Risk prediction models for enteral nutrition-related aspiration in critically ill patients: A systematic review

Abstract

BACKGROUND

Aspiration is a common and serious complication in critically ill patients receiving enteral nutrition, leading to increased morbidity and mortality. Accurate risk prediction is crucial for timely prevention and management; however, the quality and predictive performance of existing prediction models remain uncertain and require systematic evaluation.

AIM

To systematically identify, appraise, and synthesize evidence on the development, validation, and performance of published risk prediction models for enteral nutrition-related aspiration in critically ill adults.

METHODS

A comprehensive literature search was conducted in PubMed, EMBASE, Web of Science, Cochrane Library, CINAHL, CNKI, Wanfang Data, VIP, and CBM databases from inception to June 2025. Eligible studies included those reporting the development, validation, or impact assessment of multivariate models for predicting aspiration risk in critically ill adults receiving enteral nutrition. Two reviewers independently extracted data and assessed the risk of bias and applicability using the Prediction model Risk of Bias Assessment Tool.

RESULTS

Twelve studies (all from China) reporting 18 prediction models were included. Common predictors were age, Glasgow Coma Scale, mechanical ventilation, and gastric residual volume. Areas under the curve ranged from 0.771 to 0.995 in derivation cohorts, with sensitivity ranging from 0.776 to 0.952 and specificity from 0.695 to 0.952. However, only four models underwent external validation. Prediction model Risk of Bias Assessment Tool assessment rated 10 studies (83.3%) as high risk of bias, primarily due to inadequate sample size, improper handling of missing data, and lack of external validation.

CONCLUSION

Existing aspiration prediction models demonstrate good discrimination but have a high risk of bias and limited external validation. Rigorous validation in diverse populations is urgently needed.

Key Words: Critical illness; Enteral nutrition; Respiratory aspiration; Risk prediction models; Systematic review; Risk factors

Core Tip: This systematic review is the first to comprehensively evaluate risk prediction models for enteral nutrition-related aspiration in critically ill adults. We identified 18 models from 12 studies, with age, Glasgow Coma Scale score, and mechanical ventilation being the most common predictors. Although most models showed acceptable discrimination, the overall risk of bias was high, primarily due to inadequate methodological conduct in the analysis domain of the Prediction model Risk of Bias Assessment Tool. Rigorous external validation and adherence to standardized reporting guidelines are urgently needed before these models can be safely implemented in clinical practice.

INTRODUCTION

Aspiration refers to the entry of food, oropharyngeal secretions, or gastric contents into the lower respiratory tract[1]. Among critically ill patients receiving enteral nutrition, the reported incidence of aspiration can be as high as 50%[2]. Severe aspiration may result in impaired pulmonary function and the development of acute respiratory distress syndrome, thereby increasing patient mortality[3]. Clinically, aspiration is classified as overt or silent, with silent aspiration accounting for more than 75% of all cases[4]. Due to the absence of obvious clinical manifestations and the lack of reliable diagnostic indicators, silent aspiration often goes unrecognized, contributing substantially to the high incidence of pneumonia among critically ill patients[5]. Therefore, early, systematic assessment of aspiration risk, timely identification of contributing factors, and the implementation of targeted preventive measures are of paramount importance in improving patient outcomes[6].

Clinical risk prediction models are statistical tools that estimate an individual’s probability of developing a specific outcome based on multiple predictor variables[7]. By integrating diverse patient characteristics, these models enable risk stratification and inform clinical decision-making[7]. In the context of enteral nutrition-related aspiration, a well-developed prediction model could help clinicians identify patients at highest risk, tailor monitoring intensity, and guide preventive strategies such as modifying feeding protocols or considering alternative enteral access routes.

In recent years, several studies have developed risk prediction models for enteral nutrition-related aspiration in critically ill populations. These models incorporate various combinations of predictors, including patient demographics, disease severity indicators, and treatment-related factors. However, the methodological quality of these studies, the consistency of predictor selection, and the extent of model validation have not been systematically evaluated. Without such evaluation, it remains unclear which models, if any, are sufficiently robust for clinical application.

Therefore, this systematic review aims to: (1) Identify all published risk prediction models for enteral nutrition-related aspiration in critically ill adults; (2) Critically appraise their methodological quality using the Prediction model Risk of Bias Assessment Tool (PROBAST); (3) Summarize model characteristics, predictors, and performance; and (4) Provide recommendations for future model development and clinical implementation.

MATERIALS AND METHODS

Study design

This systematic review of prediction-modeling studies was conducted and reported in accordance with the CHARMS (Critical Appraisal and Data Extraction for Systematic Reviews of Prediction Modelling Studies) framework.

Eligibility criteria

Inclusion criteria: Studies were eligible if they: (1) Enrolled critically ill adult patients; (2) Developed and/or validated a multivariate prediction model for the risk of enteral nutrition-related aspiration; and (3) Used an appropriate study design, including cohort (prospective or retrospective), case-control, or cross-sectional designs.

Exclusion criteria: Studies were excluded if they: (1) Were not published in English or Chinese; (2) Were duplicate publications or lacked an accessible full text; (3) Were review articles, conference abstracts, editorials, or comments; or (4) Did not provide sufficient data for extraction or quality appraisal.

Search strategy

A comprehensive search of nine electronic databases was performed from inception to June 2025: PubMed, EMBASE, Web of Science, Cochrane Library, CINAHL, CNKI, Wanfang Data, VIP Database, and CBM. Search strategies combined controlled vocabulary (e.g., MeSH terms) and free-text keywords mapped to three core concepts: (1) Critically ill patients (e.g., “critical care”, “critically ill patients”, and “ICU patients”); (2) Enteral nutrition (e.g., “enteral nutrition”, “nasogastric feeding”, and “tube feeding”); and (3) Aspiration and prediction (e.g., “aspiration”, “microaspiration”, “risk assessment”, “prediction model”, and “forecasting”). The detailed PubMed search strategy is presented in Figure 1.

Figure 1  The literature search strategy is exemplified by PubMed.

Study selection and data extraction

Two reviewers, trained in evidence-based methods, independently screened titles and abstracts, followed by full-text assessment against the eligibility criteria. Disagreements were resolved through discussion or adjudication by a third reviewer. Data were extracted using a standardized form based on the CHARMS checklist. Extracted items included first author, year of publication, country/region, study objective, design, data source, participant characteristics, details of model development or validation, candidate and final predictors, model performance metrics (e.g., discrimination, calibration), and additional evaluation indices (e.g., sensitivity and specificity).

Quality assessment (risk of bias and applicability)

Methodological quality was independently assessed by two reviewers using the PROBAST[8]. Discrepancies were resolved by consensus or referral to a third reviewer. PROBAST evaluates risk of bias across four domains-participants, predictors, outcome, and analysis, with each domain rated as “low”, “high”, or “unclear” risk of bias. An overall study rating of “low” risk of bias required all domains to be rated “low”; if any domain was rated “high”, the overall risk of bias was deemed “high”. Applicability to the review question was assessed across the first three domains (participants, predictors, outcome) using the same rating scheme.

RESULTS

Studies included after literature search

The initial database search identified 830 records. After removing 218 duplicates using EndNote, 612 unique records remained for title and abstract screening. Following full-text assessment against the predefined eligibility criteria, 12 studies were ultimately included in this systematic review[9-20]. The detailed selection process is illustrated in Figure 2.

Figure 2  Literature screening process.

Characteristics of included studies

The 12 included studies were published between 2019 and 2025, all conducted in China, with three published in English. Regarding study design, five were prospective cohort studies. The main characteristics of the included studies are summarized in Table 1.

Table 1 Characteristics of included studies (n = 12).

Ref.	Country	Study objective	Data source	Study design	Study population
Chen et al[9], 2025	China	Model development	Single center	Retrospective cohort	ICU patients receiving EN
Zhou et al[17], 2025	China	Model development	Single center	Prospective cohort	Critically ill stroke patients
Shi et al[14], 2024	China	Development and validation	Single center	Retrospective cohort	Postoperative glioma patients in ICU
Peng et al[13], 2022	China	Development and validation	Single center	Retrospective cohort	Critically ill ICH patients
Guan et al[10], 2022	China	Development and validation	Single center	Prospective cohort	Severe acute pancreatitis
Jing and Wu[11], 2023	China	Development and validation	Single-center	Prospective cohort	Tube-fed patients
Luo et al[12], 2024	China	Development and validation	Single-center	Retrospective cohort	Mechanically ventilated children
Yu et al[15], 2022	China	Model development	Single-center	Retrospective cohort	Severe traumatic brain injury
Zhang et al[16], 2022	China	Development and validation	Single-center	Retrospective cohort	Patients receiving nasogastric feeding
Sun et al[19], 2020	China	Development and validation	Single-center	Prospective case-control	Patients receiving nasogastric feeding
Wang et al[20], 2024	China	Development and validation	Single-center	Prospective cohort	Acute ischemic stroke patients
Hou et al[18], 2024	China	Development and validation	Single-center	Retrospective cohort	Acute pancreatitis patients

ICU: Intensive care unit; EN: Enteral nutrition; ICH: Intracerebral hemorrhage.

Characteristics of risk prediction models

Across the 12 included studies, 18 distinct prediction models for enteral nutrition-related aspiration were identified. The sample sizes for model development ranged from 60 to 515 critically ill patients, with reported aspiration incidence varying from 9.46% to 57%. Detailed characteristics of these models are presented in Table 2.

Table 2 Characteristics of prediction models for enteral nutrition-related aspiration in critically ill patients (n = 12).

Ref.	Total sample	No. of candidate variables	Aspiration incidence, %	Modeling method	Model presentation	AUC	Sensitivity	Specificity	Validation method	No. of predictors	Predictor variables
Chen et al[9]	500	28	57	Logistic regression	Nomogram	0.82	0.776	0.695	Internal	6	Intubation days, body position, daily EN duration, APACHE II score, sedatives/analgesics, PaO2
Zhou et al[17]	60	8	46.67	Logistic regression						4	Age, impaired consciousness, dysphagia, gastrointestinal dysmotility
Shi et al[14]	379	12	19.44	Logistic regression		0.771			Internal	3	COPD, duration of mechanical ventilation, duration of postoperative coma
Peng et al[13]	368	21	39.95	R software	Nomogram	0.995	0.952	0.842	Internal and external	5	NG tube diameter, gastric residual volume, history of aspiration, NIHSS score, Water Swallow Test grade
Guan et al[10]	296	15	9.46	R software, Python	Random forest, neural network, decision tree, support vector machine, generalized linear regression	0.976			Internal	5	APACHE II score, level of consciousness, nutritional risk, NG tube insertion depth, PLR
Jing and Wu[11]	103	29	20.08	R software	Nomogram				Internal	4	Number of comorbidities, intubation depth, history of aspiration, sedatives/hypnotics
Luo et al[12]	330	13	31.52	R software	Nomogram	0.810			Internal and external	7	Gastric residual volume, mode of mechanical ventilation, feeding volume, level of consciousness, NG tube depth, prokinetics, sedatives
Yu et al[15]	212	11	48.11	Logistic regression	Nomogram				Internal	6	Age, diabetes mellitus, APACHE II score, impaired consciousness, nutritional risk, NG tube length
Zhang et al[16]	220	27	20.9	Logistic regression, R software	Nomogram, CART	0.895, 0.902	0.825, 0.806	0.736, 0.758	Internal	11	Age, history of aspiration, number of comorbidities, NG tube depth, NG tube duration, food source, sedatives/hypnotics, impaired consciousness, complications, serum CRP, serum albumin
Sun et al[19]	515	29	20	Logistic regression, R software	Nomogram, CART	0.93, 0.96	0.909, 0.883	0.886, 0.962	Internal	5	History of aspiration, number of comorbidities, intubation depth, sedatives/hypnotics
Wang et al[20]	359	30	16.9	Logistic regression	Nomogram	0.853			Internal and external	4	Suctioning, brainstem infarction, temporal lobe infarction, Barthel Index score
Hou et al[18]	200	11	12.5	Logistic regression	Nomogram	0.926	0.884	0.852	Internal	5	Body position, level of consciousness, nutritional risk, APACHE II score, NG tube length

APACHE II: Acute Physiology and Chronic Health Evaluation II; EN: Enteral nutrition; COPD: Chronic obstructive pulmonary disease; NG: Nasogastric; NIHSS: National Institutes of Health Stroke Scale; PLR: Platelet-to-lymphocyte ratio; CART: Classification and regression trees; CRP: C-reactive protein; AUC: Area under the receiver operating characteristic curve.

The final models incorporated between 3 and 11 predictor variables, which were categorized into the following five major domains: (1) Baseline patient characteristics: Age, diabetes mellitus, chronic obstructive pulmonary disease, number of comorbidities; (2) Disease severity indicators: Acute Physiology and Chronic Health Evaluation II (APACHE II) score, National Institutes of Health Stroke Scale score, impaired consciousness/level of consciousness, duration of postoperative coma, dysphagia, results of the Water Swallow Test, history of aspiration, nutritional risk, and serum levels of C-reactive protein and albumin; (3) Treatment and intervention-related factors: Duration of mechanical ventilation, mode of mechanical ventilation, use of sedatives, analgesics, or hypnotics, and administration of prokinetic agents; (4) Nutritional support and tube-related factors: Nasogastric tube placement days, tube diameter, insertion depth, daily duration of enteral nutrition, feeding volume, and food source; and (5) Gastrointestinal tolerance and physiological response indicators: Gastrointestinal dysmotility and gastric residual volume.

Of the 18 models, 14 models had either no validation or only internal validation[9-11,15-19], while 4 models were externally validated[12-14,20]. In terms of modeling techniques, one study by Guan et al[10] used Python software, whereas all others employed logistic regression analysis, primarily using R software. The most common model presentation format was a nomogram (n = 10), followed by classification and regression trees (n = 2), random forest (n = 1), neural network (n = 1), decision tree (n = 1), and support vector machine (n = 1). Two studies did not specify the modeling format used[14,17].

The area under the curve (AUC) values for the included models ranged from 0.771 to 0.995, indicating good to excellent discrimination. Reported sensitivity values ranged from 0.776 to 0.952, and specificity values ranged from 0.695 to 0.952. All models reporting AUC achieved an AUC > 0.7, reflecting satisfactory predictive performance; however, three studies did not provide AUC values[11,15,17].

Risk of bias and applicability assessment

The methodological quality and applicability of the 12 included studies were evaluated using the PROBAST. Overall, 10 studies were judged to have a high risk of bias[9-18].

Participants domain: Seven studies were rated high risk, primarily because they relied on retrospective, non-registry data sources, where missing information was common[9,12-16,18].

Predictors domain: Six studies were judged high risk due to the potential recall bias in retrospective data collection, potentially compromising predictor accuracy[9,12-14,16,18]. Two studies were rated as unclear because they lacked details on predictor measurement methods[10,15].

Outcome domain: All studies were rated as low risk for outcome assessment.

Analysis domain: Nine studies were rated high risk, mainly due to inadequate reporting of sample size justification, incomplete handling of missing data, and the absence of strategies to address model overfitting or underfitting[9-15,17-18]. Regarding applicability, all 12 studies were judged to have good applicability to the review question[9-20]. In summary, two studies demonstrated low risk of bias and good applicability, whereas the remaining ten studies[9-18] were characterized by high risk of bias but good applicability[19-20]. Detailed PROBAST assessments are presented in Table 3.

Table 3 Risk of bias and applicability assessment using Prediction model Risk of Bias Assessment Tool (n = 12).

Ref.	Participants	Predictors	Outcome	Analysis	Overall risk of bias	Participants	Predictors	Outcome	Overall applicability
Chen et al[9]	H	H	L	H	H	L	L	L	L
Zhou et al[17]	L	L	L	H	H	L	L	L	L
Shi et al[14]	H	H	L	H	H	L	L	L	L
Peng et al[13]	H	H	L	H	H	L	L	L	L
Guan et al[10]	L	?	L	H	H	L	L	L	L
Jing and Wu[11]	L	L	L	H	H	L	L	L	L
Luo et al[12]	H	H	L	H	H	L	L	L	L
Yu et al[15]	H	?	L	H	H	L	L	L	L
Zhang et al[16]	H	H	L	L	H	L	L	L	L
Sun et al[19]	L	L	L	L	L	L	L	L	L
Wang et al[20]	L	L	L	L	L	L	L	L	L
Hou et al[18]	H	H	L	H	H	L	L	L	L

L: Low risk of bias/good applicability; H: High risk of bias/poor applicability; “?”: Unclear risk of bias.

DISCUSSION

A rigorous assessment of the risk of bias in primary studies is essential for ensuring the scientific validity and reliability of systematic review conclusions[21]. Among the 12 studies included in this review, 10 were rated as having a high overall risk of bias. The most prominent methodological concerns were identified in the participants and analysis domains of the PROBAST tool.

In the participants domain, the high risk of bias primarily arose from the reliance of seven studies[9,12-16,18] on retrospective cohort design. Retrospective studies depend on pre-existing medical records, which are particularly susceptible to selection bias and measurement bias due to incomplete documentation, absence of standardized measurement procedures, and missing data.

The most critical concerns were observed in the analysis domain, where nine studies were rated as high risk. Several key issues were identified. Inadequate sample size: Most studies did not report any sample size calculation or justification[10,11,14-20]. The relatively small sample sizes (ranging from 60 to 515) were likely insufficient for the number of candidate predictors (8-30 variables), increasing the risk of model overfitting. Incomplete reporting of data handling methods: Nine studies failed to clearly describe procedures for handling continuous variables or missing data, undermining reproducibility. Suboptimal predictor selection: Six studies selected predictors solely through univariable analyses, a method that may introduce bias in variable inclusion. Insufficient validation and assessment of model performance: Only four models underwent external validation, while the remainder were either internally validated or not validated at all. The absence of robust external validation can lead to optimistic performance estimates and limits generalizability[22].

Although many existing models reported excellent predictive metrics, their high risk of bias serves as a cautionary signal that these results may be overly optimistic. Direct clinical application of such models may therefore result in inaccurate risk estimations and should be approached with considerable caution.

This review identified 18 prediction models reported across 12 studies, all of which were conducted in China and published within the past five years, reflecting a growing research interest in this area. However, the fact that all included studies originated from a single country introduces important considerations regarding geographic and ethnic generalizability.

Clinical practices, such as enteral feeding protocols, sedation strategies, and criteria for aspiration assessment, may differ substantially across healthcare systems. Likewise, patient characteristics including body habitus, disease spectrum, and genetic factors can vary between populations. Therefore, while these models demonstrated good predictive performance in Chinese cohorts, their applicability to non-Chinese populations remains unknown. Caution is warranted when applying these models to patients in other countries or regions without prior local validation.

Beyond geographic limitations, the evidence collectively indicates that prediction modeling for enteral nutrition-related aspiration remains at an early developmental stage. All included studies were based on single-center cohorts, inherently restricting data diversity and representativeness. Consequently, these models may reflect local clinical practices and patient characteristics, limiting their broader applicability.

Furthermore, the lack of external validation in 11 studies restricts assessment of their true performance and clinical utility. In terms of modeling methodology, most studies used logistic regression to construct nomograms. Notably, only one study by Guan et al[10] explored machine learning algorithms, including random forest, neural network, decision tree, and support vector machine. While these models demonstrated excellent discrimination (AUC up to 0.976), their complexity and “black-box” nature may limit clinical interpretability compared to traditional regression-based nomograms. Furthermore, machine learning approaches typically require larger sample sizes to achieve stable performance; the relatively small sample (n = 296) in Guan et al’s study[10] raises concerns regarding model generalizability.

Predictor variables also varied considerably across the 18 models, with no two models sharing an identical core predictor set. Nevertheless, several commonly identified predictors, such as gastric residual volume, impaired consciousness, mechanical ventilation, and APACHE II score, are consistent with well-established risk factors reported in the literature. This alignment suggests that current models capture clinically meaningful risk determinants. However, the substantial heterogeneity in predictor selection underscores the absence of consensus on a standardized core predictor set and corresponding cutoff thresholds.

Future research should therefore: (1) Prioritize external validation of existing high-performing models in diverse geographic and ethnic populations; (2) Conduct international multicenter studies to assess transportability; and (3) Systematically integrate existing evidence to identify a robust, parsimonious, and clinically relevant core predictor set.

Current research on risk prediction models for enteral nutrition-related aspiration in critically ill patients remains largely at the model development stage, with limited translation into clinical practice. Several factors may contribute to this gap, including model complexity or limited usability, variability across clinical environments, and insufficient awareness among healthcare professionals regarding data-driven decision-support tools.

We provide the following recommendations for future model development: (1) Employ prospective, multicenter study designs that adhere to established methodological frameworks such as PROBAST; (2) Conduct priori sample size calculations to minimize the risk of overfitting; (3) Provide explicit definitions and measurement protocols for all predictors and outcome variables; (4) Standardize and disclose all data handling procedures to ensure transparency and reproducibility; and (5) Incorporate both internal and external validation. Rather than developing new models indiscriminately, future research should prioritize the external validation of existing high-performance models[10,13].

The followings are recommendations for clinical integration: (1) Clinicians should carefully evaluate and apply prediction tools, considering each model’s risk of bias; (2) Preference should be given to models demonstrating low risk of bias and strong external validation[19,20]; and (3) Before widespread clinical implementation, pilot validation studies should be conducted within the target patient population and clinical context.

CONCLUSION

This systematic review identified 18 prediction models for enteral nutrition-related aspiration in critically ill patients. The models demonstrated good discriminatory ability, with AUC values ranging from 0.771 to 0.995 in derivation cohorts. Frequently identified predictors, including gastric residual volume, impaired consciousness, mechanical ventilation, and APACHE II score, align with established clinical risk factors, suggesting that current models capture clinically meaningful determinants. However, the overall methodological quality was suboptimal: 10 of 12 studies (83.3%) were rated as high risk of bias according to PROBAST, primarily due to inadequate sample sizes, inappropriate handling of missing data, and reliance on single-center retrospective designs. Furthermore, only four models underwent any form of external validation, severely limiting confidence in their generalizability.

Importantly, all included studies were conducted exclusively in Chinese populations, which significantly restricts the global applicability of these models. Clinical practices, patient characteristics, and healthcare settings vary across countries and regions; therefore, the performance of these models in non-Chinese populations remains unknown. Until rigorous external validation studies are conducted in diverse geographic and ethnic cohorts, clinicians should apply existing models with caution, particularly in populations outside China, and consider local pilot testing before widespread implementation. Future research should prioritize: (1) External validation of existing high-performing models in international multicenter settings; (2) Development of new models using prospective designs, adequate sample sizes, and transparent reporting; and (3) Establishment of a standardized core predictor set through meta-analysis of existing evidence.

ACKNOWLEDGEMENTS

The authors thank the staff of the Department of Nursing, The Second Affiliated Hospital of Naval Medical University, for their support during the conduct of this study. We are grateful to the librarians who assisted with the literature search and to the reviewers who provided valuable feedback on the manuscript. We also extend our sincere appreciation to the patients and their families who participated in the original studies included in this review.