Abstract
Objective
To systematically review the current state of research on the use of intelligent technologies for dysphagia risk prediction. This scoping review summarizes existing studies in terms of their technical approaches, data sources, and model performance, and analyzes key technological limitations and challenges in clinical translation. The goal is to provide insights for future study design and clinical implementation.
Methods
This study was developed based on the Preferred Reporting Items for Systematic Reviews and Meta-Analyses Extension for Scoping Reviews (PRISMA-ScR) checklist to comply with reporting and methodological standards. We systematically searched eight databases and clinical trial registries from their inception until February 25, 2026. Studies that underwent peer review as well as relevant grey literature were included. Subsequently, the results were comprehensively analyzed and discussed.
Results
A total of 38 studies were ultimately included, most of which were single-center retrospective cohort studies with sample sizes ranging from 50 to 59,811 participants. The number of publications has grown rapidly since 2023. The study populations mainly included patients with stroke, head and neck cancer patients undergoing radiotherapy, and older adults. Structured clinical data were the predominant data source, while only a few studies incorporated multimodal inputs such as imaging, physiological, or acoustic signals. Traditional machine learning models were most commonly used, followed by traditional statistical models, hybrid models, and deep learning methods. Overall model performance was generally good (area under the curve values mostly exceeding 0.70), yet there was considerable heterogeneity in model types, data sources, and outcome reporting.
Conclusion
Intelligent technologies show promising potential in dysphagia risk prediction. However, existing evidence remains heterogeneous and primarily exploratory. Future research should strengthen the construction of multicenter datasets, multimodal data integration, and external validation to enhance model generalizability and clinical applicability.
Keywords
1. Introduction
Dysphagia is defined as the inability to safely and effectively transport food from the oral cavity to the stomach to obtain adequate nutrition and hydration, resulting from structural and/or functional impairment of organs such as the mandible, labia, tongue, soft palate, pharynx, and esophagus—and consequently leading to feeding difficulties.1,2 Individuals with dysphagia demonstrate high clinical heterogeneity, with diverse and complex etiologies often secondary to neurological disorders, head and neck oncology, or respiratory diseases. 3 They are not only prone to directly developing severe physiological complications (e.g., aspiration pneumonia, malnutrition, dehydration) but also experience diminished eating pleasure, restricted social participation, and exacerbated negative psychological states,4,5 all of which profoundly affect quality of life and health outcomes. Studies have shown that patients with dysphagia carry a high risk of silent aspiration—up to 47.2% 6 —yet early symptoms are often subtle and underrecognized, resulting in a high rate of missed diagnosis and delayed intervention. With the accelerating aging of the population, the prevalence of dysphagia in older adults is rising markedly. Meta-analytic evidence indicates that the prevalence reaches approximately 41.0% in adults aged over 80 years, nearly doubling that in the 60–69-year age group (21.0%). 7 Aging-related physiological decline—including weakening of oropharyngeal muscle strength, diminished oral mucosal sensation, and reduced esophageal peristaltic function—renders older adults more susceptible to dysphagia and further increases the risk of complications and long-term care needs. 8 Therefore, achieving early risk identification and precise assessment of dysphagia has become an important issue in clinical practice. However, existing dysphagia screening and assessment tools remain diverse, with no standardized methodology. Their utility is often constrained by specific operational contexts and the professional expertise of healthcare providers. Moreover, the reliability, accuracy, and clinical feasibility of these tools require further validation 9 ; additionally, traditional tools generally suffer from insufficient sensitivity and weak dynamic tracking capacity, undermining the precision of assessment and intervention. 10 For instance, the water swallowing test is highly susceptible to patient compliance, 11 while the Eating Assessment Tool-10 (EAT-10) shows variability in validity across cultural contexts. 12 Moreover, current approaches are largely focused on the screening and diagnosis of established dysphagia, while early prediction of dysphagia risk remains relatively underdeveloped.13,14 In recent years, intelligent technologies have driven a shift in dysphagia risk prediction—from traditional experience-based approaches to data-driven precision modeling—through methods such as automated data collection, feature extraction, and dynamic modeling. Common modeling approaches for disease risk prediction include traditional statistical models, machine learning, and deep learning, of which the latter two are core components of artificial intelligence (AI). These methods demonstrate strong potential for handling complex data patterns and uncovering nonlinear relationships. 15 Nevertheless, most current studies on dysphagia risk prediction are limited to specific patient populations, with incomplete predictor variables, narrow application scenarios, suboptimal model generalizability, and inconsistent data quality.16–18 Furthermore, the technological characteristics, methodological diversity, and predictive performance of artificial intelligence models in this field have not yet been systematically reviewed. Therefore, this study adopts a scoping review methodology to comprehensively map existing research on dysphagia risk prediction models, summarize key information on modeling techniques, data sources, and model performance, and analyze technical features and methodological differences across studies, with the aim of informing future research and clinical applications.
2. Materials and methods
This scoping review was conducted following the five-stage methodological framework established by Arksey and O’Malley 19 : 1) identifying the research question, 2) determining relevant research, 3) selecting research, 4) charting the extracted data, and 5) collating, summarizing, and reporting the results. This study was developed based on the Preferred Reporting Items for Systematic Reviews and Meta-Analyses Extension for Scoping Reviews (PRISMA-ScR) checklist 20 to comply with reporting and methodological standards. It was prospectively registered on July 6, 2025 (Registration DOI: 10.17605/OSF.IO/2HAFP) in the Open Science Framework (OSF).
2.1. Identifying the research question
This study employed the Joanna Briggs Institute (JBI) Population–Concept–Context (PCC) framework, as outlined in the JBI Manual for Evidence Synthesis, 21 to guide the scoping review, focusing on the clinical application of intelligent technologies for predicting dysphagia in high-risk populations. The primary research questions were as follows: 1) What is the current status and effectiveness of intelligent technologies in predicting dysphagia risk among high-risk populations? 2) What types of predictive modeling techniques and methodological characteristics have been employed in existing research on dysphagia risk prediction? 3) What methodological limitations and challenges in clinical translation exist in current studies, and what future research directions warrant further attention?
2.2. Determining relevant research
A systematic search was conducted across Chinese databases—CNKI (China National Knowledge Infrastructure), CMJD (Chinese Medical Journal Full-Text Database), WanFang Data, and CBM (China Biomedical Literature Database)—and English databases (PubMed, EMBASE, CINAHL, and Web of Science), as well as clinical trial registries such as ClinicalTrials and the Chinese Clinical Trial Registry to identify potentially eligible unpublished or ongoing studies. The search employed a hybrid strategy of subject headings and free-text keywords. English search terms included “dysphagic,” “deglutition disorder,” “artificial intelligence,” “machine learning,” “neural networks,” “computers,” “complications,” “Detection,” “predictions,” and “influence factor.” The temporal scope covered records from the inception of each database to February 25, 2026. Detailed search strategies for all databases are provided in Supplementary File 1.
2.3. Selecting research
After importing the retrieved studies into EndNote X9 for deduplication, two trained researchers screened the studies strictly according to the pre-defined inclusion and exclusion criteria. Initial screening involved evaluating titles and abstracts, followed by a secondary screening of full texts for eligible studies. Any disagreements between the two researchers were resolved through discussion with a third independent reviewer. Ultimately, the final set of included studies was confirmed. Inclusion Criteria: 1) Study populations comprised adult high-risk groups for dysphagia (e.g., older adults, post-stroke patients, intensive care unit (ICU) patients, and postoperative head and neck tumor patients); 2) Studies employed data-driven predictive modeling approaches, including traditional statistical methods, machine learning, deep learning, or hybrid models. Machine learning and deep learning techniques—core components of artificial intelligence—were the main focus of this review. Traditional statistical models were also included to provide a comprehensive overview of the methodological landscape in this field; 3) The primary objective focused on risk prediction for dysphagia or related complications (e.g., predicting dysphagia occurrence risk, future aspiration risk, or complication probability), rather than solely diagnosis, assessment of existing symptoms, or evaluation of treatment intervention effects; 4) Original research literature, including peer-reviewed journal articles and relevant grey literature (e.g., theses, conference papers, technical reports, or preprints). Exclusion Criteria: 1)Full texts unavailable; 2) Non-Chinese or non-English languages; 3) Literature types lacking original research data, such as guidelines, reviews, commentaries, and editorials; 4) Studies involving pediatric populations.
2.4. Charting the extracted data
Two researchers independently extracted data from the included studies. In case of disagreements, a third researcher was consulted for resolution. Extracted information included author(s), publication year, country, study population, sample size, technology type, data input type, outcome measures, and other relevant details. All data were entered into an Excel spreadsheet for aggregation and analysis.
2.5. Data synthesis and analysis
Core characteristics of intelligent technologies for dysphagia risk prediction were identified and descriptively analyzed through comprehensive synthesis of included studies. Extracted information was presented in both tabular and narrative formats to align with the review objectives and research questions.
3. Results
3.1. Literature search results
The initial search retrieved 5490 relevant articles. After importing into EndNote X9 for deduplication, 615 duplicates were removed. Initial screening of titles and abstracts excluded 4792 studies that did not meet the pre-defined inclusion and exclusion criteria, leaving 83 articles for further review. Following full-text assessment, 45 studies were excluded, resulting in a final cohort of 38 included studies.22–59 Of these, 30 were English-language studies22–33,36–39,41–46,51–56,58,59 and 8 were Chinese-language studies.34,35,40,47–50,57 The detailed literature screening workflow is illustrated in Figure 1. Literature screening process.
3.2. basic characteristics of included studies
A total of 38 studies published between 2009 and 2026 were included, with 28 studies32–59 published after 2023, indicating growing research interest in intelligent technologies for dysphagia risk prediction in recent years (Figure 2). In terms of model category evolution, traditional statistical models were mainly concentrated in the period before 2019 (5/11, 45.5%).22–27 Traditional machine learning models began to emerge from 2020 onwards, with 14 studies published between 2024 and 202636,39,40,43,45,47–50,55–59, accounting for 58.3% of studies in this period. Deep learning studies appeared after 202437,46. A total of seven hybrid models were identified, of which five were published in the last three years.38,44,51,52,54 Geographically, studies were predominantly from China (52.6%, 20/38),31,34,35,40–43,47–59 the Netherlands (15.8%, 6/38),22–25,29,46 Korea (10.5%, 4/38),33,37,38,45 and Austria (7.9%, 3/38),28,32,36 reflecting a concentration in Asia and Europe. Most studies employed retrospective cohort studies (24 studies),25,28,30,33,35,38,40–54,56,57,59 followed by prospective cohort studies (12 studies),22–24,27,29,31,32,34,36,37,55,58 with one cross-sectional study
26
and one prospective diagnostic accuracy study.
39
Sample sizes varied widely (50 to 59,811 participants), with most studies having fewer than 500 participants,23–27,29,31,34–39,41,42,45,48–50,52,53,56,57 though some leveraged large databases for model development.28,30,32,33,44,46,51,54,55 Stroke-related patients were the most frequently studied population (14 studies),28,30,32,33,44,46,51,54,55 followed by head and neck cancer radiotherapy patients (7 studies).22,24,25,27,29,30,46 Additional studies focused on special populations, including inpatients,22,24,25,27,29,30,46 older adults,34,54,55,57 patients with chronic obstructive pulmonary disease,
48
traumatic brain injury patients,
50
ICU patients with endotracheal intubation,
44
and surgical patients (e.g., those undergoing cardiac surgery or anterior cervical discectomy and fusion).53,58 Detailed characteristics are presented in Table 1. Temporal distribution of dysphagia risk prediction model categories from 2009 to 2026. Basic Characteristics of the included literature.
3.3. Study design and data source characteristics
Included studies exhibited variations in research objectives, research center configurations, and data sources, with detailed characteristics presented in Table 1. Among them, 26 studies focused on predicting the risk of dysphagia occurrence22–26,28,30–32,35,36,38,41,43,44,46,48–55,58,59; 7 studies targeted aspiration or silent aspiration risk27,29,33,37,39,42,47; 2 addressed pneumonia risk,34,40 1 predicted swallowing function recovery 56 ; 1 focused on malnutrition risk 56 ; and 1 examined long-term prognosis or related clinical outcomes of dysphagia. 45 Regarding research center settings, most were single-center studies (28 studies, 73.7%),22,25,26,29,31–39,41–43,45,47–50,52,53,56–59 followed by multi-center studies (6 studies, 15.8%)23,28,30,44,54,55 and dual-center studies (4 studies, 10.5%).24,27,46,51 In terms of data sources, the majority (30 studies, 78.9%) utilized structured clinical data for modeling,22–25,27–30,32–36,40–45,47–51,53,54,56–59 primarily including demographic characteristics, disease-related indicators, laboratory tests, and treatment information. Four studies (10.5%) employed multimodal data (e.g., clinical data combined with medical imaging or acoustic information) for model construction.31,46,52,55 Additionally, 3 studies (7.9%)26,37,39 used physiological signal data for prediction, and 1 study (2.6%) 38 relied primarily on medical imaging data. Overall, current dysphagia prediction research is predominantly driven by single-center studies using structured clinical data, with multimodal data and multi-center studies remaining relatively scarce.
3.4. Predictive modeling technology types
Predictive modeling technology types and model performance.
3.5. Association between study variable types and modeling techniques
The predictive variables in the included studies covered multiple dimensions, including demographic characteristics, clinical assessment indicators, comorbidities and laboratory data, imaging-derived features, and swallowing-related physiological signals. These variable types not only determined the model inputs but also largely influenced the choice of modeling techniques. All conventional statistical models relied on structured clinical data (11/11, 100%),22–25,27,29,30,35,41,42,53 and were well suited to linear association analysis and risk stratification. By contrast, machine learning and deep learning are more appropriate for high-dimensional, unstructured data such as imaging, physiological signals, and multimodal inputs, as they excel in automatic feature extraction and complex pattern recognition. When handling structured data, conventional statistical models assume linear relationships and require manual feature selection. 23 Their coefficients can be directly converted into clinical risk scores. 42 Machine learning models, while free from functional-form assumptions and capable of automatically capturing high-dimensional nonlinear interactions, 33 usually require post hoc interpretability methods such as SHapley Additive exPlanations (SHAP) and are associated with a relatively higher risk of overfitting. 43 In comparison, studies based on imaging or physiological signals were limited in number, but they still highlight the potential of machine learning and deep learning in processing unstructured data. Multimodal studies suggest that cross-modal fusion may improve predictive performance, although challenges remain in data alignment, feature integration, and model interpretability. Machine learning based on structured clinical data remains the most commonly used modeling strategy, whereas deep learning based on imaging and physiological signals, as well as multimodal fusion approaches, represent emerging directions. Therefore, prediction models should not be evaluated solely on the basis of algorithmic superiority; instead, their methodological suitability and clinical value should be interpreted in light of the input data type.
3.6. Model predictive performance
Model predictive performance varied significantly across different technology types among the included studies, with details presented in Table 2. Traditional statistical models demonstrated generally good area under the curve (AUC) performance, but most studies only reported AUC or risk stratification results, with limited reporting of accuracy, sensitivity, and specificity.23–25,27,29,30 Traditional machine learning models showed substantial variation in predictive performance, heavily influenced by feature selection and sample size, with some studies omitting accuracy, sensitivity, or specificity metrics.26,31,43,49,51,55,57,58 Notably, some models demonstrated excellent performance on training sets but exhibited substantial performance degradation on validation sets, indicating a risk of overfitting. For instance, in the study by Ye F et al., 43 the LGBM model achieved an AUC of 0.841 on the training set, but only 0.50 on the validation set. Among the two deep learning models, MobileNetV3 predicted aspiration using post-meal acoustic signals, demonstrating potential for non-invasive bedside screening with comprehensive performance metrics reported. 37 The 3D ResNet model, integrating imaging and dosimetric data, outperformed traditional models in multimodal prediction but primarily reported AUC, with other metrics relatively limited. 46 In contrast, hybrid models demonstrated superior predictive performance and more comprehensive metric reporting by integrating multi-source information.34,38,44,52 Additionally, a few studies provided no quantifiable outcome metrics, describing prediction factors or model capabilities only in text.22,39,54 Overall, all model types exhibited reasonable capability for dysphagia risk prediction; however, incomplete reporting of performance metrics limits direct comparisons across techniques. Future studies should standardize model evaluation metric reporting to enhance research comparability, transparency, and clinical applicability.
4. Discussion
This scoping review systematically synthesized the application progress of intelligent technologies in dysphagia risk prediction and delineated the primary methodological characteristics and development trends in this emerging field. Overall, related research has shown a marked growth trend in recent years, particularly after 2023, indicating the rapid development of intelligent technologies in digital health and clinical risk prediction. Comprehensive analysis reveals that most studies employed traditional machine learning methods and primarily relied on structured clinical data for model construction. Although some models reported high predictive performance, the distribution of modeling techniques was uneven, single-center studies predominated, multimodal data applications were limited, and certain models lacked external validation or comprehensive performance metric reporting—indicating substantial heterogeneity across studies in this field. These findings demonstrate that intelligent technologies hold potential value for predicting dysphagia and related complications, but the research remains largely exploratory, with clinical translatability requiring further validation.
Traditional machine learning methods were the most commonly used among included studies, followed by traditional statistical models, hybrid approaches, and deep learning. Although logistic regression remains widely utilized due to its interpretability and clinical familiarity, machine learning algorithms such as RF, XGBoost, and SVM demonstrated superior predictive performance in some studies, with certain models achieving AUC values exceeding 0.95.36,43,45,49–51 In medical prediction research characterized by limited sample sizes and primarily structured clinical data, machine learning algorithms are typically more readily implemented, 60 thus maintaining their dominant position in current dysphagia risk prediction studies. However, with the accumulation of large-scale clinical data and advances in computational power, deep learning techniques may exhibit greater advantages in processing complex, high-dimensional data (e.g., medical imaging, physiological signals).37,46 Nevertheless, the “black box” nature and high computational complexity of deep learning models have limited their clinical adoption to some extent. Relevant studies indicate that lack of model interpretability is a major concern for clinicians regarding AI systems. 61 In recent years, explainable machine learning methods have gained increasing attention, as they can elucidate the contribution of individual variables during model prediction, thereby enhancing model transparency and comprehensibility. This is particularly crucial for clinical applications, as explainable models not only help healthcare providers understand model decision logic but also facilitate evaluation of algorithm fairness, equitable access, and potential biases across patient populations. 62 For instance, SHAP values—a commonly used model interpretation method—can quantify each feature’s contribution to predictions, aiding identification of key predictive factors and building clinical trust. 63 Beyond model interpretability, modeling strategies and data integration approaches have emerged as important research directions in current studies. Regarding modeling strategies, hybrid models—integrating the strengths of different algorithms to enhance model stability—have gained increasing attention from researchers. 64 Among the four hybrid model studies included in this review, the models demonstrated relatively stable predictive performance (AUC 0.82–0.99), suggesting that multi-algorithm integration strategies may improve model robustness. In terms of data integration approaches, several studies explored multimodal data fusion strategies, integrating multi-source data such as clinical information, imaging, and biosignals to more comprehensively characterize swallowing function status. 65 In the future, the combination of hybrid modeling leveraging multi-algorithm advantages with multimodal data fusion is likely to emerge as a key developmental direction for dysphagia risk prediction models.
Another significant finding of this review is the heavy reliance of existing prediction models on structured clinical data. Most studies utilized demographic characteristics, clinical indicators, laboratory results, or disease-related information for model construction. Such data, typically sourced from electronic health record systems, offers advantages including accessibility and high standardization, establishing it as one of the most common data sources in current medical AI research. 66 However, dysphagia is a multifactorial disease closely associated with multiple mechanisms, including neural regulation, muscle movement, structural changes, and aging. 67 Structured clinical data from single sources often only capture patients’ static clinical characteristics, making it challenging to comprehensively elucidate the complex physiological mechanisms underlying dysphagia. Recent studies have increasingly demonstrated the substantial value of multimodal data integration in medical artificial intelligence research. By fusing imaging information, physiological signals, swallowing kinematic metrics, and speech or acoustic features, these approaches enable more comprehensive and dynamic characterization of swallowing function status. 68 Among the included studies, although only a minority utilized multimodal data for modeling,31,46,52,55 these models generally exhibited superior predictive performance. This suggests that multimodal data fusion not only enhances prediction accuracy but also facilitates multidimensional depiction of dysphagia’s complex pathophysiological mechanisms, thereby improving the clinical interpretability of predictive models. Furthermore, advances in wearable technologies enable continuous acquisition of multimodal physiological signals, opening novel technical pathways for developing dysphagia risk prediction models with dynamic monitoring capabilities. 69
Regarding model performance, most studies reported favorable predictive results, typically using AUC as the primary evaluation metric. Multiple studies demonstrated moderate to high predictive performance, indicating the potential clinical value of intelligent technologies for dysphagia risk prediction. However, significant variations existed between studies in the selection and reporting of model evaluation metrics. Some studies reported multiple metrics including accuracy, sensitivity, specificity, and F1 score,26,28,32–37,40–45,47–53,55–59 while others reported only single metrics.23–25,27,29–31,38,46 Additionally, differences in variable selection, data collection frequency, and labeling standards across studies increased the difficulty of comparison and integration.26,38,43,52 Some studies lacked independent datasets for external validation, relying solely on internal data for model evaluation,23,42,46,55 potentially leading to overestimated model performance. These methodological differences compromised comparability across studies. In recent years, the medical AI research community has developed reporting standards such as Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis–Artificial Intelligence (TRIPOD-AI) 70 and Consolidated Standards of Reporting Trials–Artificial Intelligence (CONSORT-AI) 71 to improve prediction model reporting quality and methodological transparency. Future dysphagia risk prediction research should adhere to these reporting guidelines to enhance reproducibility and clinical translatability.
Notably, this review identified significant geographical imbalance in research distribution, with over half of studies originating from China. This phenomenon may relate to several factors. First, China’s large population base is associated with a heavy burden of stroke and aging-related diseases, where dysphagia exhibits high prevalence among stroke patients and older adults, providing rich clinical data sources for relevant research. Second, China has substantially increased investments in medical AI and digital health in recent years, with multiple healthcare institutions establishing clinical big data platforms that provide an important data foundation for machine learning research. 72 Additionally, certain countries face stricter data privacy and regulatory constraints on medical data usage, potentially increasing the complexity of large-scale clinical data sharing and research. 73 Meanwhile, dysphagia research in some countries has focused more on diagnostic techniques, imaging assessment, or rehabilitation interventions, with relatively fewer studies targeting intelligent technology-based risk prediction models.74,75 Consequently, current evidence may exhibit geographical bias. Most prediction models remain developed using data from single countries or healthcare systems, with their generalizability across diverse populations and healthcare settings requiring further validation. Future research should strengthen cross-regional, multicenter collaborations and conduct external validation using multi-regional datasets to further enhance the external applicability and clinical generalizability of prediction models.
From a clinical application perspective, the application of intelligent technologies in dysphagia risk prediction holds substantial importance. Dysphagia is associated with multiple serious complications, including aspiration pneumonia, malnutrition, dehydration, and prolonged hospital stays, particularly prevalent among older adults and patients with neurological disorders. 76 Traditional dysphagia assessment methods typically rely on specialist bedside evaluations or imaging examinations. Video fluoroscopic swallowing study (VFSS) and fiberoptic endoscopic evaluation of swallowing (FEES) are considered gold standards for diagnosing dysphagia; however, their invasive nature and stringent implementation requirements result in limited clinical adoption. 77 In contrast, intelligent prediction models can rapidly identify high-risk individuals using routine clinical data, supporting early screening and risk stratification. Such prediction tools hold particular potential value in resource-limited environments such as nursing homes and community settings, where specialized dysphagia assessment personnel are often unavailable. Furthermore, integrating prediction models into electronic health record systems or digital health platforms enables automated risk alerts and real-time clinical decision support, facilitating early identification and personalized intervention for dysphagia. 32 Therefore, future research should enhance model performance while simultaneously prioritizing model simplicity, interpretability, and clinical accessibility to facilitate the translation of intelligent prediction technologies from research into clinical practice.
4.1. implications for future research
Future research should advance in the following key areas: First, strengthen multicenter, large-sample study designs and establish cross-center datasets to enhance model generalizability and external validation reliability. Second, research should advance multimodal data fusion by integrating clinical information, medical imaging, acoustic signals, and physiological metrics, while leveraging wearable devices to enable continuous data acquisition for more comprehensive and dynamic characterization of swallowing function status. Third, implement standardized data management and model evaluation following TRIPOD-AI and CONSORT-AI guidelines to improve research transparency, reproducibility, and translatability. Fourth, enhance model interpretability and clinical usability to build healthcare provider trust and facilitate practical implementation. Finally, explore integration of intelligent prediction tools into digital health platforms to enable automated risk alerts and real-time clinical decision support, driving clinical practice translation through early screening, dynamic monitoring, and personalized interventions.
4.2. Strengths and limitations
This scoping review systematically synthesized research evidence on intelligent technologies for dysphagia risk prediction, comprehensively mapping technology types, data sources, and model performance to provide a scientific basis for understanding current research characteristics and development trends. The methodology demonstrated rigor, employing a scoping review approach with systematic bilingual database searches across multiple sources, independent data extraction and analysis by two researchers, and discrepancy resolution by a third investigator, ensuring reliability of data synthesis. Included studies spanned diverse countries, disease types, and high-risk populations, reflecting the varied applications of intelligent prediction technologies. Furthermore, our inclusion criteria focused on data-driven risk prediction models and excluded studies solely focused on diagnosis, assessment of existing symptoms, or intervention effects. Although this limited the number of included studies to some extent, it enhanced the specificity and methodological consistency of the research question. By systematically searching multiple databases and including grey literature, we ensured relatively comprehensive literature coverage.
Several limitations exist. First, the number of included studies was limited (n = 38). This limitation reflects the early developmental stage of intelligent technology applications in dysphagia risk prediction, as studies published in 2024–2026 accounted for 63.2% of the total, indicating that this field has only recently entered a phase of rapid growth. Second, most studies employed single-center retrospective designs with limited sample sizes and heterogeneous study populations, restricting model generalizability and cross-study comparability. Third, models heavily relied on structured clinical data, inadequately capturing dysphagia’s multifactorial complexity. Fourth, deep learning models face challenges in interpretability and computational complexity, hindering clinical implementation. Additionally, most studies lacked external validation, with geographical distribution concentrated in few countries, potentially introducing regional bias. Finally, some studies reported only single performance metrics with incomplete model evaluation, reducing comparability across technologies.
5. Summary
Intelligent technologies hold significant clinical value in dysphagia risk prediction, shifting the field from experience-based to data-driven paradigms while demonstrating potential to enhance prediction objectivity, efficiency, and early risk identification. Despite limitations in data sources, model evaluation, and generalizability, multimodal data fusion, deep learning advancements, and multicenter data integration show promise for improving model performance and clinical translatability. Future research should prioritize standardized data management and model assessment, enhanced interpretability, and integration of prediction tools into digital health platforms for automated risk alerts and real-time clinical decision support to facilitate practical applications in dysphagia management.
Supplemental material
Supplemental material - Application of intelligent technologies for dysphagia risk prediction: A scoping review
Supplemental material for Application of intelligent technologies for dysphagia risk prediction: A scoping review by Yuyuan Han, Jiayi Hou, Yijia Luo, Peimeng Teng, and Guijuan He in DIGITAL HEALTH.
Footnotes
Acknowledgement
We would like to acknowledge all of the participants and all of the researchers in this study. Artificial intelligence tools were used to assist with language editing and improving the clarity of the manuscript. The authors reviewed and revised all content and take full responsibility for the final manuscript.
Author contributors
Y. Han proposed the research questions, conducted data collection and analysis, and wrote the manuscript; J. Hou assisted in determining inclusion/exclusion criteria and search strategies, assisted in data analysis, and revised the manuscript; Y. Luo and P. Teng assisted in data collection and organization; G. He provided overall research guidance and secured funding support.
Funding
The authors disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This study was supported by Zhejiang Provincial Basic Public Welfare Research Plan under Grant No. LTGY23H250001. The authors received no financial support for the research, authorship, and/or publication of this article.
Declaration of conflicting interests
The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.
Guarantor
G. He
Supplemental material
Supplemental material for this article is available online.
References
Supplementary Material
Please find the following supplemental material available below.
For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.
For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.
