Implementation of an emergency nursing pathway reduces emergency department length of stay in trauma

Abstract

Timely, organized emergency care is critical for patients with severe trauma, particularly in primary hospitals where resources and workflows may be constrained. This study evaluated the effect of implementing an emergency nursing pathway (ENP) on emergency management of injured patients in a county-level hospital and developed a predictive model to identify factors associated with outcomes. We retrospectively reviewed patients with trauma treated at The First People’s Hospital of Jiande between August 2023 and July 2025. Patients were allocated at random into a training cohort (70%) and a validation cohort (30%). The Boruta algorithm was applied in the training cohort to select relevant predictors, which were then used to train 10 machine learning (ML) classifiers; model performance was evaluated by area under the receiver operating characteristic curve, calibration curves, and decision curve analysis. Model interpretability was assessed using SHapley Additive exPlanations (SHAP) values to quantify each variable’s contribution to predictions. P < .05 considered significant. Implementation of the ENP was associated with a statistically significant reduction in emergency department length of stay compared with routine care (P < .05). Using Boruta-selected predictors, the XgBoost classifier exhibited the best discrimination among tested ML models (area under the curve = 0.669; 95% confidence interval: 0.632–0.705) and was therefore selected as the primary predictive model. SHAP analysis indicated that injury severity score and ENP contributed most to model predictions, followed by mechanism of injury, severity of illness, and method of visiting the hospital. In this single-center retrospective study, application of an ENP for severely injured patients in a primary hospital was associated with shorter emergency department length of stay. These findings support ENP implementation to streamline care and suggest that ML-based tools may help risk-stratify trauma patients in resource-limited settings. Prospective multicenter validation is warranted.

1. Introduction

Severe trauma remains a leading cause of death and disability worldwide, and timely, well-organized emergency care is critical to limit early morbidity and streamline resource use – especially in county-level hospitals where staffing and diagnostic capacity are often constrained.^[1–7] Emergency nursing pathway (ENP) are structured, protocol-driven workflows designed to standardize and expedite nursing interventions during emergency resuscitation.^[8,9] By delineating role responsibilities, sequencing critical tasks, and facilitating early interventions and interprofessional communication, ENP have the potential to streamline emergency department (ED) processes, reduce time to key interventions, and improve clinical outcomes.^[8,10–12] While studies in tertiary centers have reported benefits of pathway-based care for specific emergency conditions, evidence evaluating ENP implementation for trauma in grassroots or county-level hospitals is limited.^[8–12]

Concurrently, the emergence of machine learning (ML) methods offers new opportunities to identify prognostic factors and build predictive tools that can support clinical decision-making in the ED. Robust feature-selection techniques such as the Boruta algorithm can reduce dimensionality and mitigate overfitting by identifying the most relevant predictors from routinely collected clinical variables. Integration of ML-derived risk models with Boruta algorithm can further translate statistical prediction into actionable clinical guidance.

In this retrospective single-center study, we evaluate whether implementation of an ENP is associated with shorter emergency department length of stay (EDLOS) among trauma patients presenting to a county hospital, and we develop and validate an interpretable ML model to identify factors associated with prolonged EDLOS. We hypothesized that ENP implementation would be associated with shorter EDLOS stay times and that an interpretable ML model would identify injury severity and process-related variables as principal drivers of outcome, thereby offering a pragmatic tool to support trauma care in grassroots hospitals.

2. Methods

2.1. Data sources

This retrospective cohort study used routinely collected clinical data from The First People’s Hospital of Jiande. We screened all consecutive patients with trauma who presented to the ED between August 1, 2023 and July 31, 2025.

Data were extracted from the hospital electronic medical record, the ED nursing documentation system, and the institutional trauma registry. Extracted variables included demographic information (age and sex), mechanism of injury, injury severity score (ISS) and EDLOS (defined as time from ED arrival to transfer out of the ED to the operating room, intensive care unit, general ward, or discharge).

Two trained investigators independently abstracted and entered data into a predesigned database. Discrepancies were resolved by consensus or adjudicated by a senior investigator. All patient identifiers were removed prior to analysis to ensure confidentiality. This retrospective study was conducted in accordance with the Declaration of Helsinki and was approved by the Ethics Committee of the First People’s Hospital of Jiande. The requirement for informed consent was waived due to the retrospective nature of the study and the anonymization of patient data.

2.2. Grouping, inclusion and exclusion criteria

A total of 2703 patients meeting the study case identification criteria were included and retrospectively assigned to 2 cohorts according to the documented nursing practice during initial ED management (EDLOS ≤ 2 hours vs EDLOS > 2 hours).

Inclusion criteria: presentation to the ED of the First People’s Hospital of Jiande between August 1, 2023 and July 31, 2025 with injury as defined by institutional criteria (patients presenting to the ED with traumatic injury and with complete ED arrival and disposition timestamps, while nontraumatic presentations and isolated minor injuries not requiring ED trauma management were excluded). Availability of complete ED nursing and medical records allowing ascertainment of primary outcome measures (time stamps for ED arrival and disposition).

Exclusion criteria: declared dead on arrival to the ED. Transfer to another facility prior to completion of initial ED management (interhospital transfer) or transfer into the ED after initial resuscitation at another institution. Missing or incomplete key clinical or process data such that ED length of stay or group assignment could not be reliably determined. Nontraumatic presentations, isolated minor injuries not meeting the institutional definition of trauma, or patients who refused care.

Baseline demographic and clinical characteristics were compared between the 2 groups to assess group comparability. Any discrepancies in group assignment or data completeness were adjudicated by a senior investigator prior to analysis.

2.3. Statistical analysis

All statistical analyses were performed using R software version 4.5.0 (R Foundation for Statistical Computing). The chi-square test was used for categorical variables, and the t test was used for continuous variables. Continuous variables are reported as mean ± standard deviation, and categorical variables are expressed as percentages (%).

The Boruta algorithm is a statistical method for feature selection that is widely applied in data analysis across various fields. Its core concept involves assessing the importance of each feature using the random forest algorithm and comparing it with randomly generated “shadow features” to determine which features are significant. The advantage of the Boruta algorithm lies in its ability to handle high-dimensional data and identify all features related to the response variable, not just the most significant ones. After excluding cases with missing values, the dataset was randomly divided into a training cohort (70%) and a validation cohort (30%). In the training cohort, we employed the Boruta algorithm with 1500 trees to ensure reliable feature selection and incorporated the selected variables into ML models.

Ten different ML algorithms were included: logistic regression, support vector machine, gradient boosting machine, neural network, random forest, Xgboost, K-nearest neighbors, Adaboost, light gradient boosting machine, and CatBoost. The performance of each ML model was evaluated using the area under the curve (AUC) as well as accuracy, sensitivity, specificity, precision, and F1 score.

To elucidate the relative importance of each feature in the best-performing ML model, we employed SHapley Additive exPlanations (SHAP) visualization analysis. SHAP values, derived from game theory, provide a fair attribution of each feature’s contribution to model predictions. By decomposing model outputs into contributions from individual features, SHAP analysis offers insights into which variables most significantly drive the model’s predictive power. In summary, integrating Boruta-selected variables into multiple ML models, combined with rigorous performance evaluation and SHAP-based feature importance analysis, enabled us to identify the most effective risk prediction model and elucidate the key factors influencing its performance.

2.4. Implementation of the ENP

The ENP was implemented according to the following standardized procedures:

1. Prehospital assessment and alert

Upon receipt of the dispatch task, the ambulance team (120 emergency service) rapidly assesses the casualty(ies), initiates appropriate immediate treatment, and activates the prehospital trauma warning system.

The prehospital physician reports the patient’s condition via the alert system, including the number of casualties, anatomic regions injured, and vital signs.

The ambulance is equipped with a real-time visual information exchange/telemetry system to continuously monitor the patient’s status and to transmit physiologic data and a predicted time of arrival to the receiving hospital.

1. In-hospital preactivation and preparation

Based on the information received from the prehospital warning system, the ED trauma nursing team preactivates the in-hospital trauma “green channel” for patients meeting severe trauma criteria.

Specialist physicians are summoned to initiate a multidisciplinary team consultation. A dedicated trauma wristband for the patient is printed and prepared.

Resuscitation drugs, blood products, and required equipment are prepositioned and checked in advance according to the pathway checklist.

1. Immediate actions on arrival and time stamping

On patient arrival at the trauma center, time-sensitive examinations and procedures are performed according to pathway time targets, including bedside focused assessment with sonography for trauma, urgent computed tomography when indicated, preparation for transfusion, and airway management such as endotracheal intubation. These items are reported here to describe the ENP workflow and were not included as variables in the analyses.

The trauma application system, which was already integrated into the hospital’s routine workflow during the entire study period, was used to scan the patient wristband and automatically record arrival and procedure timestamps to ensure adherence to pathway benchmarks and facilitate quality monitoring.

1. Post-resuscitation decision-making and disposition

After initial resuscitation and advanced trauma life support measures, patients who remain hemodynamically unstable are promptly evaluated for operative intervention. Those requiring operative control undergo damage-control surgery.

Patients who do not require immediate surgery are transferred directly to the intensive care unit for continued resuscitation, completion of diagnostic workup, establishment of a definitive diagnosis, and formulation of a definitive treatment plan.

3. Results

3.1. Characteristics of participants

A total of 2703 trauma patients were included in the analysis. Patients were stratified by EDLOS into an EDLOS ≤ 2 hours group (n = 685) and an EDLOS > 2 hours group (n = 2018) (Table 1). Overall, 1281 patients (53%) were managed under the ENP. Implementation of the ENP was associated with a higher proportion of patients achieving an EDLOS ≤ 2 hours (P < .001).

Table 1.

Baseline characteristics of trauma patients stratified.

Characteristic	Overall N = 2,703	EDLOS ≤ 2 h N = 685	EDLOS > 2 h N = 2,018	P value
ENP				<.001
Yes	1281 (47%)	371 (54%)	910 (45%)
No	1422 (53%)	314 (46%)	1108 (55%)
Sex				.554
Male	1624 (60%)	405 (59%)	1219 (60%)
Female	1079 (40%)	280 (41%)	799 (40%)
ICU				.539
Yes	48 (1.8%)	14 (2.0%)	34 (1.7%)
No	2655 (98%)	671 (98%)	1984 (98%)
Mechanism of injury				<.001
Falling from a height	1370 (51%)	330 (48%)	1040 (52%)
Traffic accident	918 (34%)	274 (40%)	644 (32%)
Other	415 (15%)	81 (12%)	334 (17%)
Severity of illness				.034
Non emergency	53 (2.0%)	6 (0.9%)	47 (2.3%)
Emergency	2466 (91%)	624 (91%)	1842 (91%)
Critical	138 (5.1%)	44 (6.4%)	94 (4.7%)
Endangered	46 (1.7%)	11 (1.6%)	35 (1.7%)
Method of visiting the hospital				<.001
Come to the hospital by yourself	1353 (50%)	277 (40%)	1076 (53%)
Call an ambulance	1061 (39%)	347 (51%)	714 (35%)
External hospital referral	246 (9.1%)	47 (6.9%)	199 (9.9%)
Other	43 (1.6%)	14 (2.0%)	29 (1.4%)
Age	58.77 ± (19.01)	58.53 ± (19.67)	58.86 ± (18.79)	.743
ISS	5.98 ± (5.29)	6.04 ± (5.92)	5.96 ± (5.06)	.621

EDLOS = emergency department length of stay, ENP = emergency nursing pathway, ICU = intensive care unit, ISS = injury severity score.

Sex distribution (P = .554), mean age (P = .743), mean ISS (P = .621), and intensive care unit admission rates (P = .539) showed no statistically significant differences between groups. In contrast, mechanism of injury (P < .001), severity of illness (P = .034), and method of visiting the hospital (P < .001) differed significantly between groups (Table 1).

3.2. Feature selection by the Boruta algorithm

Feature selection was performed using the Boruta algorithm to identify variables of relevance for subsequent analysis. Boruta identified 5 important predictors: mechanism of injury, ENP, severity of illness, ISS, and method of visiting the hospital (Fig. 1).

Figure 1.

Feature selection results from the Boruta algorithm. Boxplots show the distribution of importance Z-scores for each candidate variable across Boruta iterations. Colors indicate Boruta decisions: green = confirmed (important), yellow = tentative, red = rejected, blue = shadow features. Confirmed features (green) include ISS, ENP and other process/clinical variables. Labels on the x-axis: A = Mechanism of injury, B = Severity of illness, C = Method of visiting the hospital. The plot highlights which predictors were retained for downstream modeling. ENP = emergency nursing pathway, ICU = intensive care unit, ISS = injury severity score.

Figure 2 shows that the confirmed predictors consistently demonstrated higher importance than the shadow features across Boruta iterations, supporting the stability of the selected features.

Figure 2.

Score variation plot from the Boruta algorithm. Lines depict the evolution of feature importance Z-scores across Boruta classifier runs (iterations). Confirmed features show consistently high Z-scores above their shadow counterparts, whereas rejected features remain near or below the shadow feature range. This panel illustrates the stability of feature selection across iterations.

3.3. ML model development and validation

The performance metrics for all models are summarized in Table 2. Among the evaluated algorithms in the validation cohort, the Xgboost model achieved the highest discrimination (AUC = 0.669, 95% confidence interval: 0.632–0.705) and demonstrated consistently strong performance across secondary metrics (Table 2, Fig. 3A and B).

Table 2.

Assessment of performance metrics for 10 ML models.

ML model	Accuracy	Sensitivity	Specificity	Precision	F1
Logistic	0.598	0.552	0.648	0.635	0.59
SVM	0.604	0.491	0.729	0.668	0.566
GBM	0.601	0.502	0.711	0.658	0.57
Neural network	0.609	0.554	0.669	0.65	0.598
RF	0.588	0.735	0.424	0.586	0.652
Xgboost	0.612	0.507	0.729	0.675	0.579
KNN	0.559	0.596	0.518	0.579	0.587
Adaboost	0.616	0.57	0.667	0.655	0.61
LightGBM	0.609	0.763	0.438	0.601	0.672
CatBoost	0.577	0.387	0.786	0.668	0.49

GBM = gradient boosting machine, KNN = K-nearest neighbors, LightGBM = light gradient boosting machine, ML = machine learning, RF = random forest, SVM = support vector machine.

Figure 3.

(A and B). ROC curves for 10 ML models in the training cohort (A) and validation cohort (B). Each colored curve represents 1 algorithm (legend in figure). The diagonal dashed line indicates no-discrimination (AUC = 0.5). In the validation cohort, XGBoost achieved the highest discrimination among tested models. AUC = area under the curve, CI = confidence interval, GBM = gradient boosting machine, KNN = K-nearest neighbors, LightGBM = light gradient boosting machine, ML = machine learning, ROC = receiver operating characteristic, SVM = support vector machine.

Decision curve analysis indicated that the Xgboost model conferred greater net benefit across a range of clinically plausible threshold probabilities compared with alternative models and default strategies (Fig. 4A and B).

Figure 4.

(A and B) DCA comparing net benefit of the ML models in the training cohort (A) and validation cohort (B). The y-axis shows standardized net benefit and the x-axis shows risk threshold probability. Colored lines correspond to different models; the horizontal line denotes no net benefit. DCA assesses clinical utility across a range of decision thresholds and indicates which models provide greater net benefit relative to default strategies. DCA = decision curve analysis, GBM = gradient boosting machine, KNN = K-nearest neighbors, LightGBM = light gradient boosting machine, ML = machine learning, SVM = support vector machine.

Calibration plots comparing predicted probabilities with observed event rates showed acceptable agreement for the Xgboost model, with model curves approaching the ideal diagonal (y = x) across most risk strata (Fig. 5A and B).

Figure 5.

(A and B) Calibration curves for ML models in the training cohort (A) and validation cohort (B). Observed event proportions (y-axis) are plotted against predicted probabilities grouped by risk bins (x-axis). The diagonal dashed line represents perfect calibration. Deviation from the diagonal indicates miscalibration; the figure shows agreement between predicted and observed risks for the evaluated models. GBM = gradient boosting machine, KNN = K-nearest neighbors, LightGBM = light gradient boosting machine, ML = machine learning, SVM = support vector machine.

3.4. SHAP analysis

SHAP analysis was applied to the top-performing Xgboost model to quantify and visualize the contribution of individual features to model predictions at both the cohort and subject levels.

The global feature importance plot (mean |SHAP value|) indicated that ISS contributed most to model output, followed by ENP, mechanism of injury, severity of illness, and method of visiting the hospital (Fig. 6A). These findings identify ISS and ENP as the principal drivers of the model’s predictions.

Figure 6.

SHAP-based interpretation of the Xgboost model. (A) Global feature importance bar plot showing mean absolute SHAP values (feature ranking by contribution to model output). (B) SHAP summary (beeswarm) plot displaying the distribution and direction of individual feature effects; color indicates feature value (warmer = higher value and cooler = lower value). (C) Waterfall plot for a representative patient illustrating how individual feature contributions (positive or negative) combine to produce the final prediction. (D) Force plot showing the same exemplar’s additive feature effects relative to the model baseline. Categorical encoding note: 1 = yes and 2 = no (applies to binary features such as ENP). ENP = emergency nursing pathway, ISS = injury severity score, SHAP = SHapley Additive exPlanations.

The SHAP beeswarm plot (Fig. 6B) shows feature direction and distribution: higher ISS increases predicted risk, ENP has a substantial but heterogeneous effect across patients, and mechanism of injury variably shifts predictions; severity of illness and arrival method exert smaller influences. SHAP waterfall/force plots (Fig. 6C and D) provide patient-level explanations, with ISS and ENP as the main drivers of individual predictions and the other variables contributing smaller additive effects.

In summary, SHAP analysis identified ISS and ENP as the primary contributors to the Xgboost model’s predictions.

4. Discussion

In this retrospective single-center study, we evaluated whether implementation of an ENP at a county hospital was associated with shorter EDLOS for trauma patients and developed an interpretable ML model to identify factors associated with prolonged EDLOS. Implementation of an ENP for patients with trauma in a primary hospital was associated with a statistically significant reduction in EDLOS compared with routine emergency care. Several baseline demographic and clinical characteristics were similar between groups (including age, sex, and ISS). However, mechanism of injury, severity of illness, and method of visiting the hospital differed significantly. Therefore, although ENP implementation was associated with shorter EDLOS, residual confounding due to baseline differences cannot be fully excluded.

The observed reduction in ED stay is clinically relevant because delays in timely assessment and definitive interventions have been associated with increased morbidity, mortality, and greater resource use after trauma. ENP likely shortens EDLOS through several complementary mechanisms: standardized and expedited triage and preactivation of the trauma team, which reduce time to recognition and initial treatment; early mobilization and prepositioning of diagnostic and therapeutic resources (e.g., focused assessment with sonography for trauma, expedient computed tomography, transfusion, and airway equipment) through preactivation protocols; predefined role allocation and checklist-guided task sequences that reduce omissions and parallelize care; and structured communication and handover processes that facilitate rapid coordination with radiology, operating room, and intensive care.^[13–18] These mechanisms each have support in prior work showing that triage standardization, team preactivation, and checklist-guided care can accelerate assessment and time to critical interventions across emergency settings. Our finding that ISS is the strongest predictor, with ENP status as an important process variable, is concordant with prior studies demonstrating that both physiologic/injury metrics and care process factors determine ED throughput and early outcomes; prior ML studies integrating clinical and process data have similarly highlighted the utility of combining severity indicators with operational variables to predict emergency severity and disposition.^[16,17]

Predefined task allocation and checklists improve workflow efficiency and reduce omissions. Workflow optimization in the ED demonstrably enhances care efficiency and patient satisfaction. A recent study showed that an artificial intelligence-driven ML algorithm was used to identify modifiable predictors of the left-without-being-seen rate; a bundled intervention targeting these factors reduced the daily left-without-being-seen rate by approximately 60 %.^[19] These findings underscore the value of technology-enabled solutions for detecting and resolving process bottlenecks, thereby improving operational performance. The introduction of standardized workflows and communication tools has yielded comparable benefits. Implementation of a structured electronic handover protocol accelerated bed turnover and significantly curtailed patient waiting times.^[20] Likewise, the incorporation of checklist-guided care pathways – for example, in the management of epistaxis – markedly decreased interhospital referral rates.^[21] Collectively, these data indicate that standardization minimizes omissions and streamlines service delivery. Novel nursing strategies further expand the potential for process improvement. Refinement of the emergency nursing workflow shortened door-to-reperfusion intervals and increased salvage rates among patients presenting with ST-elevation myocardial infarction.^[22] An integrated medical–nursing care model produced analogous gains in timeliness and clinical outcome.^[23] Finally, the pivotal role of nursing personnel cannot be overstated. Reconfiguration of nursing leadership responsibilities augmented ED productivity and was associated with a significant reduction in the proportion of unseen patients.^[24] These observations affirm that effective organizational governance of the nursing workforce is indispensable for sustained workflow optimization.

Effective communication and rapid resource mobilization are likely key mechanisms by which the ENP reduced EDLOS in our cohort. By standardizing prehospital alerts, in-hospital preactivation, role allocation, and time-stamped task tracking, the ENP facilitated earlier mobilization of imaging, laboratory, and operative resources and reduced handover-related delays, thereby shortening time to assessment and disposition.^[25–28] These process changes align with broader evidence that structured protocols, care bundles, and preactivation strategies reduce time to critical interventions and improve operational outcomes in emergency and trauma care^[20–24]; likewise, information-sharing tools and standardized handover instruments have been associated with improved communication and fewer preventable delays.^[25,26]

Our findings should be interpreted in light of several limitations. First, the retrospective and nonrandomized design introduces potential selection bias and confounding; although measured baseline variables were similar between groups, unmeasured factors (such as staffing levels, individual clinician experience, concurrent quality improvement initiatives, or temporal trends in care) could have influenced outcomes. Second, this is a single-center study conducted in a primary hospital setting, which may limit generalizability to other institutions with different resource profiles or trauma systems. Third, the primary endpoint provides indirect evidence of clinical benefit; we did not report longer-term outcomes such as in-hospital mortality, functional status, complication rates, or cost-effectiveness. Fourth, reliance on routinely collected clinical records may be subject to information bias from incomplete or inconsistent documentation. Finally, the discrimination of our ML models was modest – the best performing model (Xgboost) achieved an AUC of 0.669 (95% confidence interval: 0.632–0.705) – which limits immediate clinical applicability. Traditional trauma scoring systems such as the ISS and physiologic scores remain well-validated tools for benchmarking injury burden and predicting mortality because they capture key anatomic and physiologic determinants of outcome. ML approaches, by contrast, can integrate a broader array of variables (including process and system factors), model nonlinear relationships and interactions, and produce individualized risk estimates. In our study, ISS was the single most influential predictor by SHAP analysis, but the ML model additionally identified ENP and other process-related variables as meaningful contributors, suggesting that ML can complement traditional scores by incorporating care delivery factors that influence ED throughput. That said, the modest discrimination of our current model (AUC = 0.669) underscores that ML models are exploratory here; they are best viewed as adjuncts to – not replacements for – established scoring systems until further refinement, richer data inputs, and external validation improve predictive performance and clinical utility.

Despite these limitations, this study has practical implications. The positive association between ENP implementation and improved process and early outcome measures supports the feasibility and potential utility of structured nursing pathways in resource-limited settings. Adoption of ENP could be a relatively low-cost intervention focused on training, protocol development, and team coordination that may yield measurable improvements in emergency care delivery. For wider implementation, attention should be paid to local adaptation, staff education, integration with existing trauma protocols, and monitoring of key performance indicators to ensure fidelity and to detect unintended consequences.

Future research should aim to confirm these findings using prospective, ideally multicenter designs with randomization or robust quasi-experimental methods to mitigate confounding. Studies should include larger sample sizes and examine hard clinical endpoints (e.g., mortality, time to definitive hemorrhage control, intensive care unit and hospital length of stay, functional outcomes), as well as economic analyses. Qualitative research exploring staff perceptions, barriers to implementation, and factors influencing adherence to the pathway would also inform implementation strategies and scalability.

5. Conclusions

In summary, ENP implementation was associated with a statistically significant reduction in EDLOS in our single-center cohort. These observational results are encouraging but do not prove causation. Prospective, ideally multicenter studies and external validation of predictive models are required to confirm the effect of ENP on clinical outcomes and to determine the generalizability and utility of ML-based tools.

Acknowledgments

The authors thank all participants and all investigators.

Author contributions

Conceptualization: Yanfang Fu, Zhongqi Chen, Kaiyu Han, Yunpeng Wang.

Data curation: Yanfang Fu, Zhongqi Chen, Kaiyu Han.

Formal analysis: Yanfang Fu, Zhongqi Chen, Kaiyu Han, Yunpeng Wang.

Funding acquisition: Yanfang Fu, Zhongqi Chen, Yunpeng Wang.

Investigation: Yanfang Fu, Zhongqi Chen, Yunpeng Wang.

Methodology: Yanfang Fu, Mingfang Wu, Yunpeng Wang.

Project administration: Yanfang Fu, Mingfang Wu, Kaiyu Han, Yunpeng Wang.

Resources: Yanfang Fu, Mingfang Wu, Kaiyu Han, Yunpeng Wang.

Software: Yanfang Fu, Mingfang Wu, Kaiyu Han, Yunpeng Wang.

Supervision: Yanfang Fu, Mingfang Wu, Kaiyu Han, Yunpeng Wang.

Validation: Yanfang Fu, Mingfang Wu, Yunpeng Wang.

Visualization: Yanfang Fu, Mingfang Wu, Yunpeng Wang.

Writing – original draft: Yanfang Fu, Yunpeng Wang.

Writing – review & editing: Yanfang Fu, Yunpeng Wang.