Multimodal Neural Assessment for Predicting Post-Stroke Dysphagia: Identifying Critical Risk Factors Using Combined Indicators

Introduction

Post-stroke dysphagia (PSD) is a common and debilitating complication of strokes, which significantly reduces patients’ quality of life [1]. Dysphagia is associated with serious health consequences, such as aspiration pneumonia, malnutrition, and increased mortality [2,3]. Early and accurate detection of PSD is critical to improving patient outcomes; however, conventional diagnostic methods such as video fluoroscopic swallowing study (VFSS) and flexible endoscopic evaluation of swallowing (FEES) are limited in their ability to detect PSD early and noninvasively [4,5]. These traditional methods primarily focus on the physical process of swallowing, but they do not assess the underlying neural mechanisms, limiting their diagnostic accuracy in the early stages of PSD development.

While several risk factors for PSD, including age, stroke location, and comorbidities, have been identified, current diagnostic approaches do not provide a comprehensive understanding of the neural dysfunctions associated with it [6,7]. Although clinical scales such as the National Institutes of Health Stroke Scale (NIHSS) and the Modified Barthel Index (MBI) are useful, they do not offer direct insights into cortical activity and motor control relevant to swallowing function [8,9]. Therefore, there is an urgent need for more advanced, noninvasive, and accurate methods for early PSD prediction.

Recent advances in neurophysiological diagnostic techniques – specifically, functional near-infrared spectroscopy (fNIRS), transcranial magnetic stimulation (TMS), and surface electromyography (sEMG) – offer a promising multimodal approach. fNIRS measures cortical oxygenation and provides real-time information about brain activity, thus revealing important neural pathways involved in motor control and sensory processing relevant to swallowing [10]. TMS evaluates motor cortex excitability and cortical plasticity, which are critical for understanding swallowing dysfunction after stroke [11]. sEMG provides data on muscle activity and coordination during swallowing, allowing the assessment of muscle function and coordination in real time [12].

According to Sasegbon et al [13], swallowing function is governed by a multilayered neural control system that includes the cortical swallowing network, the brainstem swallowing central pattern generator, and peripheral nerves. With the recent development of multimodal assessment techniques, fNIRS, TMS brain function testing, and sEMG are increasingly used in clinical practice to assess brain neuronal regulation and the neurophysiological processes of stimulus response.

The combination of TMS [14–16], sEMG [17,18], and fNIRS [19–21] offers a comprehensive view of the neural and muscular mechanisms underlying PSD. While scholars have explored individual biomarkers or diagnostic modalities separately, this multimodal combination provides a more holistic approach by simultaneously evaluating cortical function, motor excitability, and muscle activity. The integration of these methods is particularly valuable as it allows for early detection of PSD at multiple levels – from brain activity to muscle coordination – thus enhancing both the accuracy and early intervention potential of the diagnostic process.

Therefore, this retrospective study of 300 stroke patients aimed to evaluate risk factors in patients with and without post-stroke dysphagia using functional near-infrared spectroscopy (fNIRS), transcranial magnetic stimulation (TMS), and surface electromyography (sEMG).

Material and Methods

ETHICS STATEMENT:

This study was approved by the Ethics Committee of Guizhou Medical University Affiliated Hospital (no. 2023029K), and it was registered with the Chinese Clinical Trial Registry (no. ChiCTR2200065153). Written informed consent was obtained from each patient before the study began.

PARTICIPANTS:

The inclusion criteria were based on the World Health Organization’s ICD-10 diagnostic codes I63 Cerebral Infarction or I61.1 Intracerebral Hemorrhage in the Cerebral Hemisphere, Cortical, as well as on CT or MRI reports showing unilateral hemispheric stroke lesions. Retrospective data collection was conducted from the inpatient electronic medical record system of the Department of Rehabilitation Medicine at Guizhou Medical University Affiliated Hospital for all eligible cases between September 2023 and December 2024. A total of 482 patients were initially identified; further screening was carried out by 2 researchers. Binary classification using +/− was performed to identify cases with dysphagia due to unilateral brain lesions (+) or without dysphagia (−). The inclusion criteria for dysphagia cases were as follows: a first-time unilateral cerebral hemisphere lesion confirmed by cranial CT or MRI, with dysphagia detected during hospitalization through VFSS or FEES, which identified pharyngeal or oropharyngeal swallowing dysfunction; age >18 years; disease duration between 2 and 24 weeks; a Mini-Mental State Examination (MMSE) score ≥24 on admission; and having undergone fNIRS, sEMG, and TMS on admission. The inclusion criteria for non-dysphagia cases were as follows: a first-time unilateral cerebral hemisphere lesion evaluated as Grade 1 on the water-swallowing test during hospitalization; age >18 years; disease duration between 2 and 24 weeks; an MMSE score ≥24 on admission; and having undergone functional brain assessments, including fNIRS, sEMG, and TMS, on admission. The following exclusion criteria were chosen: incomplete case information; a stroke involving other brain regions (eg, brainstem, cerebellum); the presence of other neurological disorders (eg, intracranial tumors, multiple sclerosis, intracranial infections); mental disorders or cognitive dysfunction; a history of head, neck, or esophageal surgery; contraindications for TMS such as the presence of a cardiac pacemaker or other metal implants; a history of epilepsy; and recurrent admissions. After excluding individuals with disease durations exceeding 24 weeks, metal implants, and preexisting neurological disorders from the initial study cohort, a total of 300 patients were included in the final cohort (Figure 1). These patients underwent assessments with fNIRS, TMS, sEMG, NIHSS, and MBI, and there were no missing data.

SAMPLE SIZE:

This study employed binary logistic regression analysis to evaluate the impact of multiple independent variables (eg, age, sex, disease duration) on the dependent variable (presence or absence of dysphagia following stroke). Based on relevant principles [22], the following sample size calculation formula was used:

where n was the required sample size per group; Z_β was the Z value corresponding to the desired statistical power (0.84 for 80% power), and Z_α/2 was the Z value corresponding to the alpha level (1.96 for a two-tailed test at the 0.05 significance level). A literature review indicated that the incidence of PSD ranges from 35% to 78%, with an average value of 56.5% used for the P value. A preliminary pilot study conducted at our institution found that the proportion of patients with post-stroke dysphagia (P₀) was 45.0%. Therefore, by using the formula above, we determined that a minimum of 124 participants per group would be required to achieve a statistical power of 80%, with an alpha level of 0.05 and a medium (0.5) effect size (Cohen’s d). Given that the study involved 300 patients, with the observational group (those with dysphagia) being approximately twice the size of the control group (those without dysphagia), this sample size provided adequate statistical power for the analysis, despite the imbalance between the groups. The larger observational group helped to mitigate potential bias introduced by the unequal distribution, thereby ensuring that the study maintained sufficient power to detect meaningful differences.

FNIRS DATA ACQUISITION AND SWALLOWING TASK: To obtain fNIRS data on swallowing function during rest and task states, we utilized an fNIRS system with near-infrared light wavelengths of 730 and 850 nm (NirScan, Danyang Huichuang Medical Equipment Co. Ltd., Danyang, China). The system collected data at a sampling rate of 10 Hz, recording cortical activation associated with swallowing. The system’s signal sources consisted of 14 light-emitting probes and 14 detectors, arranged according to the 10–20 system to form 35 channels, with each probe spaced 3 cm apart. The regions of interest (ROI) for detection included the left/right dorsolateral prefrontal cortex (L/R PFC), middle dorsolateral prefrontal cortex (MPFC), left/right primary motor cortex (L/R M1), and left/right primary somatosensory cortex (L/R S1). The fNIRS probe configuration is shown in Figure 2A. The optodes were positioned based on 4 reference points (nasion, central zero, left, and right preauricular points) using the 10–20 system. Spatial registration was performed using a 3D localization system to match the probe positions to Brodmann areas. To minimize environmental noise, the participants sat in a quiet fNIRS evaluation room, with a 5-min rest period. Each optical signal was secured using a custom hard-plastic cap and shielded with black cloth to prevent contamination from ambient light. In our previous studies, we established a standardized swallowing function protocol [23]. First, the participants were instructed to close their eyes and relax, without falling asleep; they rested for 5 min to collect resting-state data. The swallowing-task protocol involved alternating 30-s resting periods and 15-s swallowing periods 3 times, under the guidance of a computerized voice prompt. The completion of each swallowing action was confirmed by observing full laryngeal elevation. To minimize artifacts from oral movements, the participants practiced the swallowing motion prior to the study. The entire procedure lasted approximately 15 min (Figure 2B).

FNIRS DATA ANALYSIS: The fNIRS data were automatically preprocessed using NirSpark software (NirScan, Danyang Huichuang Medical Equipment Co. Ltd., Danyang, China), which operates within MATLAB (MathWorks, USA). Based on established data processing protocols [23,24], the analysis included the following steps: setting the signal standard deviation threshold to 6 and the peak threshold to 0.5; removing motion artifacts through spline interpolation; filtering physiological noise (heartbeat, respiration, and Mayer waves) with a 0.01–0.1 Hz band-pass filter; converting the filtered optical density data to oxyhemoglobin concentration using the modified Beer-Lambert law [25]; and setting the hemodynamic response function time window from 0 to 15 s, which corresponded to a single swallowing-task blocking paradigm. The data of oxyhemoglobin concentration (HbO) from the 3 blocking paradigms were superimposed and averaged; this yielded a 15 s mean HbO during the swallowing task, which represented the activation level in each ROI. To account for variability in lesion localization, the data from the patients with left-sided lesions were mirrored, with the right hemisphere defined as the lesion side after mirroring. For the analysis of complex brain network topological properties, a random network was generated matching the actual network in terms of node count, edge count, and degree distribution to assess the reliability of the real network. The clustering coefficient (Cp) was used to quantify the local connectivity and community structure of a brain network based on resting-state data. This coefficient serves to measure the extent of functional segregation within a brain’s functional network, which reflects the network’s capacity for specialized processing of specific types of information. The formula for calculating is as follows [26]:

where N represents the number of nodes in the unweighted network G; D_i is the number of edges connected to the i-th node, and E_i denotes the number of edges in the subnetwork.

TMS BRAIN FUNCTION TESTING: This assessment was performed with the CCY-I-type magnetic stimulator coil manufactured by Wuhan Yirui Medical Equipment Co., utilizing a figure-eight coil (d=7 cm) for stimulation. The patient was placed in either a supine or seated position, with surface recording electrodes applied to the muscle belly of the bilateral suprahyoid muscles. The ground electrode was attached to the bilateral mastoid processes. The TMS coil was positioned 2–4 cm anterior to the vertex of the skull and moved within a 4–6 cm range toward the contralateral hemisphere, which corresponds to the motor cortex representation area of the suprahyoid muscles [14]. Stimulation began at 30% of the maximum output intensity, with single-pulse stimuli gradually increasing in strength. The point at which the shortest MEP latency and the largest amplitude were observed was identified as the motor cortex representation area of the contralateral suprahyoid muscles, commonly referred to as the hot spot. Following the same procedure, the hot spot for the contralateral hemisphere’s suprahyoid muscle motor cortex representation area was determined. Once the hot spot was identified, the TMS intensity was adjusted to 100% and gradually reduced. The minimal magnetic stimulation intensity that could evoke motor-evoked potential (MEP) with an amplitude exceeding 50 μV in the suprahyoid muscles in the resting state for 5 out of 10 stimulations was defined as the resting motor threshold (RMT) for that hemisphere. According to prior research, unilateral TMS stimulation can induce MEPs in both the ipsilateral and contralateral suprahyoid muscles, with the side showing the larger amplitude being considered the primary MEP (typically, the contralateral side shows a higher amplitude than the ipsilateral side) and with stimulation intensities ranging from 90% to 140% of the RMT [27]. Therefore, at 120% of the RMT, MEPs were elicited from the hot spots of both hemispheres, and the maximum amplitude of the MEP was measured 3 times. The average of these maximum amplitudes was used to calculate the amplitude of the ipsilateral motor-evoked potentials (ipsilateral MEP) and the amplitude of the contralateral motor-evoked potentials (contralateral MEP). If no response was elicited, the amplitude for that side was recorded as 0 μV [27].

SEMG: The sEMG signals from the bilateral suprahyoid muscles were recorded using the MegaWin 3.0 data acquisition system (ME6000 sEMG System, Mega Electronics Ltd., Kuopio, Finland) based on our previous study [23]. The system specifications included a measurement range of ±8192 μV, with an accuracy of ±2% for input below 4000 μV and ±5% for input up to 8192 μV. The sampling rate was 1000 Hz, with a sensitivity and resolution of 1 μV and an input impedance of 20 GΩ. Regarding electrode placement, 2 pairs of disposable Ag/AgCl electrodes (1 cm in diameter) were positioned along the posterior one-third of the suprahyoid muscles at the midline, spaced 22±1 mm apart. The reference electrode was placed on the ipsilateral mastoid [27]. Prior to electrode application, excess hair was shaved, and the skin was cleaned with an alcohol wipe. After verifying signal quality, the participants were asked to swallow saliva and repeat the procedure 3 times. Two parameters were analyzed from the sEMG data of the suprahyoid muscles. The first one was swallowing duration (SD), which was defined as the interval between the onset of swallowing (when the sEMG signal exceeded the baseline by more than 2 standard deviations within 20 ms) and the offset of swallowing (when the sEMG signal dropped below the baseline by more than 2 standard deviations within 20 ms). The duration was measured in seconds (s). The second parameter was peak amplitude (PA), which was defined as the highest sEMG value recorded during the swallowing period, measured in microvolts (μV) [28]. The mean values of SD and PA were calculated for the bilateral suprahyoid muscles.

NIHSS: The NIHSS includes assessments of consciousness, gaze, visual fields, facial palsy, upper- and lower-limb movement, ataxia, sensation, language, speech disorders, and neglect. A higher score indicates more severe neurological deficits [29].

MBI:

The MBI includes assessments of bowel control, bladder control, grooming, toileting, feeding, transfer, ambulation, dressing, stair climbing, and bathing. A lower score indicates a higher level of dependence in daily living activities [30].

STATISTICAL ANALYSIS:

Based on methodological guidelines, a discussion group composed of a physician, speech therapist, and radiologist independently reviewed all the medical records and extracted the data to ensure consistency and reliability in data collection [31]. Data on the following aspects were collected for all the patients: sex, age, disease duration, stroke type, lesion side, stroke location, medical history (hypertension and diabetes), MBI score, NIHSS score, Cp, PA, SD, ipsilateral MEP, contralateral MEP, and average HbO in all ROIs. The data were entered into Microsoft Excel spreadsheets. Regarding descriptive statistics, continuous variables that followed a normal distribution were expressed as means±standard deviations (SDs), while non-normally distributed continuous variables were expressed as medians (Ms) (interquartile ranges [IQRs]). Categorical variables were described as frequencies (percentages). Pearson’s chi-square test or Fisher’s exact test was used to compare categorical variables, while ANOVA or the Kruskal-Wallis test was employed for continuous variables. Dysphagia was considered the primary outcome variable. Chi-square tests were utilized to assess the relationship between the categorical variables and the outcome variable, and ANOVA was used to assess the relationship between the continuous variables and the outcome variable. To avoid missing the influence of important independent variables, factors with P<0.1 were subjected to variance inflation factor (VIF) testing to enhance the stability and reliability of the coefficient estimates in the logistic regression model. Factors with a VIF value <5 were included in the binary logistic regression analysis to evaluate predictive risk factors. The results were reported as odds ratios (ORs), 95% confidence intervals (CIs), and P values. Receiver operating characteristic (ROC) curves were used to calculate the area under the curve (AUC) to assess the ability of risk factors to predict PSD. The predictive ability was classified based on the AUC values as follows: 0.5< AUC <0.7 indicated very low predictive value; 0.7< AUC <0.9 signaled moderate predictive value, and 0.9< AUC <1.0 indicated very high predictive value [32]. The DeLong test was used to compare AUCs and further distinguish the predictive abilities of different diagnostic indicators. The optimal cutoff value was set as the value corresponding to the maximum Youden index (sensitivity+specificity - 1). All the analyses were performed using SPSS Statistics version 25.0, with statistical significance defined as a two-tailed P<0.05.

Results

BASELINE CATEGORICAL VARIABLE CHARACTERISTICS OF PARTICIPANTS IN THE 2 GROUPS:

A total of 300 stroke patients were included in this study, 99 in the non-dysphagia group and 201 in the dysphagia group. Fisher’s exact test comparison showed no significant differences between the 2 groups in terms of sex, stroke type, lesion side, hypertension, and diabetes (P>0.1, Table 1). However, the non-dysphagia group had a lower proportion of patients aged >65 years, with an MBI score <45 and lesions located in both the cerebral cortex and deep-brain regions, compared to the dysphagia group (P<0.1, Table 1).

BASELINE CONTINUOUS VARIABLE CHARACTERISTICS OF PARTICIPANTS IN THE 2 GROUPS:

The Kruskal-Wallis test revealed that there were no significant differences between the 2 groups in terms of disease duration and the mean HbO levels of RM1 (P>0.1). However, the non-dysphagia group had higher Cp, PA, SD, ipsilateral MEP, contralateral MEP, MPFC, and mean HbO values of all ROIs on the healthy side compared to the dysphagia group (P<0.1). The non-dysphagia group also had lower NIHSS scores, RPFC values, and mean HbO values of RS1 compared to the dysphagia group (P<0.1, Table 2).

MULTIVARIATE LOGISTIC REGRESSION ANALYSIS OF RISK FACTORS FOR PSD PATIENTS:

The factors with significant differences were subjected to VIF testing. The VIF coefficients for age, stroke location, MBI, NIHSS, Cp, PA, SD, ipsilateral MEP, and contralateral MEP were all <5, indicating that these variables were non-collinear (Table 3) and met the conditions for logistic regression analysis. After assigning values to these variables (Table 4), they were included in the multivariate binary logistic regression analysis. The results show that age >65 years, the stroke location lobar+deep, an MBI score <45, the NIHSS score, the Cp, PA, ipsilateral MEP, and contralateral MEP were independent risk factors for the occurrence of dysphagia after stroke. A forest plot was used to visualize the effect sizes of these factors in the logistic regression analysis (Figure 3).

PREDICTIVE VALUES OF PSD-RELATED RISK FACTORS FOR PSD PATIENTS:

According to the principles of ROC curve plotting, the continuous variables from the independent risk factors for PSD (the NIHSS score, the Cp, PA, ipsilateral MEP, and contralateral MEP) were selected and combined with fNIRS-TMS-sEMG detection indicators (Cp+PA+ipsilateral MEP+contralateral MEP, Cp+PA+ipsilateral MEP, Cp+PA+contralateral MEP, Cp+PA, Cp+ipsilateral MEP, Cp+contralateral MEP, PA+ipsilateral MEP, and PA+contralateral MEP) as covariates to jointly plot the ROC curve. The findings show that the AUC for predicting PSD using the combination Cp+PA+ipsilateral MEP+contralateral MEP was 0.985 (P<0.001, 95% CI 0.974–0.995), with a maximum Youden index of 0.880. The corresponding cutoff value was 0.737, with a sensitivity of 92.0% and specificity of 96.0%. The AUC for predicting PSD using Cp+PA+ipsilateral MEP was 0.939 (P<0.001, 95% CI 0.915–0.964), with a maximum Youden index of 0.725. The corresponding cutoff value was 0.759, with a sensitivity of 80.6% and specificity of 91.9%. The AUC for predicting PSD using PA+ipsilateral MEP was 0.934 (P<0.001, 95% CI 0.908–0.959), with a maximum Youden index of 0.701. The corresponding cutoff value was 0.855, with a sensitivity of 72.1% and specificity of 98.0%. The AUCs for the remaining variables were all <0.9, which indicates relatively low predictive values for PSD. These results suggest that the combined fNIRS-TMS-sEMG indicators had higher predictive values for PSD compared to the other combined variable sets and individual variables. Detailed findings are shown in Table 5 and Figure 4. Furthermore, the DeLong test was used to pairwise compare variables with an AUC >0.9 to distinguish their predictive capacities. The results show that the AUC for predicting PSD using the combination Cp+PA+ipsilateral MEP+contralateral MEP was significantly higher than that for Cp+PA+ipsilateral MEP (Z=4.029, P<0.05) and PA+ipsilateral MEP (Z=4.245, P<0.05). There was no significant difference in AUC between Cp+PA+ipsilateral MEP and PA+ipsilateral MEP (Z=1.058, P=0.290), as shown in Table 6.

Discussion

STUDY LIMITATIONS:

This study has some limitations. First, it was a retrospective study, which may have introduced selection bias. In our situation, local healthcare policies restrict dysphagia treatment to tertiary hospitals, and this could have resulted in a higher proportion of PSD patients at our institution. Additionally, a single-center study may limit the results’ generalizability. Furthermore, due to limited access to medical records, we included fewer influencing factors, which may have excluded other factors related to dysphagia (eg, cognitive impairment, speech disorders, and smoking and alcohol consumption) as well as other serological markers (eg, white blood cell count, C-reactive protein). Therefore, large-scale, multi-center prospective studies are needed to examine more comprehensive case data and variable indicators.

Conclusions

This retrospective study of 300 stroke patients demonstrates that age >65 years, cortical and deep-brain strokes, an MBI score <45, the NIHSS score, the Cp, PA, ipsilateral MEP, and contralateral MEP are independent risk factors for the occurrence of PSD by using functional near-infrared spectroscopy (fNIRS), transcranial magnetic stimulation (TMS), and surface electromyography (sEMG). The combined use of fNIRS, TMS, and sEMG indicators is a feasible and effective approach for predicting the onset of dysphagia, as it offers higher predictive value compared to individual diagnostic methods and clinical scales. Among the combined predictors, Cp+PA+ipsilateral MEP+contralateral MEP shows the highest predictive accuracy, highlighting the clinical applicability of these combined indicators for the early identification and prevention of PSD. These findings support the potential of multimodal diagnostic tools for improving the management of PSD.