Abstract
Aim Early enteral nutrition is often recommended for patients with acute stroke who have difficulty with oral intake. This study aimed to develop a predictive model to assess the need for enteral nutrition in older patients with acute cerebrovascular disorders. The model employs a machine learning algorithm using observational parameters related to swallowing ability.
Methods Ninety patients experiencing a cerebrovascular accident for the first time were included in this study. Swallowing function was assessed using the Food Intake LEVEL Scale. Nine specific variables were used to create a model for determining the need for enteral nutrition. Initially, variable selection was conducted through correlation analysis. Subsequently, the data were randomly divided into training and test groups. Five machine learning methods were applied to identify the most effective algorithm: logistic regression, decision tree, random forest, support vector machine, and XG Boost.
Results Through correlation analysis, we identified the independent variables Functional Independence Measure, motor and cognitive scores and speech intelligibility. The logistic regression model demonstrated high performance (accuracy, 0.82; area under the curve, 0.82).
Conclusion We demonstrated that a predictive model, employing machine learning and integrating Functional Independence Measure motor and cognitive scores and speech intelligibility, exhibits superior predictive efficacy and ascertains the necessity for enteral nutrition. This model can be expediently appraised even by individuals not specialized in dysphagia. Additionally, it is applicable to patients who are incapable of adhering to conventional swallowing assessment protocols owing to compromised consciousness or cognitive impairments, or those with an exceptionally elevated risk of aspiration.
1. Introduction
Stroke is a leading cause of disability and mortality worldwide that often results in complex healthcare challenges, particularly for older patients. One significant complication in acute stroke care is dysphagia, a condition impairing the ability to swallow, which is frequently observed in this patient population (Gordon et al., 1987; Horner & Massey, 1988). Dysphagia increases the risk of aspiration pneumonia (ECRI, 1999) and frequently culminates in insufficient oral nutritional intake. Therefore, it is recommended to assess and commence enteral nutrition promptly during the course of a stroke. Early initiation of enteral nutrition is crucial in stroke care, as it mitigates the risks of malnutrition and facilitates recovery (Burgos et al., 2018). Nevertheless, assessment of swallowing ability and the subsequent decision to initiate tube feeding frequently necessitate professional evaluation and experienced judgment, typically provided by a dysphagia specialist.
In numerous healthcare settings, particularly during off-hours such as weekends, access to dysphagia specialists is limited. This underscores the need for nursing staff to play a role in the preliminary assessment of dysphagia. Nurses, as primary caregivers, hold a crucial position in identifying swallowing difficulties and initiating early interventions. In addition, 24/7 dysphagia screening, conducted by proficiently trained nurses, can reduce the incidence of aspiration pneumonia in hospitalized patients who have experienced a stroke, consequently reducing the duration of hospitalization (Palli et al., 2017). The Gugging Swallowing Screen (Trapl et al., 2007), the Mann Assessment of Swallowing Ability (Carnaby-Mann & Lenius, 2008), and the Toronto Bedside Swallowing Screening Test (Martino et al., 2009) have each been corroborated by substantial evidence, affirming their efficacy for dysphagia screening.
However, the use of these assessment tools might pose challenges in patients who are incapable of adhering to testing protocols owing to diminished consciousness or cognitive impairment, or in those with a significant high risk of aspiration. Such patients are frequently encountered in the acute phase of stroke (Giacino et al., 2014; Aurore Thibaut et al., 2020). Furthermore, the scarcity of nursing professionals has become a societal issue (Ross, 2022; Yatsu & Saeki, 2022; Kolanowski et al., 2023). Numerous healthcare facilities encounter challenges in administering meticulous dysphagia screening to patients experiencing acute strokes, on a round-the-clock basis throughout the year, owing to the onerous responsibilities of nurses. Therefore, it is imperative to possess a tool capable of swiftly estimate the swallowing ability of patients with acute stroke, grounded in precise clinical observation, and that can ascertain the most effective strategy for nutritional administration (Donovan et al., 2012).
Recently, there has been a burgeoning in the number of studies documenting the application of machine learning for precise prognostication in the medical sector. Machine learning, a sophisticated theoretical framework, is employed to formulate models that infer unknown data based on historical datasets. This approach distinguishes itself by selecting the most efficacious algorithm from a plethora of predictive methodologies, thereby crafting superior predictive models. Unlike traditional statistical methods, machine learning algorithms dynamically enhance the precision of predictions for future data based on past information. Consequently, this leads to the development of more resilient and accurate models for predicting unknown data, compared to that of conventional techniques. Models employing observational parameters can be quantified in a comparatively brief period as explanatory variables. If these type of models are utilized to evaluate swallowing proficiency and ascertain the necessity of enteral nutrition in patients with acute stroke, it may play a pivotal role in decreasing the frequency of aspiration and shortening the duration of hospital stay for those patients admitted because of stroke.
The objective of this study is to develop a novel predictive model, utilizing advanced machine learning algorithms, to estimate swallowing capacity and determine the necessity for enteral nutrition in geriatric patients with acute cerebrovascular disease. By leveraging observational parameters related to swallowing ability, this model aims to provide a reliable tool for healthcare professionals, particularly nurses, to make informed decisions about initiating tube feeding, thereby improving patient outcomes in acute stroke care.
2. Methods
2.1. Study design
This study was a retrospective observational cohort analysis. Ethical approval was obtained from the Clinical Research Ethics Committee of Kanazawa medical university hospital (approval number: 78), ensuring adherence to the Declaration of Helsinki. Informed consent was obtained from all participants or their legal guardians.
We excluded from the study patients with: subarachnoid hemorrhage, brainstem or cerebellar lesions, previous treatment of head and neck tumors or esophageal cancer, a speech intelligibility difficult to measure because of the effects of aphasia or apraxia, and incomplete measurement data and details of which will be elaborated upon subsequently.
2.2. Patients
We included a total of 134 patients with first-ever stroke who were subjected to videoendoscopic evaluation of swallowing (VEES) or videofluoroscopic swallowing study (VFSS) by rehabilitation physicians at our institution between October 1, 2017 and October 31, 2020.
Based on VEES and VFSS results, which are considered reference standards for assessing swallowing ability (Langmore, 2003), a meal plan was developed for the patients (mean time from onset of stroke to examination: 5.61, standard deviation [SD]: 3.84 days). A week following the initiation of swallowing rehabilitation tailored to the specific meal plan, a seasoned speech-language pathologist with a minimum of five years of experience in the field of swallowing rehabilitation conducted an assessment of each patient’s Food Intake LEVEL Scale (FILS) (Kunieda et al., 2013) and speech intelligibility. The five-point scale of speech intelligibility, as described by Yorkston et al. (1992), serves as a critical measure specifically designed to assess speech and articulation abilities in patients with dysarthria, a condition often characterized by impaired movement of the muscles used for speech production. This scale, ranging from 1 (indicating no difficulty) to 5 (where speech is essentially unintelligible), meticulously evaluates the clarity with which a speaker’s intended message is conveyed and understood by a listener. In the context of dysarthria, where patients may experience varying degrees of speech impairment due to neurological disorders or muscle damage, this scale provides a nuanced and sensitive tool for quantifying the impact of these impairments on speech intelligibility. An occupational therapist evaluated each patient’s Functional Independence Measure (FIM) and Brunnstrom recovery stage (BRS) a day after assessing their swallowing ability. Additionally, the demographic data amassed comprised age, sex, stroke type (ischemic or hemorrhagic), Charlson comorbidity index (Charlson et al., 1987), and duration since stroke onset.
These elements were quantifiable even in patients who could not adhere to the assessor’s instructions because of diminished consciousness or cognitive dysfunction. This aspect was the fundamental aim of this study.
2.3. Model training and validation based on machine learning
Initially, based on the FILS scores, patients were systematically classified into two groups: those in need of enteral nutrition (FILS scores 1–6) and those without such a requirement (FILS scores 7–10). The designation assigned to this variable was “necessity of enteral nutrition.” Additionally, using the FILS severity classification, patients were stratified into four distinct categories: “Difficult oral intake” (FILS scores 1–3), “Combined oral intake and alternative nutrition” (FILS scores 4–6), “Oral intake only” (FILS scores 7–9), and “Normal” (FILS score 10). This variable was termed “severity.” Subsequently, we performed a comprehensive correlation analysis to evaluate the link between the two variables—”necessity of enteral nutrition” and “severity”—and various factors, such as age, duration since stroke onset, Charlson comorbidity index, FIM motor and cognitive scores, speech intelligibility, and BRS for arms, legs, and hands.
In the second phase of the research, our focus shifted to the development and validation of a predictive algorithm tailored to assess the swallowing capacity in patients with acute stroke. The methodology adopted for this investigation is systematically detailed in Figure 1. To ensure the robustness of the prediction model, a stratified five-fold cross-validation technique was employed. This involved dividing the entire dataset into five equal and random segments. In each iteration, one segment was allocated as the testing set, while the others served as training data. The combination of these sets resulted in the creation of a single comprehensive dataset. Each segment was used once as a test set and four times as training data, thus generating five distinct datasets. The model employed “necessity of enteral nutrition” and “severity” as dependent variables and selected variables that exhibited a correlation coefficient of 0.55 or higher with the dependent variable as independent variables. After z-score normalization of the independent variables, five distinct machine learning algorithms were applied to each dataset: logistic regression, support vector machine, decision tree, random forest, and XGBoost. Hyperparameters were finely tuned through a grid search methodology to construct an individual predictive model for each dataset. Grid search, an exhaustive algorithm, was used to identify the optimal combination of hyperparameters, testing every possible value combination in a predefined grid.
The flow of model training and validation based on machine learning.
The effectiveness of each combination was appraised using the mean accuracy, area under the curve (AUC), and F1 score. The AUC metric was employed to assess the binary classification of “necessity of enteral nutrition,” whereas the F1 score was used to assess the multi-class classification of “severity.” The F1 score is the harmonic mean of precision and sensitivity, providing a single metric that balances both aspects. This is crucial in situations where an equal importance is placed on minimizing false positives and negatives. This metric ranges from 0 to 1, where 1 indicates perfect precision and sensitivity, and 0 indicates the poorest precision and sensitivity. Sensitivity and specificity were gauged for the “necessity of enteral nutrition” to facilitate a comparative analysis of its efficacy over other dysphagia screening protocols.
The analysis was conducted using Python 3.10 (Python Software Foundation, Wilmington, Delaware, USA), with a custom-written code for this purpose. The code has been made publicly accessible to ensure reproducibility (https://github.com/kokamoto46/speech_audiology_prediction).
3. Results
After excluding patients with subarachnoid hemorrhage, a total of 90 patients (52 men and 38 women; mean age, 79.4 years; SD 10.5 years) were included in the study. Table 1 summarizes the patient details and predictive variables. The lesion sites were on the left hemisphere in 59 patients and on the right hemisphere in 31 patients. Furthermore, 46 patients had cerebral infarction, and 44 had cerebral hemorrhage.
Correlation analysis revealed significant correlations between the “necessity of enteral nutrition” and “severity” and key variables, such as duration since stroke onset, Charlson comorbidity index, FIM motor score, FIM cognitive score, speech intelligibility, and BRS for the arm, hand, and leg (Table 2).
Consequently, based on these outcomes, FIM motor score, FIM cognitive score, and speech intelligibility were selected as the independent variables for further investigation.
Table 3 presents the average accuracy, AUC, sensitivity, and specificity for five machine learning models concerning “necessity of enteral nutrition.” The logistic regression model exhibited superior efficacy, with an accuracy of 0.82, AUC of 0.82, sensitivity of 0.82, and specificity of 0.83. The receiver operating characteristic curves for the model are illustrated in Figure 2.
The receiver operating characteristic curves for the model predicting the necessity of enteral nutrition. The area under the curve for this model registered at 0.82, with a sensitivity of 0.82, and a specificity of 0.83.
Table 4 presents the average accuracy and mean F1 scores for the respective models in the classification of “severity.” Among these, logistic regression was the most effective, with an average accuracy of 0.67 and an F1 score of 0.63.
4. Discussion
This study demonstrates that the logistic regression model incorporating FIM motor and cognitive scores along with speech intelligibility, exhibits sufficient predictive efficacy and is apt for forecasting the necessity of enteral nutrition in geriatric patients with acute stroke. The logistic regression model demonstrated a sensitivity and specificity of 0.82 and 0.83, respectively. These metrics are comparable to those of the Modified Water Swallow Test (sensitivity, 0.70; specificity, 0.88) and the Gugging Swallowing Screen (sensitivity, 1.0; specificity, 0.50), which are established dysphagia screening tests. Consequently, the performance of the logistic regression model is not substandard relative to existing dysphagia screening methodologies (Artiles et al., 2020). A salient feature of this model is that it rules out the need for patients to perform intricate instructional tasks or to ingest food or liquids.
In the domain of machine learning methodologies, the strategic selection of variables is paramount, primarily aimed at augmenting the accuracy of predictive outcomes. However, this process encounters a significant impediment: the complexity involved in deciphering the interrelations between the chosen variables and the predictive elements within the model. Furthermore, understanding the foundational reasons underlying these predictions are challenging. To address this, alongside the meticulous analysis of the data that inform these predictive models, it is imperative to engage in a comprehensive and detailed discourse. Such an approach is essential to enhance the transparency and interpretability of the model, thereby ensuring that the predictions are accurate, intelligible, and justifiable, especially in contexts demanding high precision and reliability, such as in the medical and scientific fields.
The correlation established between swallowing proficiency and FIM scores in acute stroke patients is underpinned by a comprehensive understanding of the interplay between motor and cognitive functions and swallowing ability. The FIM, encompassing both motor and cognitive domains, serves as a holistic measure of a patient’s functional independence, directly influencing their ability to manage oral intake. The documented role of the FIM as a prognostic tool (Broadley et al., 2003) aligns with this understanding, suggesting that improvements in FIM scores can be indicative of potential enhancements in swallowing capabilities. This is particularly relevant in acute stroke patients, where motor and cognitive impairments are common and can significantly impact swallowing mechanics. In recent years, studies have indicated that comprehensive whole-body exercise regimens for stroke patients not only enhance their FIM motor scores but also improve swallowing capabilities (Yoshimura et al., 2020). Additionally, the research cited involving elderly nursing home residents (Pu et al., 2017) further substantiates this link by highlighting the association between reduced activities of daily living (ADL) capacity and dysphagia. Lower ADL scores often correlate with increased stroke severity, which can exacerbate swallowing difficulties due to affected posture, respiratory function, and alertness. This relationship suggests that ADL capacity, as a reflection of overall functional ability, can be a critical indicator of swallowing function.
The distinctive attribute of the FIM, evaluating ADL executed in everyday settings, makes it suitable for appraising patients unable to comply with the test because of impaired consciousness or cognitive functions. Furthermore, its high reliability is maintained even when administered by non-specialist evaluators, thereby reducing score variability attributable to different assessors (Hamilton et al., 1994; Ottenbacher et al., 1996). It has been established that FIM scores are significant predictors of swallowing function improvement in patients with acute stroke (Nakayama et al., 2014). Consequently, ongoing monitoring of FIM scores is advantageous for dysphagia management.
The claim that speech intelligibility is a critical factor in assessing swallowing ability, particularly in patients with acute stroke, is underpinned by several key considerations. First, acute stroke often results in motor score and sensory deficits affecting the oral organs, such as the lips and tongue, which are crucial for both speech and swallowing. This overlap suggests a common neurological basis for these functional impairments. Additionally, conditions such as velopharyngeal incompetence and respiratory dysfunction, which are common in stroke, further complicate the coordination required for safe swallowing (Wang et al., 2018). The statistical association of dysarthria with dysphagia in patients with acute stroke (45.2%) underscores this relationship, indicating that speech impairments can be a reliable predictor of swallowing difficulties (Bahia et al., 2016).
Furthermore, the simplicity and non-invasiveness of assessing speech intelligibility— achievable through basic conversation without the need for complex instructions or exposure to aspiration risks—make it a practical and safe screening tool. This approach is particularly advantageous in patients with stroke who may also suffer from higher brain dysfunctions, making more complex assessments challenging. Thus, the use of speech intelligibility as a predictor not only aligns with clinical observations and research findings but also offers a pragmatic low-risk assessment method to a vulnerable patient population (Hirsch et al., 2022).
The proposed predictive models for “necessity of enteral nutrition” and “severity” are pivotal in tailoring dysphagia treatment plans for patients with acute stroke, addressing the varying degrees of swallowing impairment. The “severity” model, in particular, allows for a nuanced categorization of patients based on their swallowing capabilities. Patients in the “Difficult oral intake” category typically exhibit significant swallowing challenges, necessitating enteral nutrition to ensure adequate nutritional intake and non-food swallowing training to enhance their swallowing mechanics. In cases classified as “Combined oral intake and alternative nutrition,” patients possess some capability for oral intake but still require enteral nutrition support, complemented by a diet of uniform consistency to minimize swallowing difficulties. For patients in the “Oral intake only” category, a single consistency diet is often sufficient, as it simplifies the swallowing process while still allowing oral feeding. Individuals deemed “Normal” can safely consume a regular diet without restrictions, indicating a near or complete recovery of swallowing function. This stratified approach, based on severity, ensures that each patient receives a personalized treatment plan, optimizing their recovery potential and meeting nutritional needs while minimizing the risk of complications, such as aspiration.
4.1. Limitations and future research
The limitations and potential improvements of the machine learning model for assessing swallowing ability in patients with acute stroke is substantiated by various key points. First, the current model excludes screening tests for swallowing assessment. This exclusion may be a potential weakness, as these tests provide critical information about a patient’s risk of aspiration, a common and serious complication in dysphagia. Incorporating parameters such as the patient’s arousal state, cough reflex strength, and results from practical screening tests could significantly enhance the model’s predictive accuracy. This holistic approach would offer a more comprehensive assessment, mitigating the risks associated with relying solely on machine learning predictions.
Additionally, the study’s findings on the negligible impact of age on swallowing function are limited by the high mean age of the study cohort (Logemann, 1990). Including a broader age range, especially younger patients, in future larger-scale studies could provide more insightful data on age-related swallowing function variations.
Furthermore, the accuracy of the models, especially the “severity” prediction model, needs refinement for clinical utility. Predictive models for the “necessity of enteral nutrition” are confronted with significant variance in the sensitivity. To address this issue, the implementation of the XG Boost model is recommended due to its relatively lower variance in the sensitivity, or alternatively, the assembly of a larger and more uniform data set is suggested. The lack of external validation and reliance on data from a single center also raise concerns about the generalizability of the findings. Employing stratified five-fold cross-validation minimizes analytical errors but does not compensate for the potential biases of a single-center study. Therefore, future studies should aim to validate these models in multi-center studies, using published code to ensure reproducibility and reliability across different clinical settings.
5. Conclusions
We illustrated that a predictive model, employing machine learning and integrating FIM motor score, FIM cognitive score, and speech intelligibility, exhibits superior predictive efficacy and is instrumental in ascertaining the necessity for enteral nutrition.
A key strength of this model is its simplicity and practicality. It circumvents the need for patients to undergo complex tasks or consume food and liquids for assessment, making it particularly suitable for those with impaired consciousness or cognitive functions. This aspect enhances its applicability in diverse clinical scenarios, especially in acute stroke care.
Furthermore, the development of predictive models for both the necessity of enteral nutrition and the severity of swallowing impairment allows for a more tailored approach to dysphagia management. This stratification ensures that each patient receives a personalized treatment plan, optimizing their recovery potential and nutritional intake while minimizing risks.
In summary, this study contributes a valuable tool to the field of acute stroke care, offering a practical, efficient, and reliable method for assessing the need for enteral nutrition. Its integration of easily assessable parameters makes it a promising model for widespread clinical use, with the potential to significantly improve patient outcomes in stroke rehabilitation.
Data Availability
The data used to support the findings of this study are available from the corresponding author upon request.
Data availability
The data used to support the findings of this study are available from the corresponding author upon request.
A web-based application is available that leverages machine learning models, specifically designed for predicting the need for enteral nutrition in patients with acute stroke. This application utilizes FIM motor and cognitive scores, along with an assessment of verbal intelligibility, to make its predictions (https://speechaudiologyprediction-c9xnrvhnld53qc7nbayfku.streamlit.app/).
References
