended CPP as a general measure for dysphonia assessment, proposing it as a preferred acoustic measure over conventional voice measures including jitter, shimmer, and harmonics-to-noise ratio31. CPP has the advantage of not relying on direct estimation of fundamental frequency, enabling reliable analysis even in voices with moderate-to-severe dysphonia29–31. Nevertheless, approaches based on individual acoustic parameters, including CPP, still require manual feature extraction and have limitations in real-time analysis of complex voice characteristics29–32. Third, current research has focused on applying deep learning models to mel-spectrograms or mel-frequency cepstral coefficients (MFCCs) derived from voice data to overcome the limitations of manual feature extraction and enable real-time monitoring33–35. In one such study, voice spectrogram analysis using short-time Fourier transform (STFT) and MFCCs was performed for various vocalization methods (phonation, cough, high-pitch phonation, and counting numbers). The multibranch model analyzing both the STFT and MFCCs achieved the highest performance with an area under the receiver operating characteristic (ROC) curve (AUC) of 0.887 ± 0.014 for counting numbers and a sensitivity of 0.882 ± 0.044 for high-pitch phonation33. Building upon these advancements, in deep learning-based approaches, our research team has been incrementally developing aspiration detection algorithms through the analysis of post-swallowing voice and a comparative analysis of pre- and post-swallowing voices34,35. Our previous study initially analyzed post-swallowing voices by converting variable-length voice data into mel-spectrograms and applying them to MobileNetV3, thereby achieving AUCs of 0.8361 for the integrated model, 0.8010 for the male model, and 0.7679 for the female model34. In a subsequent study analyzing pre- and post-swallowing voice changes using standardized 2 s segments, the models achieved AUCs of 0.7997 for the integrated model, 0.8319 for the male model, and 0.7331 for the female model35. During the development of these models, we identified several challenges in the original approach using variable-length voice samples, as outlined below. In standard clinical protocols in otolaryngology and speech-language pathology, sustained vowel/a:/phonation of approximately 3–5 s is commonly used for voice assessment31,36. However, voice recordings obtained from patients in clinical settings vary in length because sustained phonation capacity is influenced by aging and neurological impairments20,37. These variable-length recordings pose challenges for model training, as older patients and those with weakened oropharyngeal musculature often have difficulty sustaining prolonged phonation, and longer recordings tend to contain greater noise variability20,34,37. To address these issues, the present study measured the voice recording lengths in healthy controls and patients who underwent VFSS and adopted a standardized 2 s segmentation strategy as a preprocessing method20,31,34–37. Based on these findings, we hypothesized that standardized 2 s voice analysis offers several advantages: (1) inclusion of patients with limited sustained phonation capacity who cannot maintain vocalization for longer periods; (2) reduced noise variation through additional trimming, potentially improving signal quality and classification performance; (3) increased data volume by segmenting original-length recordings into 2 s segments, thereby expanding the dataset by approximately 1.5 to 3 times and potentially improving model performance; and (4) more consistent and rapid analysis applicable to patients’ daily lives and clinical settings. These non-invasive voice-based methods can be easily integrated into app- or web-based telemedicine platforms for the remote diagnosis and management of dysphagia, enabling continuous monitoring during the daily lives of patients23–35. The present study aimed to validate these hypotheses by developing and evaluating a deep learning-based algorithm for aspiration detection in dysphagia assessment using 2 s voice segments, with the ultimate goal of enhancing the clinical applicability of voice-based dysphagia assessment through an improved data sampling methodology. Methods Research design This single-center, prospective cohort study was conducted at Seoul National University Bundang Hospital between October 2021 and February 2023. All methods were carried out in accordance with the relevant guidelines and regulations, including the Declaration of Helsinki. This study was approved by the Institutional Review Board of Seoul National University Bundang Hospital (approval number: B-2109-707-303; initial approval date: September 1, 2021; actual study start date: October 7, 2021) and was registered with ClinicalTrials.gov (registration ID: NCT05149976; first release date: November 1, 2021). Written informed consent was obtained from all participants or their guardians prior to enrollment. This study only collected data on patients’ voice conditions without any intervention. Therefore, the study was conducted in accordance with the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) guidelines. We ensured the reliability and transparency of our research by adhering to these guidelines and ethical standards. The collected data were used in voice analysis research aimed at detecting aspiration in dysphagia assessment. Study participants Study participants were recruited based on informed consent. Guardians provided consent for those unable to do so owing to age or health conditions. Experienced occupational therapists and clinical dietitians provided a comprehensive explanation of the study. Participants who underwent VFSS between October 2021 and December 2022 for dysphagia-related symptoms were recruited. From January to February 2023, recruitment was expanded to include healthy controls through institutional notices and online channels, as obtaining sufficient normal samples from clinical settings alone was challenging. This expansion aimed to improve the performance and applicability of future diagnostic tools for dysphagia. Two experienced clinicians made the final selection based on age, comorbidities, swallowing-related questionnaires, and VFSS results following specific inclusion and exclusion criteria. To ensure the reliability of the VFSS results, two clinicians independently interpreted the findings. Inter-rater reliability was assessed, resulting in a Cohen’s kappa coefficient of 0.87, indicating a high level of agreement between the evaluators. The main inclusion criteria for the VFSS test and healthy control participants were the ability to vocalize sustained/a:/for approximately 5 s, the ability to follow the instructions of the researchers, and consent to participate. Additionally, VFSS test participants were required to have undergone a VFSS examination because of dysphagia-related signs or symptoms, whereas healthy controls were required to have none. The main exclusion criteria were the inability to follow the instructions of the researchers, inability to vocalize sustained/a:/, serious voice- or speech-related disorders (e.g., vocal cord muscle tension dysphonia, vocal cord paralysis, or vocal nodules), presence of tracheostomy, and inability to participate in the study based on the judgment of the researchers34,35,38. Exclusion of participants with voice and speech disorders was conducted by reviewing the formal diagnoses documented in their medical records. Initially, 323 participants were enrolled, comprising 159 healthy controls and 164 participants who underwent VFSS examination due to signs and symptoms of dysphagia. Within the VFSS group, 53 individuals were classified as normal (PAS 1), 38 as mild-to-moderate dysphagia (penetration group, PAS 2–4; 27 males and 11 females), and 73 as having an aspiration-risk (PAS ≥ 5). The PAS 2–4 group was excluded from the final analysis to optimize model performance. Due to its limited sample size and the ambiguous acoustic characteristics inherent in the intermediate range, this group could introduce technical noise, potentially hindering the model’s ability to learn distinct features and establish clear decision boundaries between normal and aspiration-risk groups. The VFSS group consisted of patients referred to the hospital for assessment based on clinical indications of dysphagia. In contrast, healthy controls did not undergo VFSS but were confirmed to have no symptoms of dysphagia through interviews conducted by clinicians and the research team. Based on previous studies, major comorbidities were categorized into six groups: neuromuscular disorders, gastrointestinal and oropharyngeal disorders, pulmonary disorders, other malignancies, age-related conditions, and no comorbid conditions34,39–41. The aspiration-risk group was significantly older; therefore, all participants younger than 40 years were excluded to maintain age homogeneity between the groups. After applying the exclusion criteria, including age < 40 years and poor recording quality, the final study population comprised 128 normal participants (healthy controls and VFSS-normal [PAS 1]) and 70 participants in the aspiration-risk group (PAS ≥ 5) (see Fig. 1 for a detailed participant flowchart). The sample size was determined by the number of eligible participants who met the inclusion and exclusion criteria and had analyzable voice recordings during the study period; no a priori sample size calculation was performed because this was an exploratory model-development study. Fig. 1Flowchart of the study population. Flowchart of the study population. Data collection methods VFSS test participants consumed 3 mL of water, along with plain yogurt, small fluid, large fluid, semi-blended diet, and level 3 thickened fluid, whereas healthy controls only consumed 3 mL of water. Since healthy controls did not undergo VFSS, it was not feasible to provide them with barium-mixed boluses. Furthermore, as the primary objective of this study was to detect voice characteristics associated with aspiration risk rather than to compare aspiration severity across different bolus viscosities—and considering that aspiration frequently occurs even with low-viscosity liquids like water in patients with dysphagia—the healthy controls performed voice recordings only after swallowing plain water42,43. For VFSS test participants, sustained /a:/ vocalization was recorded for approximately 5 s immediately after each swallow during the VFSS examination to capture potential dysphagia-related voice changes. Healthy controls were also recorded under the same vocal task conditions after water intake. All recordings were performed in a noise-isolated environment. Voice recordings of VFSS participants were performed using a Sony ICD-TX660 recorder (bit rate: 64 kbps, bit resolution: 16 bit, sampling frequency: 44.1 kHz, microphone bandwidth: 95–20,000 Hz, and recording mode: stereo) attached to the participants’ upper sleeve, whereas healthy controls used their smartphones’ recording function (Samsung or iPhone) positioned in front of their mouths or on a table. To verify the potential bias due to recording environments, cosine similarity was measured after converting the same individual’s recordings to mel-spectrograms. Compared with the Sony recorder, similarity was > 0.85 in all cases (Sony recorder upper sleeve position vs. Samsung smartphone, upper sleeve position: 0.8683, table: 0.8770, and in front of the mouth: 0.8873; iPhone, upper sleeve position: 0.9531, table: 0.9618, and in front of the mouth: 0.9618), confirming that bias due to the recording position or device was minimized. The aspiration status of VFSS participants was determined based on PAS scores derived from VFSS imaging. Notably, since both the aspiration-risk group (PAS ≥ 5) and the VFSS-normal group (PAS 1) were recorded using the same Sony recorder and protocol, any discrepancy in recording devices existed only between VFSS participants and healthy controls. This further ensured that the observed acoustic differences between PAS 1 and PAS ≥ 5 were not attributable to device-related variations. Following this verification, the duration of each original voice recording was measured to evaluate the rationale for adopting 2 s standardized segments. Data preprocessing methods Audio data cleaning and data format standardization During the data collection process, voice recordings were gathered from both the normal and aspiration-risk groups. Preprocessing involved three main stages: first, manual trimming of background voices and machine noise at the beginning and end of recordings. Second, unlike previous studies that typically used full-length recordings, we implemented an additional refinement step in which a Python script was used to segment each audio file into standardized 2 s units, followed by manual refinement. Third, all stereo recordings were split into separate left and right mono channels (mp3, 64 kbps), ensuring that acoustic information from both channels was utilized without data loss. The detailed number of voice files at each preprocessing stage is listed in Table 1. Analysis of the sample distribution after 2 s segmentation revealed distinct imbalance patterns based on sex. In the female group, the number of segments in the normal group was higher than that in the aspiration-risk group (553 vs. 134). Conversely, in the male group, the aspiration-risk group produced approximately 1.85 times more segments than the normal group (444 vs. 240). These discrepancies in data imbalance were primarily attributed to the sustained phonation duration characteristics of each group. To mitigate potential model bias arising from imbalances, specific oversampling strategies were implemented, as detailed in the ‘Data oversampling process to minimize data imbalance between groups’ section. Table 1Number of data points in each stage.FormatsSexNormalAspiration-Risk Healthy control VFSS-normal Total normal † Full-lengthRecordingsOriginal format(stereo + mono)Male382664147Female1073914646Total(Male + Female)14565210193Mono format(64 kbps, mp3)Male425294294Female1157819392Total(Male + Female)1571302873862-secondSegmentsOriginal format(stereo + mono)Male14143184222Female3956145667Total(Male + Female)536104640289Mono format(64 kbps, mp3) ††Male15486240444Female431122553134Total(Male + Female)585208793578† The normal group includes both healthy controls (who did not undergo VFSS) and VFSS-normal (PAS 1) participants.†† All audio data were standardized to mono channels prior to analysis. To minimize potential acoustic loss during 2 s segmentation, a manual refinement process was performed. Table 1Number of data points in each stage.FormatsSexNormalAspiration-Risk Healthy control VFSS-normal Total normal † Full-lengthRecordingsOriginal format(stereo + mono)Male382664147Female1073914646Total(Male + Female)14565210193Mono format(64 kbps, mp3)Male425294294Female1157819392Total(Male + Female)1571302873862-secondSegmentsOriginal format(stereo + mono)Male14143184222Female3956145667Total(Male + Female)536104640289Mono format(64 kbps, mp3) ††Male15486240444Female431122553134Total(Male + Female)585208793578† The normal group includes both healthy controls (who did not undergo VFSS) and VFSS-normal (PAS 1) participants.†† All audio data were standardized to mono channels prior to analysis. To minimize potential acoustic loss during 2 s segmentation, a manual refinement process was performed. Number of data points in each stage. Full-length Recordings Original format (stereo + mono) Total (Male + Female) Mono format (64 kbps, mp3) Total (Male + Female) 2-second Segments Original format (stereo + mono) Total (Male + Female) Mono format (64 kbps, mp3) †† Total (Male + Female) † The normal group includes both healthy controls (who did not undergo VFSS) and VFSS-normal (PAS 1) participants. †† All audio data were standardized to mono channels prior to analysis. To minimize potential acoustic loss during 2 s segmentation, a manual refinement process was performed. Design of the training and test datasets for 10-fold cross-validation In this study, separate datasets for males and females were constructed to evaluate the model performance based on sex. Each sex-specific dataset was divided into 10 segments for 10-fold cross-validation. To minimize bias owing to individual differences, data from the same participant were placed in the same fold. To build an integrated sex model, combined datasets of male and female data were also constructed in the same manner, ensuring the inclusion of voice data from both sexes while considering the sex ratio in each group. This approach allowed a direct performance comparison between the sex-specific and integrated sex models. The datasets were stored in Hierarchical Data Format version 5 (HDF5), with 10 train–test dataset pairs created for each fold combination. The HDF5 file structure was implemented with reference to the create_h5pymp3_dataset.py code from PaSST: Efficient Training of Audio Transformers with Patchout (Apache-2.0 license)44,45. Each HDF5 file was hierarchically structured as follows: (1) audio filename: information for identifying individual participants, (2) audio content (mp3 data): the recorded speech information in the mp3 format, and (3) label: binary classification information indicating the normal or aspiration-risk status. This structured dataset enabled efficient data access and model training and was designed to be consistent with our previous study34. Data oversampling process to minimize data imbalance between groups To address potential data imbalance, we applied a differentiated approach based on sex. For the male model, both the number of participants and 2 s data samples showed a relatively balanced distribution between the normal and aspiration-risk groups (Table 1); therefore, no additional oversampling was performed. In contrast, the female model showed a severe imbalance in both the number of participants and data samples, indicating insufficient learning data for patients with aspiration symptoms. Therefore, random oversampling was applied to data from the aspiration-risk group during the training of the female model. For the model integrating both male and female data, we oversampled the female aspiration-risk group data to match the level of the female normal group data during the HDF5 file generation stage. Conversion to mel-spectrogram For /a:/ voice data preprocessing, we converted mp3 data in the HDF5 format to mel-spectrograms using the preprocessing code from the Efficient Pre-Trained CNNs for Audio Pattern Recognition (Audio Tagging) (EfficientAT, MIT License) model46,47. The audio data were mostly set to a 2 s length to reduce zero padding. The mel-spectrogram conversion process proceeded in the order of pre-emphasis filter, STFT, power magnitude computation, and mel frequency filter bank. Referring to previous post-swallowing research results, we did not perform data augmentation34. The preprocessing parameters were set as follows: 128 mels, a 32,000 sampling rate, hop size of 320 (10 ms interval), window length of 640 (20 ms interval), and 640 for the number of fast Fourier transforms. Model training methods In this study, we focused on enhancing model performance by adopting 2 s preprocessed data compared to the approach used in our previous research34. By refining the preprocessing through 2 s segmentation, we aimed to improve overall model reliability. Notably, all models in this study were newly trained from scratch using the refined 2 s dataset rather than reusing models or weights from our prior work. The model was trained using the EfficientAT framework, including pre-trained models (mn40_as and mn30_as) and non-pre-trained models (mn4.0 and mn3.0)34,46,47.The pre-trained models (mn40_as and mn30_as) are standard, open-access models pre-trained on AudioSet—a large-scale general audio dataset provided by this framework. This ensures that the transfer learning process was entirely independent of the previous clinical data46,47. EfficientAT was selected for this study owing to its provision of pre-trained models trained on large-scale audio data, which facilitates transfer learning. The primary goal of this study was to develop a model capable of detecting aspiration in mobile and medical device environments using the MobileNetV3-based model. Here, “mn” in the model names referred to MobileNet, while “40”, “30”, “4.0”, and “3.0” denoted the width multiplier. The width multiplier used in lightweight neural networks such as MobileNet adjusts the number of filters in each layer, helping to scale the size of the model and control the computational load and number of parameters. Therefore, the same width multipliers used in EfficientAT, that is, mn40_as and mn30_as, were applied to the non-pre-trained models46,47. The head type of the model was implemented using a multilayer perceptron (MLP), with BatchNorm2d applied for normalization, and the data processing was set with 12 workers. To evaluate all models from the same starting point, we tracked the changes in the loss function and performance metrics using the Weights & Biases Tool, extracting the performance of each of the 10 folds at the most stable point, which was at 150 epochs. The learning rate was managed using the LambdaLR learning-rate scheduler, starting at an initial setting of 5.00 × 10− 5, maintained until the 100th epoch, then decreasing linearly from epochs 100 to 105, and finally reduced to 5.00 × 10− 7. Adam was used as the optimizer with a weight decay of 0.01 applied to prevent overfitting and to enhance generalization performance. The batch size was set to 64 for the training set. Figure 2 shows a flow diagram of voice data preprocessing and model development. Fig. 2Flow diagram of voice data preprocessing and model development. Flow diagram of voice data preprocessing and model development. Statistical analysis and model evaluation Statistical analysis was performed to examine demographic differences between the groups (normal vs. aspiration-risk groups). Owing to the different sex distributions between the groups and sex-influenced voice characteristics, separate models were trained for each sex. Participants younger than 40 years were excluded to minimize age bias, and comorbidities were categorized into six groups34,39–41. For statistical analysis, the Mann–Whitney U test was used for continuous variables, and the chi-square test was used for categorical variables, with the statistical significance set at p < 0.05. To assess the validity of standardizing voice segments to a 2 s duration in this study, phonation duration was measured using original-length data with noise reduction across the three groups described in the Study Participants section: healthy controls, VFSS-normal (PAS 1), and VFSS-aspiration-risk (PAS ≥ 5). Additionally, the distribution of phonation durations was examined by categorizing them into three intervals: <2 s, 2–5 s, and > 5 s. Prior to statistical testing, the Shapiro–Wilk test was performed to assess normality of phonation duration, revealing non-normal distributions in most groups (p < 0.05). Accordingly, the Kruskal–Wallis test was conducted to compare continuous phonation duration data across the three groups, followed by pairwise comparisons using the Mann–Whitney U test with Bonferroni correction. Categorical duration data (< 2 s, 2–5 s, > 5 s) were compared across groups using the chi-square test (Bonferroni-corrected significance threshold: p < 0.0167). The model performance was evaluated in terms of AUC (ability to distinguish between normal and aspiration-risk), accuracy (overall correct prediction rate), sensitivity (ability to correctly identify aspiration-risk cases), specificity (ability to correctly identify normal cases), positive predictive value (PPV, precision; probability that predicted aspiration-risk cases were actually aspiration-risk), negative predictive value (NPV; probability that predicted normal cases were actually normal), F1-score (harmonic mean of precision and sensitivity), loss, train accuracy, and train loss. All metrics were calculated as the mean, 95% confidence interval (CI), and maximum values across a 10-fold cross-validation. The classification threshold was set to 0.5 for normal probability, considering both the data imbalance in female samples and the importance of aspiration-risk detection. The loss function was calculated using binary cross-entropy with logits, considering that this is a binary classification problem. The performance metrics were computed using Python’s metrics module. Specifically, true positive (correct aspiration-risk predictions), true negative (correct normal predictions), false positive (incorrect aspiration-risk predictions), and false negative (incorrect normal predictions) were derived from the confusion matrix to calculate specificity, PPV, and NPV. All statistical analysis and model training were conducted using Python on Colab Pro+ (GPU A100) software between March and April 2024. Results The baseline characteristics of the study participants are presented in Table 2. The aspiration-risk group showed a significantly higher proportion of males (74.29% vs. 32.03%, p < 0.001) and a higher mean age (72.30 ± 12.03 vs. 61.16 ± 13.00 y, p < 0.001) than the normal group. Even after excluding participants under the age 40 y, significant differences in age and comorbidity distributions persisted between the two groups (p < 0.001). Sex-specific analysis revealed that the mean age of the aspiration-risk group was significantly higher for both males and females (p = 0.001). Regarding comorbidities, a significant difference was observed in males (p = 0.002), whereas no significant difference was found in females (p = 0.140). Table 2Characteristics of the study population.Characteristics†Normal (n = 128)Aspiration-risk (n = 70)p-value†† Sex, n (%) Male41 (32.03%)52 (74.29%)< 0.001(χ2 = 30.76, df = 1)Female87 (67.97%)18 (25.71%) Overall Distribution (Male + Female) Age, years (Mean ± SD)61.16 ± 13.0072.30 ± 12.03< 0.001Age distribution, n (%) 40–4932 (25.00%)5 (7.14%)< 0.001(χ2 = 31.98, df = 4)50–5938 (29.69%)7 (10.00%)60–6930 (23.44%)18 (25.71%)70–7912 (9.38%)16 (22.86%)80+16 (12.50%)24 (34.29%)Comorbidities, n (%) Neuromuscular disorders17 (13.28%)18 (25.71%)< 0.001(χ2 = 36.10, df = 5)Gastrointestinal andoropharyngeal disorders3 (2.34%)12 (17.14%)Pulmonary disorders4 (3.12%)9 (12.86%)Other Malignancies7 (5.47%)3 (4.29%)Age-related conditions12 (9.38%)8 (11.43%)No comorbid conditions85 (66.41%)20 (28.57%) Male Distribution Age, years (Mean ± SD)63.27 ± 13.5772.25 ± 11.680.001Age distribution, n (%) 40–497 (17.07%)4 (7.69%)< 0.001(χ2 = 17.53, df = 4)50–5914 (34.15%)3 (5.77%)60–6910 (24.39%)16 (30.77%)70–794 (9.76%)12 (23.08%)80+6 (14.63%)17 (32.69%)Comorbidities, n (%) Neuromuscular disorders5 (12.20%)11 (21.15%)0.002(χ2 = 18.54, df = 5)Gastrointestinal andoropharyngeal disorders1 (2.44%)12 (23.08%)Pulmonary disorders2 (4.88%)8 (15.38%)Other Malignancies2 (4.88%)2 (3.85%)Age-related conditions5 (12.20%)6 (11.54%)No comorbid conditions26 (63.41%)13 (25.00%) Female Distribution Age, years (Mean ± SD)60.16 ± 12.6672.44 ± 13.340.001Age distribution, n (%) 40–4925 (28.74%)1 (5.56%)< 0.001(χ2 = 13.53, df = 4)50–5924 (27.59%)4 (22.22%)60–6920 (22.99%)2 (11.11%)70–798 (9.20%)4 (22.22%)80+10 (11.49%)7 (38.89%)Comorbidities, n (%) Neuromuscular disorders12 (13.79%)7 (38.89%)0.140(χ2 = 8.31, df = 5)Gastrointestinal andoropharyngeal disorders2 (2.30%)0 (0.00%)Pulmonary disorders2 (2.30%)1 (5.56%)Other Malignancies5 (5.75%)1 (5.56%)Age-related conditions7 (8.05%)2 (11.11%)No comorbid conditions59 (67.82%)7 (38.89%)† Models were trained separately by sex due to voice differences. Participants aged under 40 y and those with voice or speech-related disorders (verified through medical records) were excluded to minimize potential bias.†† Categorical variables (sex, age intervals, comorbidities) are presented as n (%) and analyzed using chi-square test, continuous variable (age) as mean±standard deviation (SD) and analyzed using Mann–Whitney U test. Table 2Characteristics of the study population.Characteristics†Normal (n = 128)Aspiration-risk (n = 70)p-value†† Sex, n (%) Male41 (32.03%)52 (74.29%)< 0.001(χ2 = 30.76, df = 1)Female87 (67.97%)18 (25.71%) Overall Distribution (Male + Female) Age, years (Mean ± SD)61.16 ± 13.0072.30 ± 12.03< 0.001Age distribution, n (%) 40–4932 (25.00%)5 (7.14%)< 0.001(χ2 = 31.98, df = 4)50–5938 (29.69%)7 (10.00%)60–6930 (23.44%)18 (25.71%)70–7912 (9.38%)16 (22.86%)80+16 (12.50%)24 (34.29%)Comorbidities, n (%) Neuromuscular disorders17 (13.28%)18 (25.71%)< 0.001(χ2 = 36.10, df = 5)Gastrointestinal andoropharyngeal disorders3 (2.34%)12 (17.14%)Pulmonary disorders4 (3.12%)9 (12.86%)Other Malignancies7 (5.47%)3 (4.29%)Age-related conditions12 (9.38%)8 (11.43%)No comorbid conditions85 (66.41%)20 (28.57%) Male Distribution Age, years (Mean ± SD)63.27 ± 13.5772.25 ± 11.680.001Age distribution, n (%) 40–497 (17.07%)4 (7.69%)< 0.001(χ2 = 17.53, df = 4)50–5914 (34.15%)3 (5.77%)60–6910 (24.39%)16 (30.77%)70–794 (9.76%)12 (23.08%)80+6 (14.63%)17 (32.69%)Comorbidities, n (%) Neuromuscular disorders5 (12.20%)11 (21.15%)0.002(χ2 = 18.54, df = 5)Gastrointestinal andoropharyngeal disorders1 (2.44%)12 (23.08%)Pulmonary disorders2 (4.88%)8 (15.38%)Other Malignancies2 (4.88%)2 (3.85%)Age-related conditions5 (12.20%)6 (11.54%)No comorbid conditions26 (63.41%)13 (25.00%) Female Distribution Age, years (Mean ± SD)60.16 ± 12.6672.44 ± 13.340.001Age distribution, n (%) 40–4925 (28.74%)1 (5.56%)< 0.001(χ2 = 13.53, df = 4)50–5924 (27.59%)4 (22.22%)60–6920 (22.99%)2 (11.11%)70–798 (9.20%)4 (22.22%)80+10 (11.49%)7 (38.89%)Comorbidities, n (%) Neuromuscular disorders12 (13.79%)7 (38.89%)0.140(χ2 = 8.31, df = 5)Gastrointestinal andoropharyngeal disorders2 (2.30%)0 (0.00%)Pulmonary disorders2 (2.30%)1 (5.56%)Other Malignancies5 (5.75%)1 (5.56%)Age-related conditions7 (8.05%)2 (11.11%)No comorbid conditions59 (67.82%)7 (38.89%)† Models were trained separately by sex due to voice differences. Participants aged under 40 y and those with voice or speech-related disorders (verified through medical records) were excluded to minimize potential bias.†† Categorical variables (sex, age intervals, comorbidities) are presented as n (%) and analyzed using chi-square test, continuous variable (age) as mean±standard deviation (SD) and analyzed using Mann–Whitney U test. Characteristics of the study population. < 0.001 (χ2 = 30.76, df = 1) Age distribution, n (%) < 0.001 (χ2 = 31.98, df = 4) Comorbidities, n (%) < 0.001 (χ2 = 36.10, df = 5) Gastrointestinal and oropharyngeal disorders Age distribution, n (%) < 0.001 (χ2 = 17.53, df = 4) Comorbidities, n (%) 0.002 (χ2 = 18.54, df = 5) Gastrointestinal and oropharyngeal disorders Age distribution, n (%) < 0.001 (χ2 = 13.53, df = 4) Comorbidities, n (%) 0.140 (χ2 = 8.31, df = 5) Gastrointestinal and oropharyngeal disorders † Models were trained separately by sex due to voice differences. Participants aged under 40 y and those with voice or speech-related disorders (verified through medical records) were excluded to minimize potential bias. †† Categorical variables (sex, age intervals, comorbidities) are presented as n (%) and analyzed using chi-square test, continuous variable (age) as mean±standard deviation (SD) and analyzed using Mann–Whitney U test. The sustained phonation duration characteristics of the study participants are presented in Table 3. To further examine phonation duration characteristics, the normal group was subdivided into healthy controls (without VFSS) and VFSS-normal (PAS 1) participants. The analysis revealed significant differences across the three groups overall and for both sexes (p < 0.001). The healthy control group (overall: 6.19 ± 1.70 s; males: 6.25 ± 1.37 s; females: 6.17 ± 1.81 s) showed the longest durations, whereas the VFSS-normal (overall: 2.48 ± 2.21 s; males: 2.71 ± 2.02 s; females: 2.32 ± 2.34 s) and VFSS-aspiration-risk (overall: 2.22 ± 1.47 s; males: 2.28 ± 1.55 s; females: 2.00 ± 1.16 s) groups showed significantly shorter durations (p < 0.001), with no significant difference between VFSS-normal and VFSS-aspiration-risk (p > 0.05). The distribution of phonation duration categories also differed significantly across groups for both sexes (p < 0.001). Consequently, as approximately 57% of the VFSS participants exhibited a phonation duration of < 2 s, the number of segments generated after 2 s segmentation was more limited compared to the healthy control group. Table 3Sustained phonation duration by group and sex.SexGroup N † Mean ± SD(sec)p-value††< 2 s(n, %)2–5 s(n, %)> 5 s(n, %)p-value†††Overall(Male+Female)Healthy control (normal)*1456.19 ± 1.70< 0.0012(1.38%)23(15.86%)120(82.76%)< 0.001VFSS-normal**652.48 ± 2.2138(58.46%)23(35.38%)4(6.15%)VFSS-aspiration-risk**1932.22 ± 1.47110(56.99%)70(36.27%)13(6.74%)MaleHealthy control (normal)*386.25 ± 1.37< 0.0010(0.00%)7(18.42%)31(81.58%)< 0.001VFSS-normal**262.71 ± 2.0213(50.00%)11(42.31%)2(7.69%)VFSS-aspiration-risk**1472.28 ± 1.5582(55.78%)53(36.05%)12(8.16%)FemaleHealthy control (normal)*1076.17 ± 1.81< 0.0012(1.87%)16(14.95%)89(83.18%)< 0.001VFSS-normal**392.32 ± 2.3425(64.10%)12(30.77%)2(5.13%)VFSS-aspiration-risk**462.00 ± 1.1628(60.87%)17(36.96%)1(2.17%)† N=number of voice recordings (original length). Duration is presented as mean ± SD. Categorical data are presented as n (%).†† Kruskal–Wallis test for continuous duration data.††† Chi-square test for categorical duration data.*, ** Healthy control (normal) group differ significantly from VFSS-normal and VFSS-aspiration-risk groups (p < 0.001). There is no significant difference between VFSS-normal and VFSS-aspiration-risk groups (p > 0.05). Pairwise comparisons with Bonferroni correction (p < 0.0167): Mann–Whitney U test for continuous data, chi-square test for categorical data. Table 3Sustained phonation duration by group and sex.SexGroup N † Mean ± SD(sec)p-value††< 2 s(n, %)2–5 s(n, %)> 5 s(n, %)p-value†††Overall(Male+Female)Healthy control (normal)*1456.19 ± 1.70< 0.0012(1.38%)23(15.86%)120(82.76%)< 0.001VFSS-normal**652.48 ± 2.2138(58.46%)23(35.38%)4(6.15%)VFSS-aspiration-risk**1932.22 ± 1.47110(56.99%)70(36.27%)13(6.74%)MaleHealthy control (normal)*386.25 ± 1.37< 0.0010(0.00%)7(18.42%)31(81.58%)< 0.001VFSS-normal**262.71 ± 2.0213(50.00%)11(42.31%)2(7.69%)VFSS-aspiration-risk**1472.28 ± 1.5582(55.78%)53(36.05%)12(8.16%)FemaleHealthy control (normal)*1076.17 ± 1.81< 0.0012(1.87%)16(14.95%)89(83.18%)< 0.001VFSS-normal**392.32 ± 2.3425(64.10%)12(30.77%)2(5.13%)VFSS-aspiration-risk**462.00 ± 1.1628(60.87%)17(36.96%)1(2.17%)† N=number of voice recordings (original length). Duration is presented as mean ± SD. Categorical data are presented as n (%).†† Kruskal–Wallis test for continuous duration data.††† Chi-square test for categorical duration data.*, ** Healthy control (normal) group differ significantly from VFSS-normal and VFSS-aspiration-risk groups (p < 0.001). There is no significant difference between VFSS-normal and VFSS-aspiration-risk groups (p > 0.05). Pairwise comparisons with Bonferroni correction (p < 0.0167): Mann–Whitney U test for continuous data, chi-square test for categorical data. Sustained phonation duration by group and sex. Overall (Male+ Female) 2 (1.38%) 23 (15.86%) 120 (82.76%) 38 (58.46%) 23 (35.38%) 4 (6.15%) 110 (56.99%) 70 (36.27%) 13 (6.74%) 0 (0.00%) 7 (18.42%) 31 (81.58%) 13 (50.00%) 11 (42.31%) 2 (7.69%) 82 (55.78%) 53 (36.05%) 12 (8.16%) 2 (1.87%) 16 (14.95%) 89 (83.18%) 25 (64.10%) 12 (30.77%) 2 (5.13%) 28 (60.87%) 17 (36.96%) 1 (2.17%) † N=number of voice recordings (original length). Duration is presented as mean ± SD. Categorical data are presented as n (%). †† Kruskal–Wallis test for continuous duration data. ††† Chi-square test for categorical duration data. *, ** Healthy control (normal) group differ significantly from VFSS-normal and VFSS-aspiration-risk groups (p < 0.001). There is no significant difference between VFSS-normal and VFSS-aspiration-risk groups (p > 0.05). Pairwise comparisons with Bonferroni correction (p < 0.0167): Mann–Whitney U test for continuous data, chi-square test for categorical data. The model performance showed distinct patterns across the sex-specific and integrated models. In the male model, mn40_as achieved the highest mean AUC of 0.7586 (95% CI, 0.6499–0.8673), whereas mn30_as demonstrated superior mean performance in terms of sensitivity of 91.88% (95% CI, 86.92–96.83%) and accuracy of 82.04% (95% CI, 74.73–89.35%). For female participants, the non-pre-trained mn4.0 model consistently showed the best mean performance across all metrics, with an AUC of 0.7376 (95% CI, 0.6451–0.8301), sensitivity of 61.11% (95% CI, 40.98–81.25%), and accuracy of 77.53% (95% CI, 70.61–84.44%). In the integrated sex model, by combining both the male and female data, mn4.0 achieved the highest mean AUC of 0.8090 (95% CI, 0.7451–0.8729) and accuracy of 80.74% (95% CI, 74.76–86.72%), whereas mn40_as showed the highest mean sensitivity at 82.77% (95% CI, 74.83–90.72%). Detailed performance metrics are presented in Tables 4 and 5, and ROC curves are shown in Fig. 3. Table 4Performance metrics of males and females models for aspiration detection in post-swallowing voice analysis.ModelAUC†AccuracySensitivitySpecificityPPV†NPV†F1-ScoreLossTrain-AccuracyTrain-LossMale ModelsPre-trained modelsmn40_asmean0.7586*81.7491.7659.96*83.1472.90*0.86800.6233100.000.0032(95% CI)(0.6499, 0.8673)(74.67, 88.82)(85.59, 97.92)(40.09, 79.83)(76.62, 89.66)(57.38, 88.42)(0.8212, 0.9148)(0.3357, 0.9108)(100.00, 100.00)(0.0010, 0.0053)max0.987298.18100.0097.4495.00100.000.96971.3934100.000.0118mn30_asmean0.757282.04*91.88*59.5784.13*69.610.8750*0.6628100.000.0055(95% CI)(0.6372, 0.8773)(74.73, 89.35)(86.92, 96.83)(37.57, 81.56)(77.60, 90.66)(51.52, 87.72)(0.8318, 0.9188)(0.3267, 0.9988)(100.00, 100.00)(−0.0019, 0.0130)max1.0000100.00100.00100.00100.00100.001.00001.6734100.000.0397Non-pre-trainedmodelsmn4.0mean0.723077.4785.9158.7080.1662.280.82680.933499.850.0287(95% CI)(0.6033, 0.8427)(69.39, 85.55)(80.40, 91.42)(37.74, 79.66)(72.79, 87.54)(42.95, 81.62)(0.7716, 0.8819)(0.6295, 1.2372)(99.65, 100.05)(−0.0193, 0.0767)max0.961594.29100.00100.00100.00100.000.96001.5356100.000.2491mn3.0mean0.711876.8583.7158.6580.7660.940.81880.921999.790.0064(95% CI)(0.5778, 0.8459)(67.09, 86.60)(75.67, 91.76)(35.36, 81.94)(73.90, 87.61)(36.49, 85.38)(0.7555, 0.8821)(0.4728, 1.3710)(99.42, 100.16)(−0.0005, 0.0132)max0.882991.23100.0092.3195.45100.000.94122.3693100.000.0378 Female Models Pre-trained modelsmn40_asmean0.709076.3958.2883.5358.3184.130.56771.034599.990.0058(95% CI)(0.6125, 0.8055)(68.63, 84.16)(43.06, 73.49)(76.12, 90.94)(40.00, 76.63)(77.44, 90.82)(0.4205, 0.7150)(0.6589, 1.4101)(99.96, 100.01)(−0.0031, 0.0148)max0.919689.02100.00100.00100.00100.000.85252.1255100.000.0468mn30_asmean0.689376.0253.3284.5458.5482.530.54701.064499.940.0214(95% CI)(0.5743, 0.8043)(67.02, 85.03)(36.12, 70.51)(76.37, 92.72)(37.09, 80.00)(76.22, 88.84)(0.3711, 0.7228)(0.7570, 1.3717)(99.90, 99.99)(0.0001, 0.0428)max0.825090.9177.78100.00100.0092.860.78792.2674100.000.0946Non-pre-trainedmodelsmn4.0mean0.7376**77.53**61.11**86.41**63.29**84.30**0.5792**1.663799.970.0039(95% CI)(0.6451, 0.8301)(70.61, 84.44)(40.98, 81.25)(79.25, 93.56)(46.75, 79.83)(75.30, 93.31)(0.4417, 0.7168)(1.1316, 2.1957)(99.94, 100.01)(0.0005, 0.0073)max0.896391.46100.0098.2591.30100.000.86272.9468100.000.0157mn3.0mean0.709776.2458.0283.9260.4283.500.57261.737599.970.0137(95% CI)(0.6330, 0.7864)(69.32, 83.16)(45.62, 70.43)(76.45, 91.39)(44.28, 76.56)(77.54, 89.45)(0.4500, 0.6955)(1.0638, 2.4112)(99.75, 100.00)(−0.0035, 0.0309)max0.832589.6178.1396.4987.5093.550.78133.2840100.000.0796† AUC: area under the receiver operating characteristic curve, PPV: positive predictive value (i.e., precision), NPV: negative predictive value.* These metrics represent the highest performance achieved for each parameter in the male model. The pre-trained model mn40_as shows the highest mean performance in AUC, specificity, and NPV. The remaining metrics – accuracy, sensitivity, PPV, and F1-score – demonstrate the highest performance in mn30_as.** In the female model, the non-pre-trained model mn4.0 demonstrates the highest performance across all metrics. Table 4Performance metrics of males and females models for aspiration detection in post-swallowing voice analysis.ModelAUC†AccuracySensitivitySpecificityPPV†NPV†F1-ScoreLossTrain-AccuracyTrain-LossMale ModelsPre-trained modelsmn40_asmean0.7586*81.7491.7659.96*83.1472.90*0.86800.6233100.000.0032(95% CI)(0.6499, 0.8673)(74.67, 88.82)(85.59, 97.92)(40.09, 79.83)(76.62, 89.66)(57.38, 88.42)(0.8212, 0.9148)(0.3357, 0.9108)(100.00, 100.00)(0.0010, 0.0053)max0.987298.18100.0097.4495.00100.000.96971.3934100.000.0118mn30_asmean0.757282.04*91.88*59.5784.13*69.610.8750*0.6628100.000.0055(95% CI)(0.6372, 0.8773)(74.73, 89.35)(86.92, 96.83)(37.57, 81.56)(77.60, 90.66)(51.52, 87.72)(0.8318, 0.9188)(0.3267, 0.9988)(100.00, 100.00)(−0.0019, 0.0130)max1.0000100.00100.00100.00100.00100.001.00001.6734100.000.0397Non-pre-trainedmodelsmn4.0mean0.723077.4785.9158.7080.1662.280.82680.933499.850.0287(95% CI)(0.6033, 0.8427)(69.39, 85.55)(80.40, 91.42)(37.74, 79.66)(72.79, 87.54)(42.95, 81.62)(0.7716, 0.8819)(0.6295, 1.2372)(99.65, 100.05)(−0.0193, 0.0767)max0.961594.29100.00100.00100.00100.000.96001.5356100.000.2491mn3.0mean0.711876.8583.7158.6580.7660.940.81880.921999.790.0064(95% CI)(0.5778, 0.8459)(67.09, 86.60)(75.67, 91.76)(35.36, 81.94)(73.90, 87.61)(36.49, 85.38)(0.7555, 0.8821)(0.4728, 1.3710)(99.42, 100.16)(−0.0005, 0.0132)max0.882991.23100.0092.3195.45100.000.94122.3693100.000.0378 Female Models Pre-trained modelsmn40_asmean0.709076.3958.2883.5358.3184.130.56771.034599.990.0058(95% CI)(0.6125, 0.8055)(68.63, 84.16)(43.06, 73.49)(76.12, 90.94)(40.00, 76.63)(77.44, 90.82)(0.4205, 0.7150)(0.6589, 1.4101)(99.96, 100.01)(−0.0031, 0.0148)max0.919689.02100.00100.00100.00100.000.85252.1255100.000.0468mn30_asmean0.689376.0253.3284.5458.5482.530.54701.064499.940.0214(95% CI)(0.5743, 0.8043)(67.02, 85.03)(36.12, 70.51)(76.37, 92.72)(37.09, 80.00)(76.22, 88.84)(0.3711, 0.7228)(0.7570, 1.3717)(99.90, 99.99)(0.0001, 0.0428)max0.825090.9177.78100.00100.0092.860.78792.2674100.000.0946Non-pre-trainedmodelsmn4.0mean0.7376**77.53**61.11**86.41**63.29**84.30**0.5792**1.663799.970.0039(95% CI)(0.6451, 0.8301)(70.61, 84.44)(40.98, 81.25)(79.25, 93.56)(46.75, 79.83)(75.30, 93.31)(0.4417, 0.7168)(1.1316, 2.1957)(99.94, 100.01)(0.0005, 0.0073)max0.896391.46100.0098.2591.30100.000.86272.9468100.000.0157mn3.0mean0.709776.2458.0283.9260.4283.500.57261.737599.970.0137(95% CI)(0.6330, 0.7864)(69.32, 83.16)(45.62, 70.43)(76.45, 91.39)(44.28, 76.56)(77.54, 89.45)(0.4500, 0.6955)(1.0638, 2.4112)(99.75, 100.00)(−0.0035, 0.0309)max0.832589.6178.1396.4987.5093.550.78133.2840100.000.0796† AUC: area under the receiver operating characteristic curve, PPV: positive predictive value (i.e., precision), NPV: negative predictive value.* These metrics represent the highest performance achieved for each parameter in the male model. The pre-trained model mn40_as shows the highest mean performance in AUC, specificity, and NPV. The remaining metrics – accuracy, sensitivity, PPV, and F1-score – demonstrate the highest performance in mn30_as.** In the female model, the non-pre-trained model mn4.0 demonstrates the highest performance across all metrics. Performance metrics of males and females models for aspiration detection in post-swallowing voice analysis. Non-pre-trained models Non-pre-trained models † AUC: area under the receiver operating characteristic curve, PPV: positive predictive value (i.e., precision), NPV: negative predictive value. * These metrics represent the highest performance achieved for each parameter in the male model. The pre-trained model mn40_as shows the highest mean performance in AUC, specificity, and NPV. The remaining metrics – accuracy, sensitivity, PPV, and F1-score – demonstrate the highest performance in mn30_as. ** In the female model, the non-pre-trained model mn4.0 demonstrates the highest performance across all metrics. Table 5Performance metrics of the sex-integrated (male + female) Model.ModelAUC†AccuracySensitivitySpecificityPPV†NPV†F1-ScoreLossTrain-AccuracyTrain-LossPre-trained modelsmn40_asmean0.803180.2882.77*77.8576.0283.26*0.78660.954299.980.0027(95% CI)(0.7527, 0.8536)(75.32, 85.24)(74.83, 90.72)(72.36, 83.33)(72.27, 79.77)(75.14, 91.38)(0.7432, 0.8339)(0.6239, 1.2846)(99.95, 100.01)(0.0007, 0.0047)max0.902888.8197.2291.3584.8198.680.86962.0878100.000.0106mn30_asmean0.794979.4181.4577.5375.9281.640.78220.829999.950.0059(95% CI)(0.7466, 0.8433)(74.60, 84.21)(75.71, 87.19)(70.62, 84.44)(70.90, 80.95)(74.48, 88.80)(0.7463, 0.8180)(0.6161, 1.0436)(99.90, 100.00)(−0.0017, 0.0136)max0.916392.5490.0093.2789.3397.000.84381.3631100.000.0405Non-pre-trainedmodelsmn4.0mean0.8090*80.74*81.7880.02*78.56*81.550.7971*1.157599.920.0054(95% CI)(0.7451, 0.8729)(74.76, 86.72)(74.04, 89.52)(72.20, 87.83)(73.88, 83.24)(73.24, 89.87)(0.7508, 0.8435)(0.8748, 1.4403)(99.83, 100.01)(−0.0006, 0.0115)max0.951992.54100.0091.1187.01100.000.88551.6770100.000.0308mn3.0mean0.782278.3678.0878.3575.5879.650.76411.162599.900.0118(95% CI)(0.7378, 0.8266)(73.69, 83.03)(71.50, 84.66)(72.68, 84.02)(71.85, 79.31)(72.01, 87.30)(0.7281, 0.8000)(0.7954, 1.5296)(99.81, 100.00)(−0.0071, 0.0307)max0.890190.3086.6791.3582.9395.960.81252.1447100.000.0983† AUC: area under the receiver operating characteristic curve, PPV: positive predictive value (i.e., precision), NPV: negative predictive value.* In the integrated model combining male and female data, the non-pre-trained model mn4.0 achieves the highest values in AUC, accuracy, specificity, PPV, and F1-score. The remaining metrics – sensitivity and NPV – show the highest values in the pre-trained model mn40_as. Table 5Performance metrics of the sex-integrated (male + female) Model.ModelAUC†AccuracySensitivitySpecificityPPV†NPV†F1-ScoreLossTrain-AccuracyTrain-LossPre-trained modelsmn40_asmean0.803180.2882.77*77.8576.0283.26*0.78660.954299.980.0027(95% CI)(0.7527, 0.8536)(75.32, 85.24)(74.83, 90.72)(72.36, 83.33)(72.27, 79.77)(75.14, 91.38)(0.7432, 0.8339)(0.6239, 1.2846)(99.95, 100.01)(0.0007, 0.0047)max0.902888.8197.2291.3584.8198.680.86962.0878100.000.0106mn30_asmean0.794979.4181.4577.5375.9281.640.78220.829999.950.0059(95% CI)(0.7466, 0.8433)(74.60, 84.21)(75.71, 87.19)(70.62, 84.44)(70.90, 80.95)(74.48, 88.80)(0.7463, 0.8180)(0.6161, 1.0436)(99.90, 100.00)(−0.0017, 0.0136)max0.916392.5490.0093.2789.3397.000.84381.3631100.000.0405Non-pre-trainedmodelsmn4.0mean0.8090*80.74*81.7880.02*78.56*81.550.7971*1.157599.920.0054(95% CI)(0.7451, 0.8729)(74.76, 86.72)(74.04, 89.52)(72.20, 87.83)(73.88, 83.24)(73.24, 89.87)(0.7508, 0.8435)(0.8748, 1.4403)(99.83, 100.01)(−0.0006, 0.0115)max0.951992.54100.0091.1187.01100.000.88551.6770100.000.0308mn3.0mean0.782278.3678.0878.3575.5879.650.76411.162599.900.0118(95% CI)(0.7378, 0.8266)(73.69, 83.03)(71.50, 84.66)(72.68, 84.02)(71.85, 79.31)(72.01, 87.30)(0.7281, 0.8000)(0.7954, 1.5296)(99.81, 100.00)(−0.0071, 0.0307)max0.890190.3086.6791.3582.9395.960.81252.1447100.000.0983† AUC: area under the receiver operating characteristic curve, PPV: positive predictive value (i.e., precision), NPV: negative predictive value.* In the integrated model combining male and female data, the non-pre-trained model mn4.0 achieves the highest values in AUC, accuracy, specificity, PPV, and F1-score. The remaining metrics – sensitivity and NPV – show the highest values in the pre-trained model mn40_as. Performance metrics of the sex-integrated (male + female) Model. Non-pre-trained models † AUC: area under the receiver operating characteristic curve, PPV: positive predictive value (i.e., precision), NPV: negative predictive value. * In the integrated model combining male and female data, the non-pre-trained model mn4.0 achieves the highest values in AUC, accuracy, specificity, PPV, and F1-score. The remaining metrics – sensitivity and NPV – show the highest values in the pre-trained model mn40_as. Fig. 3ROC curves of the best-performing models by sex. The darker colored curves represent the mean receiver operating characteristic (ROC) curves across 10 folds, while the lighter colored curves indicate individual ROC curves for each cross-validation fold. For the male model, using the pre-trained mn40_as model, the mean AUC ± SD across 10 folds is 0.7586 ± 0.1442. For the female model and the integrated sex model, using the non-pre-trained mn4.0 model, the results are 0.7376 ± 0.1226 and 0.8090 ± 0.0847, respectively. (A). Male models (mn40_as). (B). Female models (mn4.0). (C). Integrated sex (male + female) models (mn4.0). ROC curves of the best-performing models by sex. The darker colored curves represent the mean receiver operating characteristic (ROC) curves across 10 folds, while the lighter colored curves indicate individual ROC curves for each cross-validation fold. For the male model, using the pre-trained mn40_as model, the mean AUC ± SD across 10 folds is 0.7586 ± 0.1442. For the female model and the integrated sex model, using the non-pre-trained mn4.0 model, the results are 0.7376 ± 0.1226 and 0.8090 ± 0.0847, respectively. (A). Male models (mn40_as). (B). Female models (mn4.0). (C). Integrated sex (male + female) models (mn4.0). Discussion This study evaluated whether standardizing post-swallowing voice recordings into 2 s segments could improve aspiration-risk detection in patients with dysphagia. Compared with our previous variable-length approach, the 2 s segmentation strategy was associated with improved and more consistent classification performance34,35. The male model showed a mean AUC ranging from 0.7118 to 0.7586 and accuracy from 76.85% to 82.04%, with consistently high sensitivity from 83.71% to 91.88%. In the female model, the AUC ranged from 0.6893 to 0.7376 and accuracy from 76.02% to 77.53%, whereas sensitivity was lower, ranging from 53.32% to 61.11%. The integrated model (male+female) demonstrated the best overall performance, with an AUC range of 0.7822 to 0.8090, accuracy from 78.36% to 80.74%, and sensitivity from 78.08% to 82.77%. Although the limited number of female patients with aspiration symptoms likely contributed to the lower sensitivity of the female model, these findings support the clinical utility of the refined 2 s segmentation approach, particularly for screening scenarios in which sensitivity is prioritized. The 2 s segmentation approach appeared to confer advantages beyond a simple increase in sample number48. By standardizing input length, it reduced variability across recordings and better matched the phonation capacity of the target clinical population. In our cohort, participants who underwent VFSS had significantly shorter phonation durations than healthy controls (p < 0.001), whereas no significant duration difference was observed between the VFSS-normal and aspiration-risk groups (Table 3). This finding is in line with previous studies indicating that patients with dysphagia or age-related neuromuscular decline often exhibit reduced maximum phonation time18,49–51. Among VFSS participants classified as at aspiration-risk, 55.78% of males and 60.87% of females produced phonation shorter than 2 s, whereas only 8.16% and 2.17%, respectively, sustained phonation longer than 5 s. These observations are consistent with previous reports that age-related declines in pulmonary elastic recoil and respiratory muscle strength lead to shortened maximum phonation time and reduced utterance length52–55. In this context, a short 2 s input window is clinically feasible and supports concise recording protocols suitable for real-time screening and telemedicine applications. Furthermore, segmentation into shorter segments of approximately 2 s allowed more precise trimming of noise at the beginning and end of each segment, likely improving overall signal quality alongside the increase in sample size, in line with prior evidence that audio noise reduction can enhance machine learning performance48,56. However, the relative contributions of increased data volume and improved signal quality could not be disentangled in this study and warrant further investigation, and the lack of systematic comparison of pre- and post-swallowing phonation durations calls for prospective validation. Beyond the methodological advantages of the 2 s scheme, the non-pre-trained mn4.0 model showed performance comparable to or slightly better than that of the pre-trained models across most metrics, particularly in the female and integrated groups. Although these differences were modest, they are noteworthy because the pre-trained EfficientAT models were developed on AudioSet, which differs substantially from clinical voice data reflecting subtle dysphagia-related pathology46,47. Such domain mismatch can lead to negative transfer, especially when the available training data are limited, as highlighted in prior transfer learning research32,57,58. Accordingly, our results suggest that, in specialized clinical applications with restricted and highly specific datasets, learning features directly from the target domain may be as effective as, or occasionally preferable to, relying on large-scale pre-trained models. Potential contributors to the observed performance differences, including domain mismatch, data imbalance, and sex-based variability, are discussed further in the Limitations section. We also considered sex-related differences in voice characteristics and therefore trained both sex-specific and sex-integrated models. Previous studies have shown that male and female voices differ in frequency distribution, voice quality, and resonance owing to anatomical and physiological differences, and that sex-specific modeling can improve performance in some voice-based disease classification tasks59–61. To examine whether such differences were present in our cohort, we analyzed several mel-spectrogram-derived acoustic features, including mean amplitude, amplitude standard deviation, energy, spectral centroid, and spectral bandwidth, using Mann–Whitney U tests. In the normal group, mean amplitude (p < 0.001), energy (p = 0.018), and spectral bandwidth (p < 0.001) showed significant sex-based differences, whereas in the aspiration group most features did not differ significantly between sexes, with the notable exception of spectral centroid (p< 0.001). These findings suggest that, although clear sex-related acoustic differences are evident in normal phonation, aspiration-related changes may attenuate the usual distinctions between male and female voices in this clinical context59,61. Consistent with this pattern, the sex-integrated MobileNetV3 models outperformed the sex-specific models when applied with the same architecture (e.g., mn4.0: integrated AUC 0.8090 vs. male AUC 0.7230 and female AUC 0.7376; mn40_as: integrated AUC 0.8031 vs. male AUC 0.7586 and female AUC 0.7090), indicating that a sex-integrated approach may be more robust and clinically practical in our setting46,47. At the same time, these results support the relevance of sex-based acoustic differences in voice-based disease classification and justify continued sex-aware sampling and performance monitoring in future work59–61. In this study, penetration-aspiration was graded using the PAS, and scores ≥ 5 were classified as aspiration-risk because they are expected to affect vocal cord vibration and are clinically associated with airway invasions and aspiration pneumonia4,5,7–10,62–66.Because the proposed model is intended as a non-invasive screening tool rather than a confirmatory diagnostic instrument, sensitivity was prioritized, and the aspiration-risk group should be interpreted as encompassing high-risk levels of airway invasion requiring further instrumental evaluation such as VFSS. The proposed approach offers several clinical advantages over existing non-invasive dysphagia screening methods. Unlike expert-dependent tools such as the 3-ounce water swallow test, the Gugging Swallowing Screen, the Functional Oral Intake Scale, the Eating Assessment Tool-10, and the Mann Assessment of Swallowing Ability, the proposed method requires only a brief voice recording, enabling rapid screening without specialist involvement67–72. Notably, the algorithm is best viewed as a preliminary screening tool rather than a confirmatory diagnostic instrument, and when aspiration risk is detected, confirmatory evaluation through instrumental assessments such as VFSS by medical professionals is essential. The short recording time reduces patient burden, and integration with mobile applications enables longitudinal monitoring and personalized management, potentially improving quality of life for patients with dysphagia73. Regarding the choice of voice elicitation method, this study used sustained vowel phonation (/a:/), which can be performed by anyone regardless of their native language. Although connected speech may better capture wet voice characteristics, sustained vowels are less affected by linguistic and cognitive variability, making them more suitable for diverse multinational settings73–76. Future research is warranted to compare the detection performance between sustained vowels and connected speech to further evaluate the clinical efficacy of different utterance types. Additionally, the current study relied on 10-fold cross-validation without an independent hold-out set or external validation cohort. As a result, the reported performance should be interpreted as internal validation only, and the generalizability of the model to other institutions, devices, and patient populations remains uncertain. External validation using multi-institutional cohorts will therefore be essential to confirm the robustness of the proposed algorithm across diverse clinical settings. Limitations This study has several limitations that should be acknowledged. First, data imbalance, particularly in the female aspiration-risk group, remained a significant challenge. Although participants aged < 40 y were excluded, a significant difference in age distribution persisted between the two groups. This reflects the clinical reality that dysphagia is more prevalent in older populations, but the influence of age on vocal characteristics may not have been fully controlled. Second, the normal group in this study consisted of healthy controls (who did not undergo VFSS) and VFSS-normal (PAS 1) participants. As recruiting sufficient VFSS-normal controls within the clinical setting was challenging, additional healthy controls who were confirmed free of clinical indications of dysphagia were recruited; however, silent aspiration cannot be entirely ruled out in this group. Owing to clinical setting limitations, normal participants recorded their voices at home using smartphones, whereas the voices of VFSS participants were recorded using a specialized recorder in the hospital. Although we confirmed minimal device-related bias through cosine similarity analysis (> 0.85), differences in recording environments remain a limitation. Individual variability in vocalization timing and susceptibility of omnidirectional microphones to ambient noise are additional confounding factors; the use of USB-C-compatible unidirectional microphones should be considered in future studies77. The distinction between inpatients and outpatients among the VFSS group was not documented. As a result, this study could not perform a stratified analysis regarding the impact of patient severity or clinical environment. Third, although VFSS test participants swallowed solutions of various consistencies during testing, the normal participants swallowed only water, creating a potential bias in food consistency exposure. The impact of different bolus types on PAS scores and model performance could not be examined. Importantly, VFSS reflects aspiration status at a single time point under specific conditions, which may not capture real-world aspiration risk during daily meals73. This underscores the need for mobile-based applications that allow for continuous, real-world screening to capture aspiration risks during daily activities. Fourth, participants with mild dysphagia (penetration group, PAS 2–4) were excluded due to limited sample size; however, focusing on clear classification between normal (PAS 1) and aspiration-risk (PAS ≥ 5) groups allowed for more robust model training. As shown in the supplementary analyses (Supplementary Note 1–3, Table S1–S8), incorporating this intermediate group into multi-class or modified binary classification tasks resulted in a noticeable decline in performance when it was treated as a separate or normal-adjacent category, likely due to both limited sample size and inherent acoustic ambiguity of this category. Nevertheless, excluding this ‘gray zone’ (PAS 2–4) may involve a trade-off with clinical sensitivity for early detection. More broadly, the underlying mechanism of how the presence of a bolus at PAS ≥ 5 affects vocal cord vibration remains to be further validated through prospective studies. Fifth, although the sex-integrated model outperformed sex-specific models as demonstrated in the Results, the performance discrepancies between pre-trained and non-pre-trained models likely reflect a combination of domain mismatch, data imbalance, and sex-based acoustic differences, and require large-scale validation32,46,47,57,58. Clinical diagnoses were not incorporated as input variables, and including factors such as diagnosis and sex has been reported to improve the prediction of airway invasion, highlighting the need for multimodal models that integrate acoustic data with diverse clinical information38. Sixth, model performance was assessed using a 10-fold cross-validation without a separate internal hold-out set or external validation cohort. Consequently, the reported performance should be regarded as internal validation only, and the generalizability of the model to other institutions, devices, and patient populations remains uncertain. To address these limitations, we are expanding our research to multiple institutions, recruiting additional participants, implementing a dedicated validation set, and re-evaluating the penetration group (PAS 2–4). In addition, we plan to enhance the investigation by using various recording devices and food consistencies. Conclusions This study demonstrated that a standardized 2 s voice segmentation approach can enable effective detection of aspiration-risk in patients with dysphagia, with the sex-integrated model achieving particularly favorable performance. The proposed approach offers several practical advantages, including a brief recording time, the applicability to patients with limited phonation capacity, and compatibility with different mobile and telemedicine platforms for continuous monitoring. Limitations remain, including data imbalance, exclusion of the penetration group (PAS 2–4), and variability in recording environments, which will be addressed through ongoing multi-institutional data collection and model refinement. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Supplementary Material 1 Supplementary Material 2 Supplementary Material 2 Supplementary Material 3 Supplementary Material 3