Machine learning-based model for prediction of swallowing recovery in post-stroke patients

Abstract

Background

Post-Stroke Dysphagia (PSD), as a common complication of cerebrovascular accidents, seriously affects patients’ quality of life and prognosis. This retrospective study aims to provide a reliable machine learning model for predicting the prognostic factors of PSD.

Methods

The clinical data from patients admitted to the Fourth Affiliated Hospital of Soochow University from January 2021 to December 2023 were collected. The dataset was chronologically split into a training set (January 2021-May 2023, n = 377) and a temporal validation set (June-December 2023, n = 91). Feature variables selection was performed using correlation analysis and logistic regression. Five machine learning models were developed and evaluated and subsequently evaluated on the temporal validation set using precision, recall, F1-score, and area under the receiver operating characteristic curve (AUC).

Results

Eight feature variables were selected for model construction. On the temporal validation set, the Gradient Boosting Decision Tree (GBDT) demonstrated the best performance with a precision of 0.937, recall of 0.795, F1-score of 0.863, and an area under the receiver operating characteristic curve (AUC) of 0.940. During 5-fold cross-validation on the training set, Random Forest (RF) achieved the highest average AUC (0.949), but its performance decreased to 0.890 on the temporal validation set. The superior performance of GBDT on the temporal validation set indicates its stronger generalization capability compared to other models.

Conclusions

The GBDT model showed robust performance on the temporal validation set, suggesting its potential clinical utility for predicting swallowing function recovery in post-stroke patients.

Introduction

Stroke, characterized by its high incidence, disability rate, recurrence rate, and mortality rate, has become the main cause of death and disability among adults in China [1]. Post-stroke dysphagia (PSD) is a prevalent complication that significantly affects the quality of life and prognosis of patients [2–4].Several studies have reported dysphagia in 30%-50% of stroke patients [5, 6]. Despite a favorable recovery rate from dysphagia within the first month post-stroke, approximately 13%-18% of patients continue to experience persistent dysphagia up to 6 months following the onset of stroke [7]. Patients are at increased risk of inadequate oral intake, malnutrition, dehydration, aspiration, stroke-related pneumonia and even asphyxiation [8–11]. Therefore, the prediction of dysphagia duration, in conjunction with prognostic factors related to rehabilitation and preventive interventions, holds the potential to enhance the recovery of swallowing function.

Previous studies have identified factors affecting the recovery of PSD, including age [12], activities of daily living (ADL), hyperlipidaemia, non-cardioembolic stroke, national institutes of health stroke scale (NIHSS) [13], oral hygiene [14, 15], muscle mass [15, 16], cognitive level [17], dysphagia severity scale (DSS) and body mass index (BMI) [18]. There were significant variations in, the study population, inclusion/exclusion criteria, swallowing function assessment methods, variable settings, model construction methods, interventions and outcome indicators. In response to the considerable size, high dimensionality and structural complexity of medical data, the conventional statistical approach of constructing predictive models based on logistic regression has become inadequate for data analysis.

Machine Learning (ML) uses computer algorithms to build models automatically, learn patterns from data through an autonomous learning process, and make accurate predictions based on these pattern [19]. Compared with traditional logistic regression models, ML can not only better handle multicollinearity in large, high-dimensional data (albeit often at the cost of reduced interpretability) but also select feature variables in a way that is less affected by subjective factors, showing great potential in the diagnosis, classification, and prognosis of diseases.

Therefore, this study aims to explore the predictive factors related to the recovery of swallowing function from five aspects: demographic information, clinical indicators, laboratory test indicators, assessment metrics, and rehabilitation treatment measures, construct several post-stroke swallowing function prediction models based on machine learning, and select the optimal model through internal and temporal validation of each model. This study could potentially provide a digital tool for future clinical personalised selection of nutritional regimens and early swallowing rehabilitation interventions, with the aim of improving swallowing function prognosis.

Methods

Study design and source of data

The study was reviewed and approved by the ethics committee of the host institution before the investigation was carried out. Subjects are exempt from signing the informed consent form as it was a retrospective study without additional tests and treatment manipulations. In this study, consecutive datasets of stroke patients were included as an internal training set for constructing the prediction model from January 2021 to May 2023 in the Fourth Affiliated Hospital of Soochow University. The dataset of stroke patients from June 2023 to December 2023 was collected as a temporal validation set to validate the model. Inclusion criteria: (1) patients (> 18 years old) diagnosed with stroke by cranial CT, MRI and other imaging examinations [20]; (2) patients with swallowing disorders assessed by fluoroscopic swallowing function examination and fibreoptic endoscopic swallowing function examination for the first time after admission to the hospital; (3) sustained rehabilitation treatment for ≥ 2 weeks. Exclusion criteria: (1) exclusion of other causes of dysphagia; (2) patients with long-term dysphagia; (3) patients with cognitive or psychiatric disorders who were unable to cooperate with the completion of the assessment; (4) incomplete medical records; the study was reviewed and approved by the ethics committee of the host institution prior to the implementation of the survey.

The medical records of stroke patients included demographic information, clinical examination data, and laboratory test indicators. A total of 27 feature variables were collected and included in the analyses as follows: (1) demographic variables: age, gender, and duration of disease (defined as the time in days from stroke onset to the initial swallowing assessment at hospital admission); (2) clinical variables: BMI, stroke type, stroke localisation (Side of stroke injury, stroke site), surgical history (recorded as a binary variable: Yes/No), dysarthria, artificial airway, hypertension, diabetes mellitus, pulmonary infection; (3) laboratory variables: haemoglobin, white blood cell count, albumin, total protein, triglycerides, blood glucose, D-dimer, fibrinogen. These were measured from routine fasting blood tests performed within 24 h of hospital admission; (4) assessment indicators: modified Bathel Index (MBI), water swallow test, penetration aspiration scale (PAS), modified bariums wallow impairment profile (MBSImP); (5) rehabilitation measures to improve swallowing function: electrical stimulation (20 min/session, 1 session/day, 5 sessions/week), repetitive transcranial magnetic stimulation (rTMS) (the motor cortex representative area of the hyoglossus muscle on the affected side, 5 Hz, 20 min/session, 1 session/day, 5 sessions/week) and surface EMG biofeedback (20 min/session, 1 session/day, 5 sessions/week). Swallowing function recovery assessment criteria, meet the Kubota drinking test = grade 1, VFSS or FEES review results show that the patient swallows without aspiration or choking, the nasal feeding tube can be removed, and complete oral feeding is resumed without the need for special food preparation.

This study conducted standardized training for data collectors during the quality control phase, rigorously adhered to the inclusion and exclusion criteria for data collection, entered data by two researchers and cross-checked it to ensure the assessments were reliable and accurate. Assessments were regularly conducted during the patient’s hospital stay and through regular outpatient or remote follow-up within 6 months after discharge; MBI assessments, the initial VFSS/FEES, and related swallowing scale evaluations were completed upon admission. Swallowing function recovery assessment criteria: Assessed by the therapist, meeting the criteria of Yanagida’s water drinking test = level 1, and the VFSS or FEES re-examination shows that the patient does not aspirate or choke while swallowing, allowing for the removal of the nasogastric tube, full recovery to oral feeding, and no need for specially prepared food.

Data preprocessing

The raw dataset underwent preprocessing involving data cleaning, imputation, and transformation. Samples missing > 30% of features and features missing > 20% of samples were excluded. Remaining missing values were imputed using the k-nearest neighbors method, which estimates missing entries based on similar cases. Categorical variables were converted into numerical format using integer label encoding, implemented with the factorize() function in Python. For ordinal variables such as the Water Swallow Test grade, the integer assignments preserved the natural order of the categories. Continuous variables remained in their original numerical form. Associations among features were evaluated using Spearman rank correlation coefficient, and the correlation matrix was visualized with a heatmap.

Model development and validation

(1) Characteristic variable screening: Variables pre-selected through correlation analysis were subjected to univariate screening; those demonstrating statistical significance (p < 0.1) were subsequently processed using Lasso-regularized logistic regression and elastic net-logistic regression for further characteristic screening to identify favourable and unfavourable factors in the feature variables. (2) Predictive models were constructed using five machine learning algorithms—L₂-regularized logistic regression (L₂-NLR), support vector machine (SVM), random forest (RF), neural networks (NN), and gradient boosting decision tree (GBDT)—based on feature variables selected from the training set. (3) Predictive model validation: The performance of the constructed predictive model is comprehensively evaluated using the validation dataset, and the indicators analysed include precision rate, recall rate, F1 score, AUC value, and receiver operating characteristic curve (ROC), the area under the ROC curve, whose value is larger to indicate the better performance of the model.

Statistical analysis

Continuous variables were described by mean ± standard deviation (SD) or median and interquartile range (IQR), while categorical variables were described by counts and percentage. In the analysis of correlation, Pearson correlation coefficient is used to assess the correlation of continuous variables that follow a normal distribution; for continuous data that do not follow a normal distribution or categorical variables, Spearman rank correlation coefficient is used to assess correlation. In order to maximally incorporate potential predictive variables, conditions with statistical significance differences were set at p < 0.1 in this study; all models, validations, and evaluation metrics used were implemented using the Sklearn package in Python.

Results

Characteristics of study population

The study finally included 462 patients in the training set and 117 patients in the temporal validation set. According to the exclusion criteria, 111 patients were excluded, and 468 patients were finally included. Of these, 377 patients were included in the training set, a total of 236 (62.59%) patients with recovery of swallowing function and 141 (37.4%) patients with non-recovery of swallowing function. In the temporal validation set, a total of 91 patients were included, with 47 patients (51.65%) recovering swallowing function and 44 patients (48.35%) not recovering swallowing function. The specific screening process and groupings are shown in Fig. 1.

Fig. 1.

Sample inclusion and exclusion process

Data pre-processing and variable screening

Based on the training set, the correlation analysis among the feature variables was performed and the correlation coefficient matrix was plotted (Fig. 2). Among them, there were 4 sets of feature variables with strong correlation. Combined with clinical knowledge and practice, the 4 variables of total protein, haemoglobin, lung infection and BMI were excluded from this study. All patients in this study were given electrical stimulation, so this variable has been excluded.

Fig. 2.

Correlation matrix plotof feature variables. The colour or intensity of the cells represents the strength and direction of the linear relationship between each pair of variables, the darker the colour of the cells in the graph the stronger the corresponding positive correlation and vice versa. The diagonal of each variable shows a perfect correlation with itself. Abbreviations: TOS, Types of Stroke; SOSI, Side of stroke injury; SOS, Site of Stroke; SH, Surgical History; AA, Artificial airway; HTN, Hypertension; T2DM, Type 2 Diabetes Mellitus; LI, Lung infections; Hb, Haemoglobin; WBC, White blood cell; ALB, albumin; TP, Total Protein; TG, Triglyceride; BG, Blood Glucose; Fib, fibrinogen; PAS, Penetration-Aspiration Scale; MBSI mP, Modified bariums wallow Impairment Profile; MBI, modified Bathel Index; WST, Water Swallowing Test; rTMS, repetitive transcranial magnetic stimulation; sEMGBF, surface electromyographic biofeedback; DOD, Duration of disease; BMI, Body Mass Index

The remaining 23 feature variables were included in the univariate analysis, and a total of 16 variables were initially screened as statistically different between the recovered and non-recovered groups (p < 0.1). The mean age of the individuals in the training set was (65.43 ± 13.41) years, comprising 271 males (71.88%). The average duration of illness was (59.11 ± 47.15) days. Among the patients, 236 (62.59%) regained swallowing function, whereas 141 (37.41%) did not. Ischemic stroke occurred in 108 (28.64%) cases, and hemorrhagic stroke was present in 269 (71.36%) cases. Detailed demographic and clinical characteristics for both the recovery and non-recovery groups within the training set are presented in Table 1. The Lasso-Logistic regression model and the elastic net-logistic regression model were used in the follow-up study to perform further screening of the 16 feature variables after the initial selection. The Lasso-Logistic regression model (λ = 0.306) selected five variables: age, artificial airway, MBSImP, MBI, and rTMS (Fig. 3A-B). The elastic net-logistic regression model (λ = 0.642) selected eight variables: age, stroke site, dysarthria, albumin, MBI, MBSImP, and rTMS (Fig. 3C-D). Univariate analysis had indicated significant differences in stroke site, dysarthria, and albumin between the groups (p < 0.1). To construct a model that balances predictive performance with clinical utility, we prioritized the feature set from the elastic net-logistic regression model. This decision was based on, on the one hand, its good predictive performance as demonstrated by cross-validation, and on the other hand, its inclusion of a broader range of clinically relevant predictors, such as stroke site, albumin, dysarthria, MBI, and rTMS. This comprehensive predictor set enhances the clinical interpretability of the model results. Consequently, the eight variables selected by the elastic net-logistic regression model were used for constructing subsequent predictive models.

Table 1.

Univariable analysis of the training set

variable	recovery team (N = 236)	Unrecovered group (N = 141)	P-value
Gender (n, %)			0.085*
Male	152 (64.41)	119 (84.4)
Female	84 (35.59)	22 (15.6)
Age (years, MD [IQR])	61 [54,72]	69 [56,73]	0.012*
Duration of disease (days, MD [IQR])	54 [21,61]	67 [26,72]	0.863
Type of stroke (n, %)			0.614
Ischaemic	64 (27.12)	44 (31.21)
Haemorrhagic	172 (72.88)	97 (68.79)
Stroke localisation
Side of stroke injury (n, %)			0.647
Left side	31 (13.14)	18 (12.76)
Right side	38 (16.10)	15 (10.63)
Bilateral	167 (70.76)	108 (76.60)
Site of Stroke (n, %)			0.081*
Cerebral hemisphere	115 (48.72)	79 (56.03)
Thalamus	33 (13.98)	26 (18.43)
Brainstem	56 (23.72)	48 (34.04)
Cerebellum	47 (19.92)	48 (34.04)
Basal ganglia	54 (22.88)	34 (24.11)
History of surgery (n, %)			0.660
Yes	61 (24.77)	45 (31.91)
No	175 (75.23)	96 (68.09)
Dysarthria (n, %)	94 (39.83)	82 (58.15)	< 0.001*
Artificial airway (n, %)	35 (14.83)	52 (36.87)	0.067*
Type 2 diabetes (n, %)	87 (36.86)	47 (33.33)	0.098*
Hypertension (n, %)	171 (72.46)	112 (79.43)	0.879
White blood cell count (MD[IQR])	7.21 [5.22,11.39]	7.48 [4.87,9.65]	0.315
Albumin (MD[IQR])	38.86 [34.46,38.02]	33.06 [32.33,37.40]	0.008*
Triglycerides (MD[IQR])	1.25 [0.93,1.69]	1.37 [1.00,1.94]	0.600
Blood glucose (MD[IQR])	6.72 [5.12,7.73]	6.03 [5.40,8.22]	0.094*
D-Dimer (MD[IQR])	0.87 [0.55,1.85]	1.23 [0.62,1.66]	0.079*
Fibrinogen (MD[IQR])	3.52 [2.49,4.21]	3.92 [2.04,4.55]	0.083*
MBI (MD[IQR])	40 [20,60]	30 [10,40]	< 0.001*
Water Swallow Test(n, %)			< 0.001*
2 level	49 (20.76)	0
3 level	57 (24.15)	4 (2.83)
4 level	59 (25.0)	17 (12.06)
5 level	71 (30.08)	120 (85.11)
PAS (n, %)			< 0.001*
1 ~ 2	85 (36.01)	3 (2.12)
3 ~ 5	58 (24.58)	80 (56.74)
6 ~ 8	93 (39.41)	58 (41.13)
MBSImP (n, %)			< 0.001*
1	66 (27.97)	0
2	61 (25.85)	2 (1.41)
3	45 (19.07)	37 (26.24)
4	33 (13.98)	42 (29.79)
> 4	31 (13.13)	60 (42.55)
rTMS (n, %)	221 (93.64)	92 (65.24)	0.063*
sEMG (n, %)	44 (18.64)	13 (9.21)	0.081*

Abbreviations: MD, indicates median; IQR, indicates interquartile range; * indicates p < 0.1 and statistically significant results

Fig. 3.

Feature variable screening process and results. (A)-(B): Screening process and results of feature variables for Lasso-Logistic regression model; (C)-(D): Screening process and results of feature variables for elastic net-logistic regression model; The bar graph’s length indicates the importance of the variable. The bigger the coefficient of the characteristic variable, the longer the bar graph, showing the greater significance within the model. Abbreviations: SOS, Site of Stroke; AA, Artificial airway; T2DM, Type 2 Diabetes Mellitu; ALB, albumin; BG, Blood Glucose; Fib, fibrinogen; PAS, Penetration-Aspiration Scale; MBSImP, Modified bariums wallow Impairment Profile; MBI, modified Bathel Index; WST, Water Swallowing Test; rTMS, repetitive transcranial magnetic stimulation; sEMGBF, surface electromyographic biofeedback; BMI, Body Mass Index

Construction and validation of predictive models

Based on the eight feature variables screened above, prediction models were constructed using five machine learning algorithms. To ensure optimal model performance and reliability, a systematic modeling and optimization strategy was implemented. Hyperparameter tuning for all models was performed via grid search within predefined parameter spaces, with evaluation conducted through five repetitions of 5-fold cross-validation. The mean AUC from the cross-validation served as the criterion for selecting the optimal parameter set. Given the class distribution in the dataset, particular attention was paid to model performance on the minority class during training and evaluation. In the context of this study, the minority class corresponded to patients who did not achieve swallowing recovery. Robust metrics suited for imbalanced data, such as the F1-score and AUC, were prioritized to ensure clinical utility. Finally, to assess model stability, all performance metrics were calculated as means with 95% confidence intervals based on these five repeated runs.

In the internal validation of the prediction models, the 5-fold cross-validation was used to compare performance. Figure 4A reports the ROC curves and AUC values for the internal validation of the five predictive models. Among them, the RF model exhibited the best predictive performance, with an AUC of 0.949, followed by the GBDT with an AUC of 0.941. The Neural Network model showed a slightly lower AUC of 0.912, while the L₂-norm Logistic Regression and SVM models had AUC values of 0.757 and 0.759, respectively. This study comprehensively evaluated the performance of different prediction models from four perspectives: accuracy, recall, F1-score, and AUC. As shown in Table 2, the Random Forest model achieved the best results across all evaluation metrics, with the GBDT closely following as the second-best performer.

Fig. 4.

Performance in internal and temporal validation. (A): ROC curves and AUC values for the five prediction models in internal validation; (B): ROC curves and AUC values for the five prediction models in temporal validation

Table 2.

Comprehensive evaluation of prediction models for internal and Temporal validation

Models	Precision (95% CI)	Recall (95% CI)	F1 (95% CI)	AUC (95% CI)
Comprehensive Evaluation of Prediction Models for Internal Validation
L2NLR	0.745 (0.720–0.770)	0.567 (0.540–0.594)	0.646 (0.622–0.670)	0.757 (0.734–0.780)
SVM	0.717 (0.692–0.742)	0.572 (0.545–0.599)	0.638 (0.614–0.662)	0.759 (0.736–0.782)
RF	0.958 (0.948–0.968)	0.812 (0.791–0.833)	0.879 (0.865–0.893)	0.949 (0.938–0.960)
NN	0.871 (0.851–0.891)	0.727 (0.704–0.750)	0.797 (0.779–0.815)	0.912 (0.898–0.926)
GBDT	0.951 (0.940–0.962)	0.801 (0.780–0.822)	0.874 (0.860–0.888)	0.941 (0.929–0.953)
Comprehensive Evaluation of Prediction Models for Temporal Validation
L2NLR	0.720 (0.682–0.758)	0.552 (0.510–0.594)	0.628 (0.590–0.666)	0.735 (0.698–0.772)
SVM	0.707 (0.669–0.745)	0.541 (0.499–0.583)	0.619 (0.581–0.657)	0.738 (0.701–0.775)
RF	0.864 (0.832–0.896)	0.788 (0.752–0.824)	0.837 (0.805–0.869)	0.890 (0.860–0.920)
NN	0.828 (0.794–0.862)	0.557 (0.515–0.599)	0.665 (0.627–0.703)	0.866 (0.834–0.898)
GBDT	0.937 (0.915–0.959)	0.795 (0.759–0.831)	0.863 (0.831–0.895)	0.940 (0.916–0.964)

Abbreviations: L₂NLR, L₂-norm Logistic regression; SVM, Support Vector Machine; RF, Random Forest; NN, Neural Networks; GBDT, Gradient Boosting Decision Tree

In the temporal validation of predictive models, this study comparatively analyses the ability of different predictive models based on the temporal validation set. Figure 4B reports the ROC curves and AUC values for the temporal validation of five predictive models, with the AUC value of 0.940 for GBDT being the only one greater than 0.9, indicating that gradient boosting decision tree have a strong generalization ability for the prognostic assessment of dysphagia after stroke.

Table 2 presents the comprehensive evaluation results of the five prediction models on both the internal validation set and the temporal validation set. All performance metrics are reported as the mean and its corresponding 95% confidence interval based on five repeated runs. From the internal validation results, ensemble models, including Random Forest and Gradient Boosting Decision Tree, achieved the best performance, with AUC values exceeding 0.94. However, model performance diverged on the independent temporal validation set. The Gradient Boosting Decision Tree model demonstrated exceptional generalization capability, maintaining the highest AUC of 0.940 on temporal validation, which represented only a 0.1% decrease compared to its internal validation result of 0.941. Its 95% confidence interval lower bound of 0.916 was the highest among all models, indicating the most robust performance estimation. In contrast, the generalization performance of other models declined to varying degrees. The AUC of the Random Forest model decreased from 0.949 to 0.890, a drop of 5.9%. Moreover, its precision on the temporal validation set declined significantly from 0.958 to 0.864, accompanied by a wider confidence interval, suggesting a potential degree of overfitting on the training set. The Neural Network, Support Vector Machine, and L₂-norm Logistic Regression models also exhibited noticeable performance degradation. Integrating the results from both internal and temporal validation, the Gradient Boosting Decision Tree model exhibited the best stability and generalization ability while maintaining high discriminative power. It performed optimally across all evaluation metrics, indicating that the predictor it constructed is the most robust and holds the greatest potential for assisting clinical decision-making.

As can be seen from the above analysis, gradient boosting decision tree have achieved excellent performance results both in internal validation and temporal validation, providing empirical support for its use in practical clinical settings to assist medical work.

Discussion

Baseline medical records of 455 stroke rehabilitation patients were collected, and age, stroke site, artificial airway, dysarthria, albumin, MBSImP, MBI, and rTMS were identified as predictors of prognosis in PSD patients by variable screening. Based on the above influencing factors, our study modelled the prognosis of dysphagia in different stroke patients at hospital discharge based on five machine learning algorithms. In the temporal validation of this study, the gradient boosting decision tree model demonstrated superior predictive performance and generalization ability compared to other models, indicating promising clinical utility.

It is noteworthy that RF achieved the highest AUC in internal validation but showed a more pronounced decline on the temporal validation set compared to the GBDT. This difference may be explained by their distinct algorithmic designs regarding bias-variance tradeoff. RF reduces variance via bagging and often excels on training data, yet its higher model complexity may increase overfitting risk on temporally shifted samples. In contrast, the GBDT reduces bias through iterative error correction, often leading to stronger generalization, especially when modeling complex variable interactions relevant to multifactorial outcomes like swallowing recovery. The greater stability of the GBDT in temporal validation highlights its potential reliability in clinical use.

A meta-analysis showed that history of stroke patients, age and stroke severity were risk factors significantly associated with PSD [21]. In elderly patients, physical deterioration, abnormal swallowing reflexes and cognitive impairment affect the recovery of swallowing function [22]. The present study found that the inclusion of age as a numerical variable did negatively affect the prognosis of swallowing function by performing a univariate analysis of the training set, in line with the findings of other scholars [23]. Some studies have identified brainstem injury, severe stroke and bilateral hemispheric injury as significant predictors of poor prognosis in PSD [24], in contrast to the present study, it has been reported that the site of stroke does not have an effect on the occurrence and severity of dysphagia [25]. Therefore the prognostic impact of stroke site on dysphagia needs to be investigated with a more scientific classification.

The prevalence of aphasia, dysarthria, respiratory infections and pneumonia in patients with PSD is 2–4 times higher than in those without PSD [21]. The degree of dysphagia has been found to be related to aphasia, dysarthria and level of cognitive functioning [26]. While, dysarthria increases the risk of aspiration, which can further exacerbate the risk of aspiration in patients with PSD [27]. Studies have reported that patients with a history of tracheal intubation have a significantly increased risk of persistent dysphagia at discharge compared to patients without a history of intubation [28], which may be related to impaired tongue-laryngeal complex motility [29]. Some scholars have assessed the amount of residue in the pharyngeal recess using VFSS and found that the severity of swallowing function is closely related to the duration of intubation [30]. In the assessment of dysphagia, VFSS and FEES are the gold standard for the diagnosis of dysphagia, but the PAS only assesses leakage and misaspiration during the patient’s examination, and is unable to evaluate other dimensions of the swallowing process. Therefore, the MBSImP, introduced in this study, is a swallowing imaging assessment developed by a team led by Professor Martin-Harris in the United States, who organised experts from several disciplines after repeated studies [31], which can quantify and stratify swallowing function, and the results of the variable screening in the present study corroborate this fact.

Studies have shown that malnutrition on admission adversely affects the recovery of swallowing function at discharge [32, 33], and the risk of pneumonia in patients with dysphagia is 2.4 times higher than in those with non-dysphagia [33]. Therefore, serum albumin levels may independently predict the risk of pneumonia in stroke patients [34, 35].

BMI was found to be a predictor of severe dysphagia after stroke [36]. Scholars found the correlation between the Functional Independence Measure (FIM) scale and the prognosis of swallowing function, confirming that the ability to perform activities of daily living affects the prognosis of swallowing function in patients with PSD [17], which is are consistent with the results of present study. rTMS is a therapeutic technique that uses magnetic pulses to modulate cerebral activity. This stimulation has been shown to enhance excitability, promote neuromediator release [37, 38] and facilitate neural remodelling [39]. A number of studies have confirmed that rTMS can promote swallowing function in PSD patients, and it has gradually been widely used in the treatment of dysphagia [40, 41]. rTMS was shown to be a very important variable in the variable screening step, so the inclusion of rTMS interventions in a comprehensive rehabilitation programme had a positive impact on promoting recovery in patients with PSD.

Limitations and future directions

This study has several limitations. First, its single-center retrospective design and the modest size of the temporal validation set may affect the statistical power and generalizability of the findings. Second, while temporal validation was performed, external validation across diverse populations and settings is required to confirm clinical transportability. Third, the calibration of the model’s predicted probabilities, a key aspect for individualized risk assessment, was not evaluated and should be quantified in future studies. Fourth, some predictive variables, for example stroke lesion location, could benefit from more granular classification in larger cohorts. Finally, prospective interventional studies are needed to test whether risk stratification based on this model can guide effective, personalized rehabilitation.

Conclusion

This study identified eight most important predictors of swallowing function recovery by incorporating 27 characteristic variables, including unfavorable factors such as age, stroke location, dysarthria, artificial airway, and MBSImP, and favorable factors such as albumin, MBI, and treatments involving rTMS. Five predictive models for the prognosis of dysphagia patients were constructed using machine learning. Through internal and temporal validation, the GBDT demonstrated the most robust performance with the smallest decline in AUC on the temporal validation set. These findings suggest that GBDT is a highly promising model for predicting post-stroke swallowing recovery. Future multi-center external validation is required prior to its clinical translation.

Author contributions

HT: Writing–original draft, Conceptualization, Investigation, Methodology. CL: Writing–review &editing, Conceptualization, Methdology, Software, Formal analysis. YF: Writing–review &editing, Data curation, Supervision, Investigation. YY: Writing–review &editing, Methodology, Project administration, Investigation. CX, SC and HC: Data curation, Investigation, Methodology. HD and MJ: Supervision, Project administration, Writing–review & editing. MS: Conceptualization, Funding acquisition, Writing–review & editing.

Funding

This work was supported by National key Research and Development Program of China (No.2022YFC2009700 & No.2022YFC2009706), National Natural Science Foundation of China (No.82272594) and Horizontal project of Soochow University (No.H201173). The funding bodies played no role in the design of the study and collection, analysis, and interpretation of data and in writing the manuscript.

This work was supported by Qilu Medical Research Fund, Soochow Medical College, Soochow University (24QL200219), National key Research and Development Program of China (No.2022YFC2009700 & No.2022YFC2009706), National Natural Science Foundation of China (No.82272594) and Horizontal project of Soochow University (No.H201173). The funding bodies played no role in the design of the study and collection, analysis, and interpretation of data and in writing the manuscript.

Data availability

The datasets used and/or analysed during the current study are available from the corresponding author on reasonable request.

Declarations

Ethical approval

This study conformed to the ethical guidelines of the Declaration of Helsinki and. The study was approved by the Ethics Committee of The Fourth Affiliated Hospital of Soochow University (Dushu Lake Hospital Affiliated to Soochow University) (241126). Informed consent was waived by the Ethics Committee of the Fourth Affiliated Hospital of Soochow University due to the retrospective nature of the study.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.