Aspiration occurs when oropharyngeal contents, such as food, liquid, saliva, or secretion, are accidentally misdirected into the larynx, lower respiratory tract, or lung (Ebihara et al., 2016), which may result in aspiration pneumonia if infection or inflammation develops. Aspiration could be life-threatening when the airway is blocked (i.e., asphyxiation), and aspiration pneumonia was ranked as the third leading cause of injury deaths in older people (Kramarow et al., 2014). A study on 784 patients reported that 65.2% demonstrated pharyngeal residue-related dysphagia (Seo et al., 2021). Aspiration pneumonia resulted in more than 58,000 annual deaths in the United States, with an age-adjusted mortality rate of 21.85 per 100,000 people (Gupte et al., 2022). Another study found that the median hospitalization charge for aspiration pneumonia was US$30,526 (Wu et al., 2017). Patients who suffered from aspiration reported fear, depression, and frustration that they might aspirate again, which affected their mental health and quality of life (Martino et al., 2009).

Dysphagia or deglutition disorder (i.e., difficulty swallowing) is the main cause of aspiration (Morley, 2015). Individuals who suffered from dysphagia were about nine times more likely to develop aspiration pneumonia (van der Maarel-Wierink et al., 2011). Likewise, dysphagia was present in 92% of pneumonia patients (Almirall et al., 2013). While the prevailing dysphagic aspiration has imposed a heavy burden on the healthcare and hospitalization systems (Allen et al., 2020; Lesa et al., 2021), early diagnosis and screening of dysphagia and aspiration risks are essential to facilitate effective management and reduce subsequent risks of pulmonary complications (Hines et al., 2016; Wirth et al., 2016). Nevertheless, aspiration resulting from dysphagia is often referred to as “silent aspiration”. Some patients may remain clinically asymptomatic, without presenting coughing or choking signs, and self-report swallowing difficulties (Wakasugi et al., 2008; Miller et al., 2009; Suiter et al., 2020). Confirmation of apparent aspiration cannot be obtained clinically (Teramoto, 2022). Bedside assessment frequently misses dysphagia patients with silent aspiration (Wakasugi et al., 2008). Meanwhile, the gold standards for assessing dysphagic aspiration are the videofluoroscopic swallowing study (VFSS) and the fiber-optic endoscopic evaluation of swallowing (FEES). Nevertheless, both fluoroscopy and endoscopy induce pain, discomfort, nausea, and radiation exposure, especially in children (Ingleby et al., 2021), which are not feasible to facilitate continuous or frequent assessments. There standards also require high costs and professionals to operate (Lancaster, 2015).

As dysphagia or neurodegenerative problems deteriorate gradually and aspiration could occur unexpectedly (Lim et al., 2023), it is necessary to develop accessible and reliable instrumental screening tools that enable continuous or frequent assessments of aspiration risks. Wearable technology with computer-aided diagnosis/screening might be a potential alternative to bedside questionnaires and instrumental diagnostic instruments (such as VFSS and FEES) (So et al., 2023). Accelerometers are among the most common sensors used in wearable technology, and have been used to evaluate levels of physical (Karas et al., 2022), ambulatory (Steins et al., 2014), and behavioral information (Cheung et al., 2022). In the case of swallowing, accelerometers can trace the biomotion of the laryngeal region that manifests swallowing abnormalities and thus aspiration risks (So et al., 2023). In the same vein, soft sensors with flexible electronics or artificial skin could serve the same purpose (Chen J. et al., 2021; Chen et al., 2021b; Gao et al., 2021). On the other hand, aspirated patients may demonstrate a wet voice (Warms and Richards, 2000) and attenuated breathing and swallowing sounds (Shaw et al., 2004; Kang et al., 2017), in which acoustic features could be recognized by microphones.

Computer-aided screening, using machine learning and deep learning, can enhance the assessment of swallowing functions and, hence, dysphagia or aspiration in older adults. Park et al. (2023) attempted to predict aspiration by applying machine learning models to a bedside screening questionnaire (GUSS test). They attained an area under the receiver operating characteristics curve (AUC) of 0.81. Through the examination of videofluoroscopic hyoid motions, Lee et al. (2016) detected swallowing impairment with strong discriminative power (AUC = 0.93) using the support vector machine (SVM). In addition, Roldan-Vasco et al. (2021) categorized swallowing dysfunctions by speech variations using the random forest and obtained a 91.0% sensitivity. Nevertheless, current wearable technology was deemed insufficiently reliable to recognize swallowing and non-swallowing, which hindered real-world applications (So et al., 2023).

To this end, our review question is: how were wearable technologies with computer-aided screening techniques were utilized to identify aspiration risks in dysphagia, and how accurate were these techniques or systems, in general? The objective of this review is to summarize the evidence on the testing techniques, protocols, and accuracy performances for the assessment of aspiration risks. The Preferred Reporting Items for Systematic Review and Meta-Analyses (PRISMA-DTA) extension for diagnostic test accuracy was adopted to frame the reporting of this review. The review was registered in PROSPERO (reference number: CRD42023408960).

The PRISMA flowchart shown in Figure 1 illustrates the search and screening process for the review. The initial search identified 178 articles from the six databases, and 96 duplicate articles were removed. Screening on the title and abstracts excluded 37 records, for the following reasons: violation of inclusion criteria of article types, such as reviews and conference abstracts, n = 2; duplicate publications (articles published as full conference papers were published again in journals with the same content. In such cases, only publications in journals were retained since they contained full, detailed information), n = 2; irrelevant to dysphagia and aspiration, n = 14; not related to the classification of aspiration, n = 7; invasive instrument, such as manometry, n = 9; non-instrumental, such as questionnaires, n = 3. Subsequently, the full-text screening was further performed to exclude 36 articles, of which 30 were not related to the classification of aspiration, 4 involved invasive instruments, and 2 were targeted for non-instrumental screening. In the end, 9 articles were eligible for data synthesis (Lee et al., 2006; Lee et al., 2011; Merey et al., 2012; Sarraf Shirazi et al., 2012; Sejdic et al., 2013; Sarraf Shirazi et al., 2014; Frakking et al., 2022; Park et al., 2022; Shu et al., 2022).

There were five studies led by research institutes from Canada (Lee et al., 2006; Lee et al., 2011; Merey et al., 2012; Sarraf Shirazi et al., 2012; Sarraf Shirazi et al., 2014), two from the United States (Sejdic et al., 2013; Shu et al., 2022), one from Australia (Frakking et al., 2022) and one from South Korea (Park et al., 2022). In addition, three of the leading teams were from clinical institutes/departments (Lee et al., 2006; Frakking et al., 2022; Park et al., 2022), while the other eight studies were either from departments of bioengineering/biomedical engineering (Lee et al., 2011) (Merey et al., 2012; Sarraf Shirazi et al., 2012) or departments of electrical and computer engineering (Sejdic et al., 2013; Sarraf Shirazi et al., 2014; Shu et al., 2022). The included studies were published in Dysphagia (Frakking et al., 2022), Journal of Neuroengineering and Rehabilitation (Lee et al., 2006; Merey et al., 2012), Artificial Intelligence in Medicine (Lee et al., 2011), Medical and Biological Engineering and Computing (Sarraf Shirazi et al., 2012; Sarraf Shirazi et al., 2014), IEEE Transaction of Biomedical Engineering (Sejdic et al., 2013), IEEE Journal of Biomedical and Health Informatics (Shu et al., 2022), and Scientific Reports (Park et al., 2022) that spanned across the clinical, engineering, and interdisciplinary science fields. It should be noted that the majority of the work has been published or is connected to the same research team.

Out of the seven items, the average point of the studies was 5.44, with a standard deviation of 1.13 (Figure 3). All items under applicability concerns were scored since all studies provided physician diagnosis and benchmarking instrument data to justify the patients and/or events. Nearly all papers lost points on the patient selection domain without clarifying whether the participants were recruited consecutively or by random sampling. All except one study conducted the index test and reference test simultaneously, while some studies lost points for not using the same reference standard. Besides, a risk of bias was also found for studies that excluded patients because they could not complete the test or discarded data with problems.

Among the nine included studies, five were further processed for meta-analysis. One study was discarded due to the lack of sample count information (Sejdic et al., 2013). Two studies were discarded because they were not classifying risky swallows (i.e., not per-sample level) but individuals at risk of aspiration (i.e., per-individual level) (Sarraf Shirazi et al., 2014; Park et al., 2022). We eliminated one study because the leave-one-out validation cannot be used to estimate the sample counts (Sarraf Shirazi et al., 2012).

The pooled diagnostic odds ratio was 21.5 (95%CI, 2.7–173.6), which was higher than the cut-off of 10.00 (Deeks, 2001) but not significant. The coupled forest plot and the forest plot of the log diagnostic odds ratio (Figure 4) demonstrated that there were high standard errors within studies that might be due to small sample sizes, in addition to variations between studies. Sensitivity could be as low as 21% (95%CI, 10%–37%) while specificity could be as high as 100% (95%CI, 93%–100%), in individual studies. By observing the SROC plot (Figure 4), it could be seen that the study-level data points dispersed over the ROC space, far away from the summary line, and with a large confidence region, which demonstrated substantial heterogeneity. In view of this, we decided not to conclude the meta-analysis result.

The significance of this study lies in its ability to summarize the accuracy performance and assessment techniques used in computer-aided screening for dysphagia risks using biophysical sensors. While overall accuracy has often been reported as an indicator of model performance, sensitivity is a more clinically important attribute for screening tools to identify those at greater risk (Wirth et al., 2016). Our review demonstrated that the current systems in our included studies were insufficient, with only two studies (with their best optimized models) demonstrating a sensitivity of more than 90%. Interesting, more information seemed not to produce better results (e.g., concatenating clinical history data (Park et al., 2022) and integrating tri-axial accelerometer and acoustic signals (Shu et al., 2022)). There is a need to improve the generalizability of the system with a larger dataset, and optimize the signal processing, segmentation, feature extraction, classifier, and their combinations to improve the accuracy performance.

Clinical heterogeneity could be sought from gender and age-group, the source of dysphagia/aspirations, and the assessment protocols. Gender could be a significant confounder in this scenario, with the Adam’s apple and deeper voice in adult males apparently influencing the throat biomotion and acoustic signals (So et al., 2023). We found no study that submitted gender as an input feature. On the other hand, dysphagic aspiration could be sourced from different pathophysiologies (Wirth et al., 2016). For example, age-related dysphagia (i.e., presbyphagia) demonstrated reduced tongue pressure and delayed triggering of swallow reflection (Rofes et al., 2010), while dementia was characterized by a compulsive eating pattern and a large bolus size (Langmore et al., 2007). Besides, post-stroke patients experienced decreased activation of swallowing because of the lesion of motor neurons (Teismann et al., 2011). These differences in abnormalities might render different signal patterns of aspiration risks, which could be the reason for the large standard error of the study and the high heterogeneity between studies. Lastly, while studies followed a “routine swallowing assessment” protocol by taking different constituencies and thicknesses of food/liquid, the procedure details were vague, and we are uncertain whether all or some swallow trials were selected for the development of computer-aided screening. In addition, only one study accounted for the non-swallowing task. A previous review commented that protocol heterogeneity might hinder the translational potential of wearable technology on swallowing assessment (So et al., 2023) and that a unified framework was necessary to account for both swallowing and non-swallowing activities (Lim et al., 2023).

Besides methodological heterogeneity in terms of instruments, feature extraction, and modeling, some technical issues might exist. The performances between studies were very extreme, ranging from 21.7% to 100%. It is skeptical when accuracy falls below 50%, which is worse than random guessing. While achieving good accuracy is desirable, obtaining perfect accuracy (i.e., 100%) from predictive models is not possible in practice, since they are designed to approximate underlying constitutive relationships by fitting with the stochastic nature of data and algorithms (i.e., a simplified construct related to a part of reality). Skeptical performances could be due to underfitting (datasets too small), overfitting, imbalance classes, misspecification of hyperparameters, and regularization (Boulesteix and Schmid, 2014; Lever et al., 2016; Kaur et al., 2019; Nichols et al., 2019; Weerts et al., 2020). While several studies did not conduct hyperparameter tuning, cross-validation techniques were often used and believed to relieve overfitting but might be prone to data snooping or peeking (Bzdok et al., 2017) and generate biased estimates, especially with small sample size (Vabalas et al., 2019). Five studies recruited fewer than 50 participants (i.e., independent samples). Data samples were subsequently pooled through repeated measurements and data augmentation techniques. Small datasets may produce strongly spurious patterns. As a rule of thumb, 50 samples or 10 samples per feature (Pedregosa et al., 2011; Riley et al., 2020; Scikit-learn developers, 2023) are minimally needed to fit predictive or machine learning models. In practice, more samples are required with higher data dimensionality and the complexity of learning algorithms (Bzdok et al., 2017). Lastly, it is important to scrutinize flaws in the data and models.

There were some limitations in this study. Only English studies were included in this study, which might lead to language bias. Besides, the number of included studies was relatively small, especially since several of them were from the same research team. Their findings might not be independent. Sterne et al. (2011) advised that a minimum of ten studies be reviewed to achieve sufficient power to assess small-study effects as a rule of thumb. On the other hand, we anticipated that flexible electronics, or soft sensors (Jung et al., 2020; Chen et al., 2021c; Gao et al., 2021; Guan et al., 2021), would be included in this review in the first place, but in vain because most of them were still on the research bench from clinical studies. Moreover, we did not conduct an in-depth data synthesis on the signal processing techniques, which warranted another technical review. For the meta-analysis, with the small dataset and unclear risk of consecutive/random sampling, it is likely that the pooled estimates of the meta-analysis lack generalizability and could be misleading. While we endeavored to provide an overall estimate of the area, we discovered that studies had unique methodological characteristics and major differences in the sets of parameters/thresholds. It might not be appropriate to summarize their test performance using meta-analysis. Subgroup analyses were not conducted on different instruments (accelerometers and microphones) and populations (older adults and children) because of the small number of available studies. Besides, studies using a cross-validation approach that did not have well-defined counts of testing sets approximated the confusion matrix based on the averaged results of cross-validation folds and the fold proportion, which might not be viewed as a pertinent method in meta-analysis. With increasing research using machine learning models for diagnostic or screening purposes, there is a need for developing a new meta-analytic approach targeting cross-validation and data resampling.

More effort is yet necessary to improve the accuracy performance of the computer-aided screening systems to identify aspiration risks, in addition to tests on larger sample sizes to ensure generalizability. A similar conclusion has been reached by another review that targeted on systems classifying swallowing and non-swallowing (e.g., speaking, yawning) events (So et al., 2023). Deep learning models were not implemented, which might be due to an insufficient dataset or a lack of a pretrained model. One study utilized the GAN approach to “generate” more data. Future work may consider improving the robustness and establishing protocols for pragmatic exploitation and implementation. Compliance could be a problem, especially with sensors that have to be stuck on the neck (e.g., accelerometers). Patient-centered designs and feasibility studies could be necessary to promote acceptance among patients and caregivers, especially those with dementia (Merilahti et al., 2009; Gold et al., 2018). Furthermore, these biophysical sensors could be incorporated and improve swallowing therapy through gaming, virtual reality, and biofeedback (Li et al., 2016; Mizoguchi et al., 2021; So et al., 2022).