Symptom clustering of old adult dysphagic patient phenotypes by unsupervised machine learning using multidimensional characteristics: a cross-sectional study

Abstract

Background

Dysphagia severely affects the old adult population. Its signs and symptoms are variable, which poses challenges to diagnosis and treatment. Existing assessment tools fall short in considering the coordination of relevant muscle groups in the pathology of dysphagia and fail to characterize dysphagia in a comprehensive manner. The aim of this study was to identify different phenotypes of dysphagia based on multidimensional patient characteristics.

Methods

This work was a cross-sectional study. Senior dysphagia patients (> 65 years old) diagnosed by Videofluoroscopic Swallow Study at seven tertiary hospitals in Suzhou were enrolled from June 2022 to October 2024. They were grouped into different phenotypes by the k-means clustering algorithm based on their respiratory, swallowing, and articulatory functions. The characteristics of demographics, medical history, and two dysphagia outcomes were compared across the different clustering results.

Results

The 1,301 patients were divided into four distinct phenotypes: minimal impairment (MI, n = 462), respiratory-dominant impairment (RD, n = 305), articulatory-dominant impairment (AD, n = 301), and multimodal impairment (MT, n = 233). Each phenotype exhibited a unique combination of respiratory, swallowing, and articulatory impairments, and the demographics (e.g., age, BMI, income, etc.) and medical history (e.g., head and neck tumors, masticatory disorders) differed significantly between phenotypes (P < 0.05). The eating ability and the incidence of swallowing-related complications both differed significantly (P < 0.001) between phenotypes. The degree of dysphagia fell in the order MT > RD > AD > MI and was consistent with the Penetration-Aspiration Scale score.

Discussion

This study identified different phenotypes of dysphagia based on multidimensional patient characteristics with unique clinical features and outcomes. The findings support the development of targeted therapeutic strategies to improve patient prognosis.

Introduction

Swallowing is a complex process regulated by neural circuits in the brainstem and higher brain centers, coordinating muscle activity in the mouth, pharynx, and esophagus [1]. Disruptions in this coordination result in dysphagia, impairing both swallowing safety (risk of aspiration) and efficacy (incomplete bolus clearance). [2] The hyoid muscles play a crucial role in swallowing, as their mobility influences the elevation and depression of the hyoid bone and larynx. They also contribute to respiratory and articulatory functions by supporting the upright posture of the trunk and skull and affecting breathing mechanics and pharyngeal space adjustment [3, 4]. During swallowing, the central nervous system temporarily inhibits respiration for 0.5 to 1.0 s [5, 6]. The exhalation before and after swallowing, which constitutes the Exhalation-Swallow-Exhalation pattern, helps protect the airway from residue or leakage [7]. However, improper timing of the swallow within the respiratory cycle increases the risk of post-swallow aspiration [8]. A reduced vital capacity can lead to shortness of breath, and alterations in respiratory rhythm and frequency may result in dysfunctional breathing and swallowing, further elevating aspiration risk [9].

Dysphagia profoundly impacts both health and social well-being, contributing to psychological distress such as anxiety, depression, and frustration, which can further lead to social withdrawal [10]. From a health perspective, it increases the risk of severe complications, including malnutrition, dehydration, and aspiration pneumonia [11]. These risks are particularly concerning for vulnerable populations, such as the old adults and individuals with neurological conditions (e.g., stroke, Alzheimer’s disease, Parkinson’s disease), where dysphagia significantly raises morbidity and mortality rates [12]. In addition, the cost of managing dysphagia and its complications places a substantial burden on healthcare systems and caregivers, emphasizing the urgent need for effective intervention strategies.

Dysphagia diagnosis requires a combination of clinical assessments and instrumental evaluations. Clinical assessments include patient history, physical examination, and bedside swallowing tests. In terms of instrumental evaluation, the videofluoroscopic swallow study (VFSS) and Fiberoptic Endoscopic Evaluation of Swallowing (FEES) are considered gold standards [13], although they are invasive, resource-intensive, and dependent on patient cooperation [14]. Early symptoms of dysphagia can be subtle [15, 16] with patients often experiencing reduced respiratory function, alterations in eating habits, hoarseness, or globus sensation [17]. The variability in signs and symptoms underscores the complexity of dysphagia diagnosis [18], making it difficult to achieve a comprehensive characterization with a single clinical assessment tool [19].

In clinical research, symptom clustering focuses on identifying and analyzing groups of symptoms that occur together [20, 21]. Since patients often experience multiple, concurrent symptoms rather than isolated ones, analyzing these clusters can improve symptom management and patient outcomes. The relationships between clustered symptoms may reflect common underlying mechanisms and thus provide a fresh perspective on the pathology. Symptom clustering has been applied to degenerative cervical myelopathy, seizures, systemic sclerosis, and cancers to advance the understanding and management of the conditions [22–25]. Within the dysphagia population, patients often report symptoms and functional limitations not immediately related to swallowing, but symptom clusters have not been well studied [26, 27]. For example, many dysphagia patients experience altered articulation, and acoustic indicators have previously been applied to predict the development of dysphagia [28, 29]. Identifying the symptom clusters of dysphagia is likely key to developing more effective treatment strategies, as it enables personalized treatments that target specific patterns to enhance efficacy.

K-means clustering, a powerful unsupervised machine learning algorithm, has been applied in various medical studies to characterize disease phenotypes, including those for acute ischemic stroke and COVID-19 patient stratification [30, 31]. Therefore, this work applied k-means clustering to identify distinct phenotypes of swallowing disorders, thus addressing the following research questions:

1. Can unsupervised machine learning using multidimensional characteristics effectively identify different phenotypes of dysphagia in the Chinese old adult population?

2. Do these phenotypes have distinct clinical features and outcomes?

3. Can the identification of these phenotypes provide a basis for targeted treatment strategies?

Methods

Figure 1 describes the data collection and assessment procedures for this cross-sectional study.

Fig. 1.

Information collection and assessment procedures

Participants

Patients were recruited at seven tertiary hospitals in Suzhou from June 2022 to October 2024 using convenience sampling.

Inclusion criteria

One at least 65 years old, 2) diagnosed with dysphagia by rehabilitation physicians through VFSS, 3) in a stable condition, conscious, and able to cooperate with the assessment, 4) capable of oral intake.

Exclusion criteria

One refused to join, 2) could not cooperate with the examination due to moderate-to-severe visual or auditory impairments (according to WHO definition) or cognitive impairments (Mini-Mental State Examination (MMSE) score <24) according to the assessment by a neurologist, 3) had fever or lung infection, 4) had serious internal diseases (e.g., severe heart failure (New York Heart Association class III-IV), end-stage liver disease (Child-Pugh class C), advanced kidney disease (chronic kidney disease stage 4-5), or other serious systemic conditions (uncontrolled cancer, severe respiratory failure, etc.)).

Ethical approval and informed consent

This study was approved by the Ethics Committee of the First Affiliated Hospital of Soochow University (No. 2024271) and followed the STrengthening the Reporting of OBservational Studies in Epidemiology (STROBE) guidelines. Written informed consent was obtained from all recruited patients.

Instrumental diagnostics

The VFSS was performed by a rehabilitation therapist and a radiologist as follows:

1. The therapist prepared barium sulfate contrast agents of different consistencies (liquid, thick liquid, paste, solid) and mixed them with water as appropriate.

2. The patient sat upright with their head erect to maintain proper posture for imaging and swallowing assessment.

3. Following the therapist’s guidance, the radiologist assessed the patient’s swallowing function while recording X-ray fluoroscopic images. Swallowing trials were conducted in the following order: thick liquid, paste, liquid, water, and solid.

4. Swallowing function was qualitatively assessed using the Penetration-Aspiration Scale (PAS) [32].

Any abnormal features observed during the VFSS were documented for further analysis.

Demographic and medical data

The collected socio-demographic information included age, sex, body mass index (BMI), education, marital and work status, and income level. The medically relevant information included dental conditions and comorbidities.

Assessments

The assessments of respiration, swallowing, and phonation were all completed in 24 h after the completion of VFSS.

Respiration

Respiratory parameters were assessed through a structured protocol conducted by a rehabilitation physician with extensive clinical experience. The patient was instructed to relax and breathe naturally before the tests started. Respiratory rate was measured using a stopwatch by counting the number of breaths per minute, with three repeated measurements averaged to minimize error. Vital capacity was measured using a calibrated spirometer (Vitalograph, Buckingham, UK), and the results (unit: mL) from two valid forced expiratory maneuvers were averaged. The breathing mode was classified as a binary variable (nasal breathing vs. mouth breathing) from the observed airflow direction during resting respiration and chewing. Subsequently, during a snack intake, the patient’s respiratory behavior during swallowing was observed and categorized as follows:

1. Normal pause of breath during swallowing, defined as a 0.5–1.0 s airflow cessation during swallowing.

2. Nasal or oral respiration during chewing, reflecting whether the patient primarily engaged in nasal or mouth breathing while processing food.

3. Breathing when swallowing, identified by visible chest/abdominal movements during the swallow indicating that swallowing was accompanied by inhalation/exhalation.

4. Tachypnea during eating, if respiratory rate exceeded 20 breaths/min or labored breathing (e.g., nasal flaring, excessive chest wall movement) was observed.

Swallowing

Swallowing muscle strength was assessed using the Standardized Evaluation Scale for Dysphagia developed by Kubota in 1982. This scale focuses on the clinical assessment of swallowing muscles, categorizing muscle strength impairment into four grades (1 = normal strength, 4 = complete dysfunction). It is a scientifically validated tool widely adopted in dysphagia rehabilitation and management. Additional muscle assessments were performed by an experienced rehabilitation physician who had undergone standardized training. In a quiet room, the patient was asked to press the tongue against the hard palate and the gums on each side, and the doctor observed the completion of the movement to rate the strength of the tongue muscles. To rate the strength of the buccal and masticatory muscles, the patient was asked to puff up the cheeks, and the doctor tapped the puffed cheeks to detect air leakage. The patient was next asked to bite a tongue depressor firmly with the molars, and the doctor pulled the tongue depressor to test if it could be held tightly. Both sides were tested. Finally, the patient was asked to open the mouth and make an “ah” sound, and the doctor rated the function of the pharyngeal muscles based on the elevation and symmetry of the soft palate.

After these tests, the patient chose a meal on a menu and finished the food in the usual way. While the patient ate, the doctor watched closely and recorded any posture change and relevant dietary habits. Postural change during mealtime, recorded as a binary variable (yes/no), was determined if any noticeable shift in the patient’s position occurred, such as leaning forward, reclining backward, or transitioning from sitting to a semi-sitting position.

The Ohkuma questionnaire [33, 34, 35], which contained 15 entries related to the progression of dysphagia, was used to further evaluate the swallowing pattern. The entries belonged to six groups: history of pneumonia (entry 1), nutritional status (entry 2), pharyngeal function (entries 3–7), oral function (entries 8–11), esophageal function (entries 12–14), and airway protection (entry 15).

Phonation

In the examination room with < 40 dB ambient noise, the patient talked to three doctors about their medical history in a natural tone. The doctors independently rated the patient’s voice using the GRBAS scale, including grade (G), roughness (R), breathiness (B), asthenia (A), and strain (S). Each dimension was scored on a 4-point Likert scale: 0, normal; 1, mild impairment; 2, moderate impairment; 3, severe impairment [36]. The scores from the three doctors were then averaged. Before the study, the GRBAS scorers underwent standardized training and achieved inter-rater reliability through pre-study calibration (Cohen’s kappa > 0.8).

In addition, three recordings of the patient’s voice were obtained as follows. The patient was asked to take a deep breath, then pronounce the “ah” sound in the normal way of speaking and keep it steady for as much time as possible. The most stable recording was analyzed using the Praat software to determine the maximum phonation time (MPT), fundamental frequency (F0), fundamental frequency perturbation (jitter), amplitude perturbation (shimmer), and harmonic-to-noise ratio (HNR). Figure 2 illustrates the determination of the acoustic parameters .

Fig. 2.

Extraction of acoustic parameters. A MPT. B F0, jitter, shimmer, and HNR

Dysphagia outcomes

The Penetration-Aspiration Scale (PAS) was used to reflect the severity of dysphagia: grade 1, no penetration or aspiration; grades 2–4, penetration; grades 5–8, aspiration. Higher grades indicate more severe dysphagia [32]. The eating assessment tool (EAT-10) questionnaire developed by Belafsky et al. was given to the patients, and their subjective self-evaluation was collected [37–39]. Incidences of swallowing-related complication (aspiration, aspiration pneumonia, malnutrition, choking) within seven days post-assessment were also recorded. Aspiration was confirmed clinically (e.g., persistent coughing, wheezing, sensation of throat obstruction) and/or via instrumental exams (e.g., fiberoptic bronchoscopy) when needed [40]. Aspiration pneumonia was diagnosed by new-onset chest CT scan infiltrates with fever, cough, or elevated white blood cell count [40]. Malnutrition was defined as a BMI below 18.5 [41]. Choking was defined as a sudden airway obstruction during swallowing that caused breathing difficulty and coughing, or gagging [42]. Two experienced physicians independently verified the incidences.

Phenotyping by unsupervised machine learning

The K-means clustering algorithm based on the Euclidean distance was used to divide the patients into different clusters. Categorical data were given numerical values as appropriate, and continuous variables were standardized by Z-score normalization. After randomly selecting the initial centroids, the algorithm iterated by assigning patients to clusters and then recalculating and re-centering the centroids. Convergence was accomplished when the total squared distances between each patient and its nearest centroid were minimized. The elbow method was used to determine the optimal number of clusters [43].

Statistical analysis

All statistical analyses were run in R 4.3.3. Continuous variables were presented as mean ± standard deviation (SD), and categorical variables were expressed as case count (ratio). The Shapiro–Wilk test was used to assess the normality of variable distributions. After the patients were divided into separate clusters, their demographic variables, medical conditions, and assessment results (respiration, swallowing, phonation, dysphagia) were compared across phenotypes. The analysis of variance (ANOVA) was applied to continuous variables and the chi-square test was applied to categorical variables. Differences were considered statistically significant if P <0.05.

Results

Patient profile

Table 1 shows the basic information of the 1301 patients enrolled in this study. The patients were 75.01 ± 7.49 years of age. Male and female patients accounted for 56.42% and 43.58% of the total, respectively. Most of them were retired. While 240 (18.45%) did not have any tooth loss, others all had various dental conditions, including fixed prosthesis (104, 7.99%), removable prosthesis (123, 9.45%), tooth loss without repair (741, 56.96%), and edentulous (93, 7.15%). The top three comorbidities were stroke (841, 64.64%), masticatory dysfunction (209, 16.06%), and Alzheimer’s disease (111, 8.53%).

Table 1.

Demographics and medical conditions of enrolled patients

Characteristics	All patients (n = 1301)
Demographics
Age† (years)	75.01 ± 7.49
Sex§
Male	734 (56.42%)
Female	567 (43.58%)
BMI†¶ (kg/m2)	22.58 ± 3.54
Highest level of education§
Primary school and below	612 (47.04%)
Middle school	402 (30.90%)
High school	225 (17.29%)
Professional degree	52 (4.00%)
Bachelor and above	10 (0.77%)
Employment§
Employed	3 (0.23%)
Retired	995 (76.48%)
Unemployed	303 (23.29%)
Monthly income§ (CNY¶)
1000–3000	23 (1.77%)
3001–5000	329 (25.29%)
5001–7000	554 (42.58%)
7001–9000	366 (28.13%)
> 9000	29 (2.23%)
Marital status§
Married	853 (65.56%)
Divorced	12 (0.92%)
Widowed	436 (33.51%)

Medical conditions
Tooth loss§
No tooth loss	240 (18.45%)
Fixed prosthesis	104 (7.99%)
Removable prosthesis	123 (9.45%)
Tooth loss without repair	741 (56.96%)
Edentulous	93 (7.15%)
Comorbidities§
Stroke	841 (64.64%)
Alzheimer’s disease	111 (8.53%)
Parkinson’s disease	73 (5.61%)
Amyotrophic lateral sclerosis	16 (1.23%)
Multiple sclerosis	6 (0.46%)
Head and neck tumors	96 (7.38%)
Local radiotherapy for head and neck tumors	45 (3.46%)
Coronary heart disease	45 (3.46%)
Brain injury	31 (2.38%)
Masticatory dysfunction	209 (16.06%)

¶ BMI body mass index, CNY Chinese Yuan

^† Expressed as mean ± standard deviation

^§ Expressed as case number (ratio)

Table 2 lists the assessment results of the patients. The respiratory rate was 18.39 ± 1.75 breaths per minute, and the vital capacity was 932.93 ± 644.77 mL. Nasal and oral breathing accounted for 56.80% and 43.20% of the total, respectively. Most patients (889, 68.33%) showed normal pause of breath during swallowing, but 286 (21.98%) had tachypnea during eating and 107 (8.22%) had oral respiration during chewing. Their muscle strength was rated as follows: tongue muscles, 1.77 ± 0.77; masticatory and buccal muscles, 1.70 ± 0.74; pharyngeal muscles, 1.67 ± 0.69. Their voice quality was rated as follows: grade, 0.78 ± 0.71; roughness, 0.65 ± 0.70; breathiness, 0.36 ± 0.57; asthenia, 0.39 ± 0.58; strain, 0.49 ± 0.64. The parameters extracted from their voice recording were as follows: MPT, 5.43 ± 4.12 s; F0, 164.21 ± 44.37 Hz; jitter, 1.20% ± 0.93%; shimmer, 2.95% ± 1.55%; HNR, 17.02 ± 4.36 dB.

Table 2.

Dysphagia characteristics of enrolled patients

Characteristics	All patients (n = 1301)
Respiration
Rate† (breaths per minute)	18.39 ± 1.75
Vital capacity† (mL)	932.93 ± 644.77
Type§
Trans nasal	739 (56.80%)
Oral	562 (43.20%)
Breathing mode§
Normal pause of breath during swallowing	889 (68.33%)
Nasal respiration during chewing	17 (1.31%)
Oral respiration during chewing	107 (8.22%)
Breathing when swallowing	2 (0.15%)
Tachypnea during eating	286 (21.98%)
Swallowing
Posture changes while eating§
No	1063 (81.71%)
Yes	238 (18.29%)
Eating position§
Sitting	837 (64.34%)
Semi-sitting	442 (33.97%)
Reclining	22 (1.69%)
Diet§
Regular diet	381 (29.29%)
Soft diet	556 (42.74%)
Minced diet	152 (11.68%)
Liquid diet	212 (16.30%)
Quantity of food consumed§
≥ 1 bowl	573 (44.04%)
< 1 bowl	728 (55.96%)
Reduced appetite§
No	758 (58.26%)
Yes	543 (41.74%)
Ohkuma questionnaire†
History of pneumonia	1.64 ± 0.69
Nutritional status	1.53 ± 0.70
Pharyngeal function	8.12 ± 2.30
Oral function	6.22 ± 1.93
Esophageal function	3.81 ± 1.24
Airway protection function	2.04 ± 0.89
Strength of swallowing muscles †
Tongue muscles	1.77 ± 0.77
Masticatory and buccal muscles	1.70 ± 0.74
Pharyngeal muscles	1.67 ± 0.69
Phonation
GRBAS†¶
Grade	0.78 ± 0.71
Roughness	0.65 ± 0.70
Breathiness	0.36 ± 0.57
Asthenia	0.39 ± 0.58
Strain	0.49 ± 0.64
Acoustic metrics†
Maximum phonation time (MPT, s)	5.43 ± 4.12
Fundamental frequency (F0, Hz)	164.21 ± 44.37
Jitter (%)	1.20 ± 0.93
Shimmer (%)	2.95 ± 1.55
Harmonic noise ratio (HNR, dB)	17.02 ± 4.36

^† Expressed as mean ± standard deviation

^§ Expressed as case number (ratio)

¶ Grade, roughness, breathiness, asthenia, and strain

Cluster analysis

The k-means clustering algorithm, which considered the Euclidean distance between the patients for all the listed attributes, was then applied to create clusters that represent distinct phenotypes. Figure 3 shows the initial clustering results with different numbers of preset clusters. Figure 4 shows that the decrease in the within-group sum of squares (WSS) passed an inflection point when there were four clusters. That is, according to the elbow rule, increasing the number of clusters more than four did not significantly improve the model fit. Indeed, when there were five clusters, good separation was no longer attained and there was substantial overlap (Fig. 3). The discussion below was based on dividing the patients into four clusters.

Fig. 3.

Separation of patients into 2, 3, 4, or 5 clusters by the k-means clustering algorithm. A K=2, B K=3, C K=4, D K=5

Fig. 4.

Determination of the optimum number of clusters

Four dysphagia phenotypes

Minimal impairment( (n = 462)

This phenotype, referred to hereafter as MI, was the most common, and it accounted for 35.5% of the total patients (Table 3). Overall, the patients had the least impairment in all aspects compared to the patients in other clusters. The rate of respiration was the lowest (17.58 ± 1.07 breaths per minute), and the vital capacity was the highest (1274.10 ± 633.31 mL). The ratio of transnasal breathing was the highest (71.21%), and most patients (93.51%) had a normal pause of breathing when they swallowed. Most of the patients ate while sitting upright (80.09%), and the ratio of the patients having a normal diet (48.05%) was the highest compared to other clusters. They also had the best appetite and consumed more food than other patients. They were superior to other patients in the Ohkuma questionnaire scores and the strength of swallowing muscles. Phonation was also impaired the least.

Table 3.

Dysphagia characteristics of different clusters

Characteristics	MI¶ (n = 462)	RD¶ (n = 305)	AD¶ (n = 301)	MT¶ (n = 233)	P value
Respiration
Rate† (breaths per minute)	17.58 ± 1.07	19.05 ± 1.82	17.76 ± 1.26	19.95 ± 1.93	< 0.001
Vital capacity† (mL)	1274.10 ± 633.31	578.82 ± 382.68	1192.88 ± 588.87	384.15 ± 324.73	< 0.001
Type§					< 0.001
Trans nasal	329 (71.21%)	140 (45.90%)	202 (67.11%)	68 (29.18%)
Oral	133 (28.79%)	165 (54.10%)	99 (32.89%)	165 (70.82%)
Breathing mode§					< 0.001
Normal pause of breath during swallowing	432 (93.51%)	152 (49.84%)	261 (86.71%)	44 (18.88%)
Nasal respiration during chewing	5 (1.08%)	7 (2.30%)	4 (1.33%)	1 (0.43%)
Oral respiration during chewing	21 (4.55%)	41 (13.44%)	23 (7.64%)	22 (9.44%)
Breathing when swallowing	0 (0.00%)	1 (0.33%)	0 (0.00%)	1 (0.43%)
Tachypnea during eating	4 (0.87%)	104 (34.10%)	13 (4.32%)	165 (70.82%)
Swallowing
Posture changes while eating§					< 0.001
No	434 (93.94%)	220 (72.13%)	274 (91.03%)	135 (57.94%)
Yes	28 (6.06%)	85 (27.87%)	27 (8.97%)	98 (42.06%)
Eating position§					< 0.001
Sitting	370 (80.09%)	133 (43.61%)	247 (82.06%)	87 (37.34%)
Semi-sitting	91 (19.70%)	160 (52.46%)	52 (17.28%)	139 (59.66%)
Reclining	1 (0.22%)	12 (3.93%)	2 (0.66%)	7 (3.00%)
Diet§					< 0.001
Regular diet	222 (48.05%)	37 (12.13%)	109 (36.21%)	13 (5.58%)
Soft diet	207 (44.81%)	137 (44.92%)	140 (46.51%)	72 (30.90%)
Minced diet	26 (5.63%)	33 (10.82%)	32 (10.63%)	61 (26.18%)
Liquid diet	7 (1.52%)	98 (32.13%)	20 (6.64%)	87 (37.34%)
Quantity of food consumed§					< 0.001
≥ 1 bowl	327 (70.78%)	50 (16.39%)	175 (58.14%)	21 (9.01%)
< 1 bowl	135 (29.22%)	255 (83.61%)	126 (41.86%)	212 (90.99%)
Reduced appetite§					< 0.001
No	407 (88.10%)	82 (26.89%)	210 (69.77%)	59 (25.32%)
Yes	55 (11.90%)	223 (73.11%)	91 (30.23%)	174 (74.68%)
Ohkuma questionnaire†
History of pneumonia	1.44 ± 0.56	1.82 ± 0.70	1.49 ± 0.65	1.98 ± 0.75	< 0.001
Nutritional status	1.19 ± 0.45	1.79 ± 0.70	1.38 ± 0.58	2.08 ± 0.76	< 0.001
Pharyngeal function	6.67 ± 1.57	9.23 ± 2.33	7.83 ± 1.73	9.94 ± 2.06	< 0.001
Oral function	5.98 ± 1.87	6.44 ± 1.90	5.57 ± 1.61	7.28 ± 1.98	< 0.001
Esophageal function	3.38 ± 0.81	3.80 ± 1.31	4.19 ± 1.49	4.21 ± 1.20	< 0.001
Airway protection function	1.38 ± 0.64	1.80 ± 0.82	2.67 ± 0.60	2.86 ± 0.40	< 0.001
Strength of swallowing muscles †
Tongue muscles	1.48 ± 0.66	1.85 ± 0.75	1.65 ± 0.72	2.39 ± 0.74	< 0.001
Masticatory and buccal muscles	1.55 ± 0.65	1.79 ± 0.72	1.50 ± 0.63	2.17 ± 0.81	< 0.001
Pharyngeal muscles	1.27 ± 0.48	1.75 ± 0.63	1.72 ± 0.64	2.31 ± 0.59	< 0.001
Phonation
GRBAS†¶
Grade	0.27 ± 0.45	0.51 ± 0.53	1.30 ± 0.54	1.46 ± 0.52	< 0.001
Roughness	0.18 ± 0.40	0.41 ± 0.51	1.19 ± 0.64	1.21 ± 0.60	< 0.001
Breathiness	0.04 ± 0.21	0.18 ± 0.44	0.57 ± 0.58	0.96 ± 0.62	< 0.001
Asthenia	0.06 ± 0.26	0.28 ± 0.49	0.48 ± 0.59	1.05 ± 0.55	< 0.001
Strain	0.08 ± 0.30	0.21 ± 0.46	0.89 ± 0.62	1.12 ± 0.58	< 0.001
Acoustic metrics†
Maximum phonation time (MPT, s)	7.86 ± 4.31	3.47 ± 2.54	6.28 ± 3.70	2.08 ± 1.73	< 0.001
Fundamental frequency (F0, Hz)	164.44 ± 38.51	166.16 ± 42.24	163.80 ± 47.39	161.73 ± 53.23	0.761
Jitter (%)	0.59 ± 0.41	0.78 ± 0.50	1.85 ± 0.90	2.11 ± 0.85	< 0.001
Shimmer (%)	1.85 ± 0.87	2.32 ± 1.09	4.10 ± 1.19	4.47 ± 1.29	< 0.001
Harmonic noise ratio (HNR, dB)	20.10 ± 2.99	18.33 ± 3.46	14.37 ± 2.94	12.65 ± 3.55	< 0.001

¶ MI minimal impairment, RD respiratory-dominant impairment, AD articulatory-dominant impairment, MT multimodal impairment, GRBAS grade, roughness, breathiness, asthenia, and strain

† Expressed as mean ± standard deviation

§ Expressed as case number (ratio) Bold font denotes statistical significance (P<0.05) for comparisons among clusters

Respiratory-dominant impairment (n = 305)

This phenotype, referred to hereafter as RD, accounted for 23.4% of the total (Table 3). The impairment was the most obvious in the respiratory aspect. The patients had the second highest respiratory rate (19.05 ± 1.82 breaths per minute), the second lowest vital capacity (578.82 ± 382.68 mL), and the second highest ratio of oral respiration (54.10%), and shortness of breath upon swallowing was common (34.10%). Many patients consumed a soft diet (44.92%) with low intake (≥ 1 bowl, 16.39%) and reduced appetite (73.11%). The Ohkuma assessment scores revealed moderate impairment, and the patients had weaker swallowing muscles than the MI patients. All GRBAS scores were the second lowest among the four clusters. Among the acoustic metrics, MPT was the second worst among the four clusters (3.47 ± 2.54 s), F0 was not statistically different, and all others were the second best.

Articulatory-dominant impairment (n = 301)

This phenotype, referred to hereafter as AD, accounted for 23.1% of the total (Table 3). The patients exhibited severe dysphonia. Among the four clusters, the GRBAS scores were the second highest. While their F0 did not differ from that of the patients of other phenotypes, their other acoustic parameters were significantly different. Specifically, MPT was 6.28 ± 3.70 s, jitter was 1.85% ± 0.90%, shimmer was 4.10% ± 1.19%, and HNR was 14.37 ± 2.94 dB. These acoustic metrics indicated that the AD patients had a significantly higher degree of dysarthria than the MI and RD patients (P <0.001). In contrast, they experienced less impairment in the respiratory functions compared to the RD patients. The respiratory rate was 17.76 ± 1.26 breaths per minute, the vital capacity was 1192.88 ± 588.87 mL, and more patients (67.11%) were breathing through the nose. Many patients (86.71%) could pause their breath normally during swallowing. Compared to RD, AD was less impaired in all assessments of swallowing except the esophageal function in the Ohkuma questionnaire.

Multimodal impairment (n = 233)

This phenotype, referred to hereafter as MT, accounted for 17.9% of the total (Table 3). The patients had worse assessment results than those of RD and AD in all aspects. They had the highest respiratory rate (19.95 ± 1.93 breaths per minute) and the lowest vital capacity (384.15 ± 324.73 mL). Most of them (70.82%) needed to breathe through the mouth, and shortness of breath while eating was common (70.82%). When eating, many patients changed their posture (42.06%), and only a small number could sit upright (37.34%). They ate a soft (30.90%) or liquid (37.34%) diet, mostly less than one bowl (90.99%), and the loss of appetite was very common (74.68%). They ranked the worst in all categories of the Ohkuma questionnaire and all muscle tests. They had the worst dysphonia among all four clusters, with the highest GRBAS scores, shortest MPT (2.08 ± 1.73 s), highest jitter (2.11% ± 0.85%) and shimmer (4.47% ± 1.29%), and lowest HNR (12.65 ± 3.55 dB).

Comparison of demographics and medical conditions across clusters

Significant differences existed in the following aspects of the demographic characteristics and the medical background of the patients (Table 4). MT and RD had a higher mean age (76.70, 76.12 years, respectively) and a lower mean BMI (21.47, 21.89, respectively) than AD and MI. The highest level of education, monthly income, and marital status also differed significantly across clusters. MT had the highest ratios of patients who had tooth loss without repair (63.09%) or were edentulous (12.02%). In terms of comorbidities, head and neck tumors were the most frequent among AD patients (13.62%), and masticatory dysfunction occurred more often in MT and RD patients (23.61% and 18.03%, respectively). Other demographic characteristics and comorbidities did not differ significantly between clusters.

Table 4.

Demographics and medical conditions of different clusters

Characteristics	MI¶ (n = 462)	RD¶ (n = 305)	AD¶ (n = 301)	MT¶ (n = 233)	P value
Demographics
Age† (years)	74.40 ± 7.45	76.12 ± 7.66	73.49 ± 6.53	76.70 ± 8.01	< 0.001
Sex §					0.120
Male	274 (59.31%)	164 (53.77%)	177 (58.80%)	119 (51.07%)
Female	188 (40.69%)	141 (46.23%)	124 (41.20%)	114 (48.93%)
BMI†¶ (kg/m2)	23.37 ± 3.16	21.89 ± 3.87	22.95 ± 3.36	21.47 ± 3.56	< 0.001
Highest level of education§					< 0.001
Primary school and below	182 (39.39%)	159 (52.13%)	137 (45.51%)	134 (57.51%)
Middle school	162 (35.06%)	84 (27.54%)	92 (30.56%)	64 (27.47%)
High school	81 (17.53%)	53 (17.38%)	62 (20.60%)	29 (12.45%)
Professional degree	34 (7.36%)	6 (1.97%)	9 (2.99%)	3 (1.29%)
Bachelor and above	3 (0.65%)	3 (0.98%)	1 (0.33%)	3 (1.29%)
Employment§					0.258
Employed	3 (0.65%)	0 (0.00%)	0 (0.00%)	0 (0.00%)
Retired	362 (78.35%)	230 (75.41%)	230 (76.41%)	173 (74.25%)
Unemployed	97 (21.00%)	75 (24.59%)	71 (23.59%)	60 (25.75%)
Monthly income§ (CNY¶)					0.004
1000–3000	9 (1.95%)	3 (0.98%)	6 (1.99%)	5 (2.15%)
3001–5000	97 (21.00%)	94 (30.82%)	66 (21.93%)	72 (30.90%)
5001–7000	200 (43.29%)	132 (43.28%)	131 (43.52%)	91 (39.06%)
7001–9000	139 (30.09%)	68 (22.30%)	95 (31.56%)	64 (27.47%)
> 9000	17 (3.68%)	8 (2.62%)	3 (1.00%)	1 (0.43%)
Marital status§					0.001
Married	317 (68.61%)	178 (58.36%)	221 (73.42%)	137 (58.80%)
Divorced	5 (1.08%)	2 (0.66%)	3 (1.00%)	2 (0.86%)
Widowed	140 (30.30%)	125 (40.98%)	77 (25.58%)	94 (40.34%)
Medical conditions
Tooth loss§					< 0.001
No tooth loss	88 (19.05%)	66 (21.64%)	62 (20.60%)	24 (10.30%)
Fixed prosthesis	52 (11.26%)	12 (3.93%)	29 (9.63%)	11 (4.72%)
Removable prosthesis	48 (10.39%)	21 (6.89%)	31 (10.30%)	23 (9.87%)
Tooth loss without repair	249 (53.90%)	179 (58.69%)	166 (55.15%)	147 (63.09%)
Edentulous	25 (5.41%)	27 (8.85%)	13 (4.32%)	28 (12.02%)
Comorbidities§
Stroke	290 (62.77%)	211 (69.18%)	181 (60.13%)	159 (68.24%)	0.059
Alzheimer’s disease	39 (8.44%)	35 (11.48%)	16 (5.32%)	21 (9.01%)	0.059
Parkinson’s disease	22 (4.76%)	20 (6.56%)	16 (5.32%)	15 (6.44%)	0.687
Amyotrophic lateral sclerosis	2 (0.43%)	1 (0.33%)	0 (0.00%)	3 (1.29%)	0.173
Multiple sclerosis	8 (1.73%)	4 (1.31%)	2 (0.66%)	2 (0.86%)	0.566
Head and neck tumors	29 (6.28%)	13 (4.26%)	41 (13.62%)	13 (5.58%)	< 0.001
Local radiotherapy for head and neck tumors	14 (3.03%)	9 (2.95%)	17 (5.65%)	5 (2.15%)	0.111
Coronary heart disease	62 (13.42%)	52 (17.05%)	39 (12.96%)	44 (18.88%)	0.133
Brain injury	8 (1.73%)	7 (2.30%)	7 (2.33%)	9 (3.86%)	0.384
Masticatory dysfunction	65 (14.07%)	55 (18.03%)	34 (11.30%)	55 (23.61%)	0.001

¶ MI minimal impairment, RD respiratory-dominant impairment, AD articulatory-dominant impairment, MT multimodal impairment, BMI body mass index, CNY Chinese Yuan

† Expressed as mean ± standard deviation

§ Expressed as case number (ratio) Bold font denotes statistical significance (P<0.05) for comparisons among clusters

Dysphagia outcomes

The relationship between dysphagia severity, assessed via VFSS, and phenotype characteristics was studied. The MI patients had a significantly lower PAS score (2.98 ± 2.44) compared to other phenotypes (P < .001), reflecting milder dysphagia. This aligns with the MI phenotype’s least impairment across all aspects. The overall impairment ranked in the order of MT > RD > AD > MI. The eating ability and the incidence of swallowing-related complications both differed significantly (P < .001) between phenotypes (Table 5).

Table 5.

Dysphagia outcomes

Characteristics	All (n = 1301)	MI¶ (n = 462)	RD¶ (n = 305)	AD¶ (n = 301)	MT¶ (n = 233)	P value¶
PAS scores†	4.82 ± 2.81	2.98 ± 2.44	5.46 ± 2.48	5.08 ± 2.51	6.58 ± 2.26	< 0.001
EAT-10 score†	5.18 ± 3.37	2.93 ± 2.29	6.78 ± 3.19	4.61 ± 2.47	8.39 ± 3.11	< 0.001
Incidence of complications §	549 (42.20%)	68 (14.72%)	202 (66.23%)	82 (27.24%)	198 (84.98%)	< 0.001

¶ MI minimal impairment RD respiratory-dominant impairment AD articulatory-dominant impairment MT multimodal impairment

The P value assessed the differences between the four phenotypes

† PAS, penetration-aspiration scale; EAT-10, eating assessment tool-10

Expressed as mean ± standard deviation

§ Expressed as case number (ratio) Bold font denotes statistical significance (P<0.05) for comparisons among clusters

Discussion

The severity of dysphagia directly affects the quality of life and leads to serious psychological and physical health problems [44]. There are a dozen swallowing assessment tools currently available (e.g., dysphagia outcome and severity scale, Gugging swallowing screen, water swallow test, standard swallowing assessment) [45], but they focus on swallowing in a rather specific manner and do not consider the interrelations of relevant muscle groups in the pathology of dysphagia [26]. Our findings, in line with prior studies, [28, 46] show that dysphagia rarely occurs in isolation. Given such complexity, a comprehensive multidimensional assessment of dysphagia patients is expected to facilitate diagnosis and treatment.

Through unsupervised k-means clustering, Chinese old adult patients with dysphagia were categorized into four distinct phenotypes that differed significantly not only in the swallowing function but also in the respiratory and articulatory functions. The worst impairment was found for MT patients, who mainly consumed a soft diet, had a loss of appetite, changed posture when eating, had weak swallowing-related muscles, and showed severe functional impairments of the oral cavity, pharynx, and esophagus. MT patients tended to be older and had lower levels of education and income. Their socioeconomic status negatively impacts dysphagia in two ways. First, aging is associated with a decline in swallowing physiology, exacerbating nutritional deficiencies. Second, lower income and education levels may restrict access to professional rehabilitation services, such as swallowing therapy or nutritional counseling. These findings align well with previous studies of community-dwelling elderly populations [47, 48]. In contrast, the MI patients were associated with more food intake, better appetite, stronger muscles related to swallowing, and less impairment of swallowing functions in all dimensions. Indeed, previous research has suggested that aging-related decline in swallowing function can be subtle and difficult to detect [49]. The RD and AD patients had moderate levels of impairment, but RD tended to be more impaired than AD. The overall level of impairment ranked as MT > RD > AD > MI. The outcomes, including the eating ability score and the occurrence of complications, matched the clustering result perfectly (Table 5). Among comorbidities, while neurogenic diseases like stroke and Parkinson’s were found in all clusters, only head and neck tumors and masticatory dysfunction differed across clusters. This result does not conform to previous studies [12] implying that traditional etiological classifications may not fully capture functional heterogeneity. Our machine-learning-based clustering reveals unique phenotypic subgroups that may correspond to distinct endotypes defined by common physiological patterns, potentially enabling personalized treatments targeting each phenotype’s specific physiological dysfunctions.

Previous studies suggested that the decline in respiratory function and abnormal breathing patterns during swallowing is a primary factor in the deterioration of eating ability and the occurrence of adverse outcomes [50–52]. In this work, the respiratory function played a key role in the differentiation of the phenotypes. MT and RD patients, with more severe dysphagia and higher PAS scores, showed severe respiratory dysfunction that raised aspiration risk. The manifestations included decreased vital capacity, increased respiratory rate, tachypnea during eating, etc. In contrast, AD and MI patients had only mild respiratory dysfunction. Thus, respiratory function is key in impairment assessment and therapeutic planning for different dysphagia subgroups.

Articulatory dysfunction has been reported as a predictor of disease progression and complication [4, 53]. Dysphonia and dysphagia co-occur frequently as they share some underlying pathological mechanisms, such as muscle tension, neurological disorders, and structural abnormalities [4, 17, 54] Both conditions involve the muscles of the larynx and pharynx, which play crucial roles in voice production and swallowing. For example, the intrinsic laryngeal muscles control the vocal cords’ tension and position, and the extrinsic laryngeal muscles stabilize and move the larynx during swallowing and speaking. In our study, both MT and AD phenotypes presented significant articulatory dysfunction. In the AD phenotype, patients showed severe dysarthria, with elevated jitter (1.85% ± 0.90%) and shimmer (4.10% ± 1.19%). A history of head and neck tumors was significantly more prevalent in the AD phenotype (13.62%) than in other clusters (≤ 6.28%), aligning with clinical evidence that laryngeal/pharyngeal tumors disrupt articulatory muscles and lead to dysarthria [55].

Personalized treatment strategies are likely to benefit patients of different phenotypes [46, 56]. For RD patients, the focus is on improving swallowing function, restoring normal breathing patterns during swallowing, and reducing respiratory dysfunction. For AD patients, the focus will be specialized treatments for articulation instead. For example, it is possible that, in the swallowing therapy, AD patients can benefit more from the Mendelsohn maneuver, whereas RD patients can benefit more from jaw thrust. Among the four phenotypes, MT patients had the most severe impairments, with significant problems in breathing, swallowing, and articulation, and a single treatment modality cannot address such multi-faceted impairments. Previous studies have shown that comprehensive interventions are more effective in improving the overall functioning in patients with multiple disorders [57, 58]. For example, respiratory support can improve their breathing function and reduce the risk of aspiration during swallowing [59]. Swallowing training can enhance the strength and coordination of swallowing muscles, thereby improving swallowing efficiency [60]. Articulation training can help improve voice quality and reduce the impact of dysphonia on quality of life [56, 61]. It is crucial for clinicians to adopt comprehensive approaches, including respiratory support, swallowing training, and/or articulation training, tailored to the specific functional impairments of each phenotype, to improve patient prognosis and quality of life.

The limitations of this study include the following. As a cross-sectional study, this work does not elucidate the long-term outcomes of different dysphagia phenotypes, particularly regarding functional decline and survival, and future prospective longitudinal studies are needed. Further clinical trials of personalized interventions are also needed to validate the accuracy and utility of the present findings. In terms of machine learning, there were some missing data when the multidimensional characteristics were collected, and to prevent bias in the eventual analysis, the random forest method was adopted to interpolate the missing values. In addition, we included the most straightforward and relevant variables in phenotyping the dysphagia patients, but it is still possible that some omitted variables are important and may sway the results. For example, we disregarded cognitive function and frailty, since these variables were uniform in our patients after the exclusion criteria were applied. Furthermore, although we followed standardized protocols, some assessments, such as muscle strength grading and respiratory patterns during swallowing, depended on subjective evaluations. Thus, despite offering valid clinical data, they might not have captured subtle functional impairments. Future studies should integrate multimodal objective assessments (e.g., nasoendoscopy, high-resolution manometry, and surface electromyography) to enhance the reliability and validity of muscle strength evaluation, and the integration of nasal airflow thermistors with VFSS will aid in precise pattern classification and personalized interventions. Nevertheless, this work still provides an essential reference in recognizing the diverse phenotypes of dysphagia in older adults.

Conclusions

Four phenotypes of dysphagia were recognized by symptom clustering for Chinese older adult patients, namely minimal impairment, respiratory-dominant impairment, articulatory-dominant impairment, and multimodal impairment. The phenotyping considered the respiratory, articulatory, and swallowing functions of patients and was accomplished using unsupervised machine learning. The key differences between the phenotypes were analyzed. The results may facilitate the diagnosis and treatment of dysphagia. Future longitudinal studies and treatment data are needed to validate the impact of the present phenotyping on patient outcomes.

Data Availability

Additional data may be available upon reasonable request, pending any necessary consultation with the Ethics Committee for approval. Requests to access the datasets should be directed to 15,106,205,726@163.com.

Abbreviations

VFSS

Videofluoroscopic swallow study

STROBE

STrengthening the Reporting of OBservational Studies in Epidemiology

BMI

Body Mass Index

PAS

Penetration-Aspiration Scale

G

Grade

R

Roughness

B

Breathiness

A

Asthenia

S

Strain

MPT

Maximum phonation time

F0

Fundamental frequency

Jitter

Fundamental frequency perturbation

Shimmer

Amplitude perturbation

HNR

Harmonic-to-Noise Ratio

SD

Standard Deviation

CNY

Chinese Yuan

WSS

Within-group sum of squares

MI

Minimal impairment

RD

Respiratory-dominant impairment

AD

Articulatory-dominant impairment

MT

Multimodal impairment

EAT-10

Eating assessment tool-10

Authors’ contributions

YZ and YD were co-primary investigators. YZ, YD, JC and HW designed the study, and contributed to interpretation of data. YZ, YD and JC wrote the first draft of the article. YZ, YD, HW and JC did the statistical analyses, data validation and interpretation of data. YZ, YD, HW, ZG,and JC were responsible for phenotype cleaning, preparation of the tables and figures, and provided further data interpretation. YZ and YD have accessed and verified the underlying data.YZ, YD, HW, ZG, and JC have accessed and verified the data. All authors had full access to all the data in the study, and had final responsibility for the decision to submit the manuscript for publication.

Funding

This research was funded by the Suzhou Science and Technology Bureau Project(Grant No. SKY2022120, SKY2023050, SSD2024079) and the Scientific Research Project of Chinese Nursing Association (Grant No. ZHKY202402). The funders were not involved in the design study, the collection, analysis, interpretation of data, or the decision to submit the manuscript for publication.

Data availability

Additional data may be available upon reasonable request, pending any necessary consultation with the Ethics Committee for approval. Requests to access the datasets should be directed to 15106205726@163.com.

Declarations

Ethics approval and consent to participate

This study was approved by the Hospital ethics committee in Suzhou (No. 2024271). Written informed consent was obtained from all recruited patients.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.