Development and Temporal Validation of a Logistic Regression–Based Risk Score for Identifying Life-Threatening Chest Pain at Emergency Department Triage: A Retrospective Cohort Study

Abstract

Background

Chest pain is a frequent emergency department (ED) presentation with highly time-dependent outcomes. Widely used risk tools such as HEART and GRACE incorporate laboratory results that are unavailable at first nurse contact, limiting their value for immediate triage. In mainland China, acuity assignment follows the Chinese Emergency Triage Scale (CETS), yet nurse triage accuracy remains variable, particularly among less experienced staff.

Objective

To develop and temporally validate a simplified, non-laboratory clinical risk score that identifies life-threatening conditions (LTCs)—defined as final physician-assigned CETS Level I or II—among adults presenting to the ED with chest pain, and to map score thresholds onto the four-level CETS framework for nurse-facing decision support.

Methods

This was a single-centre retrospective observational cohort study, reported in accordance with TRIPOD+AI, at an ED of a large tertiary academic hospital in Hefei, Anhui Province, China. The cohort was split by calendar time into an internal development cohort (n = 2,005; July 2020–February 2021) and a temporal validation cohort (n = 501; February–March 2021). The internal cohort was randomly divided into a training set (n = 1,391) and a test set (n = 614). Candidate predictors were restricted to information routinely available at nurse triage. Least absolute shrinkage and selection operator (LASSO) regression identified parsimonious predictors, which were entered into a multivariable logistic regression (LR) model. Six machine-learning (ML) algorithms were trained as benchmarks. Discrimination (AUC, AUPRC), calibration (calibration curves, Brier score), clinical utility (decision-curve analysis, standardised net benefit), and DeLong tests for AUC differences were evaluated. An integer risk score was derived from LR coefficients. Feature contributions were examined with SHAP values based on the LR model.

Results

LTC prevalence was 44.8% (899/2,005) in the internal cohort and ED mortality was 1.7% (34/2,005). LASSO retained 13 predictors: age, sex, mode of arrival, symptom onset time, cardiogenic chest pain (CCP), chest tightness, hypertension, diabetes, coronary artery disease (CAD), chronic kidney disease (CKD), cerebrovascular disease (CVD), shock index (SI), and Modified Early Warning Score (MEWS). The LR model showed stable discrimination—AUC 0.864 (95% CI 0.844–0.883) in training, 0.876 (0.849–0.904) in the test set, and 0.860 (0.825–0.896) in temporal validation—without statistically significant AUC superiority of more complex ML models on the test or validation sets (DeLong P > 0.05). In the test set, LR achieved the highest AUPRC (0.858), F1 score (0.792), and standardised net benefit (0.690), with the lowest Brier score (0.142). The integer score ranged from −3 to 49. At the Youden-optimal cut-off of 18.8, four CETS-aligned tiers were defined: Level IV (score < 6.4), Level III (6.4–18.8), Level II (18.8–32.1), and Level I (score > 32.1). SHAP analysis of the LR model ranked transportation, CCP, MEWS, SI, sex, and age among the strongest contributors.

Conclusion

An internally and temporally validated, laboratory-free LR-based integer score showed discrimination and calibration comparable to more complex ML models, with clearer interpretability for ED nursing practice. The score is intended to support—not replace—clinical judgement. Multicentre external validation and prospective implementation studies are required before routine clinical adoption.

Introduction

Chest pain accounts for a substantial share of emergency department (ED) attendances worldwide and spans aetiologies from benign musculoskeletal pain to immediately life-threatening cardiovascular emergencies.¹ For acute myocardial infarction (AMI), acute aortic syndromes, and high-risk pulmonary embolism, delays in recognition and treatment are directly linked to excess morbidity and mortality.^2,3 These vascular and cardiopulmonary emergencies therefore demand reliable identification at the earliest point of ED contact, typically nurse-led triage.²

Despite this clinical urgency, triage accuracy for high-risk chest pain remains imperfect. Studies of nurse triage for myocardial infarction report accuracy in the range of roughly one-half to four-fifths of cases, with both under-triage of true emergencies and systematic over-triage driven by defensive practice.^4–6 Under-recognition is more likely when symptoms are atypical, particularly in older adults and women,^5,7,8 and when triage depends heavily on individual experience rather than structured decision support.^4,9 Over-triage rates exceeding 70% have been described in chest-pain pathways designed to avoid missed diagnoses, contributing to crowding, resource strain, and delayed care for other critically ill patients.² In mainland China, unrestricted ED access policies further amplify volume and acuity mix, increasing the premium on rapid, consistent triage.^10,11

In Chinese EDs, acuity is assigned using the Chinese Emergency Triage Scale (CETS), a four-level system (I, Resuscitation; II, Critical; III, Urgent; IV, Non-urgent) based on chief complaint, vital signs, and clinical assessment.¹² Multicentre work has shown that CETS can be applied with high inter-rater reliability,¹¹ yet performance still depends on nurse experience, shift context, and the availability of objective aids.^12–14 CETS therefore provides a nationally familiar scaffold for decision support, but does not itself supply a chest-pain–specific, quantitative risk estimate at first contact.

Most influential chest-pain scores were developed for physician assessment after initial work-up. The HEART score discriminates short-term major adverse cardiac events with AUCs commonly around 0.82–0.86, but requires troponin.¹⁵ GRACE,¹⁶ TIMI¹⁷ and related scores likewise depend on laboratory and ECG data unavailable during pure nurse triage. EDACS and accelerated diagnostic protocols improve ED throughput once biomarkers are known, yet remain downstream of first-contact prioritisation.¹⁸ To address this limitation, several researchers have attempted to develop triage-level prediction tools. HEAR (HEART without troponin) can support very-low-risk identification in some ED settings, but performance is cut-off–dependent and prehospital generalisability is limited.¹⁹ Yuwares²⁰ proposed an online formula to predict acute coronary syndrome (ACS) risk using readily available parameters without cardiac biomarkers; however, this tool requires external validation and has not been widely adopted. The Sanchis Risk Score,²¹ which incorporates clinical characteristics and medical history without laboratory tests, demonstrated superior predictive performance for major adverse cardiac events compared to the TIMI score at 14 days (C-statistic 0.78 vs 0.66, P = 0.0002). More recently, prehospital and triage-focused models have emerged: Wibring²² developed a prehospital risk score using ambulance data, while Goto et al²³ applied Machine-learning (ML) approaches to predict critical illness in a sample of 52, 037 children aged 18 years or younger in ED triage, yet gains over well-specified logistic regression (LR) are often modest. Tools designed closer to first contact, including non-laboratory clinical scores and prehospital models, show promise but have not been purpose-built for CETS-aligned nurse triage in Chinese adults.

Several gaps therefore remain. First, few models are calibrated specifically to Chinese ED populations, in whom emergency medical service (EMS) utilisation, comorbidity patterns, and care-seeking behaviour differ from many Western cohorts. For example, in published ST-segment elevation myocardial infarction (STEMI) cohorts, ambulance use has been reported to be substantially lower in China than in England and Wales (approximately 14.1%²⁴ vs 81.4% of ED arrivals²⁵). Mode of arrival, ambulance versus private transport, is a simple prehospital signal that may carry prognostic information, but its role in Chinese chest-pain triage is incompletely characterised and must be interpreted cautiously as a possible proxy for perceived severity rather than a purely causal factor. Second, no widely adopted tool maps a laboratory-free risk estimate directly onto CETS Levels I–IV for chest pain. Third, although ML is frequently promoted for triage prediction, whether it outperforms transparent regression for this specific task, and whether any incremental gain justifies reduced interpretability, has not been rigorously tested with calibration, decision-curve, and formal AUC comparison metrics recommended by contemporary prediction-model guidance.²³

Accordingly, this study aimed to: (1) identify independent, triage-available predictors of LTCs among adults with chest pain; (2) develop logistic regression and benchmark ML models using only non-laboratory first-contact data; (3) derive a simple integer risk score from the best-balanced model; and (4) propose CETS-aligned thresholds to support nurse triage. We predefined LTC as final physician-assigned CETS Level I/II after complete ED evaluation, thereby targeting a clinically meaningful acuity endpoint while acknowledging that it is not identical to hard outcomes such as confirmed infarction or death. The intended product is decision support for nursing practice, not replacement of clinical judgement.

Methods

Study Design and Setting

We conducted a retrospective observational cohort study in the ED of The First Affiliated Hospital of USTC, a tertiary academic medical centre in Hefei, Anhui Province, China. The study is reported in accordance with the TRIPOD+AI statement for prediction model development and validation.²⁶ Although predictors and initial triage impressions were recorded at a single encounter, patients were followed through the ED episode to ascertain the final acuity outcome; the design is therefore best classified as a retrospective cohort rather than a purely cross-sectional survey.

Participants

We screened all adult ED attendances with a primary complaint of chest pain or chest-pain equivalents between July 2020 and March 2021.

Inclusion criteria: age ≥ 18 years; chief complaint of chest pain or equivalents (pressure, tightness, or discomfort from the chin to the upper abdomen, or dyspnoea/fatigue judged potentially cardiac at triage), consistent with contemporary chest-pain evaluation guidance.²⁷

Exclusion criteria: (1) pregnancy; (2) death on arrival; (3) advanced malignancy with documented do-not-resuscitate status; (4) bypass of triage (eg, direct specialty admission); (5) left without being seen; and (6) incomplete core physiological records at triage, defined a priori as missing ≥ 3 of blood pressure, heart rate, respiratory rate, oxygen saturation, or temperature.

Of 6,189 registered cases, 3,683 (59.5%) were excluded: incomplete physiological records (n = 2,582; 41.7%), triage bypass (n = 882; 14.2%), left before evaluation (n = 207; 3.3%), and prehospital death (n = 12; 0.2%). The final analytic cohort comprised 2,506 patients (Figure 1). The high exclusion fraction, driven largely by missing vitals, raises selection-bias concerns that are addressed in the Limitations and were mitigated where possible by multiple imputation for variables with lower-level missingness (Missing Data).

Figure 1.

Flow diagram of participant selection.

Data Collection

A standardized data extraction form was developed based on literature review and expert consensus among emergency medicine physicians and triage nurses. All triage nurses received standardized training on CETS application and data recording protocols prior to study initiation. Data were retrospectively extracted from the hospital’s electronic medical record system, including emergency triage records, nursing documentation, physician notes, and laboratory/imaging reports.

Demographics and context: age; sex; presentation period (day 08:00–20:00 vs night 20:00–08:00); symptom onset time (≤ 6 h vs > 6 h before arrival); mode of transportation (private vehicle vs “120” ambulance via the national EMS system).

Clinical presentation at triage: mutually exclusive nurse impression categories adapted from chest-pain guidelines:²⁷ (1) suspected cardiogenic chest pain (CCP):²⁷ typical substernal or left-sided chest pain with radiation to the left arm, jaw, or back, associated with exertion, diaphoresis, or dyspnea; (2) suspected non-cardiogenic chest pain (NCCP):^27,28 clearly associated with pleuritic, musculoskeletal, gastrointestinal, or psychogenic features; and (3) nonspecific chest pain:^1,27,29 chest discomfort not clearly fitting CCP or NCCP criteria, including chest tightness, palpitations, dizziness, presyncope, or general malaise.

Comorbidities (record-documented or patient-reported): hypertension (HTN), diabetes, coronary artery disease (CAD), chronic kidney disease (CKD), chronic obstructive pulmonary disease or other chronic lung disease, cerebrovascular disease (CVD), prior percutaneous coronary intervention (PCI), active malignancy, chronic liver disease, arrhythmias, connective tissue disease (CTD), and organic non-ischemic heart disease.

Physiological parameters: systolic and diastolic blood pressure, heart rate, respiratory rate, oxygen saturation (SpO₂), temperature, and mental status (Alert, Voice, Pain, Unresponsive—AVPU scale) measured by certified triage nurses with standardized equipment.

From these, two composite scores were calculated: (1) Modified Early Warning Score (MEWS),³⁰ ranging from 0 to 14, with higher scores indicating greater physiological instability; and (2) Shock Index (SI), calculated as heart rate divided by systolic blood pressure, with values ranged from 0.5 to 0.7 considered normal.³¹

Outcome Definition

The primary outcome was life-threatening condition (LTC), defined as the final CETS Level I or II assigned by the treating emergency physician at ED disposition, after diagnostic evaluation, clinical course, interventions, and disposition (ICU, ward, or discharge) were known.^2,12,22,23

CETS¹² levels were operationalised as: Level I (Resuscitation): immediate threat to life requiring resuscitation; Level II (Critical): potential threat requiring urgent intervention; Level III (Urgent): urgent but stable, evaluation within ~30 minutes; Level IV (Non-urgent): suitable for delayed assessment or outpatient pathway.

Reference standard and independence from predictors. Final CETS level was physician-assigned at disposition and is distinct from the initial nurse triage category. Predictors were those recorded at first nurse contact. We did not use initial nurse triage level as a model predictor. Nonetheless, because some physiological predictors (eg, MEWS components, SpO₂, SI) can also inform human acuity judgements, residual incorporation bias cannot be excluded; this limitation is discussed explicitly. LTC is an acuity/resource-need construct, not a composite of hard endpoints (eg, angiographically confirmed infarction, 30-day mortality). ED mortality is reported descriptively to contextualise label severity.

Missing Data

Variables with < 30% missingness were imputed using multiple imputation by chained equations (MICE; m = 5). Patients missing ≥ 3 core vital signs were excluded under the a priori rule above rather than imputed, because such records were judged clinically incomplete for triage modelling. Imputed datasets were analysed and results combined using Rubin’s rules where applicable. The variables imputed, missingness fractions, and exclusion threshold are reported to address transparency concerns regarding the coexistence of exclusion and imputation.

Sample Splitting and Sample Size

To reduce optimistic bias from a single random split of the entire cohort, we used a two-stage scheme: (1) Temporal validation cohort (n = 501): consecutive attendances from February to March 2021; (2) Internal development cohort (n = 2,005): July 2020–February 2021, randomly split into training (n = 1,391; model development and LASSO selection) and test (n = 614; hyperparameter monitoring and internal performance).

All splits used identical eligibility criteria. With 615 LTC events in the training set, the events-per-candidate-predictor ratio substantially exceeded conventional rules of thumb for stable logistic modelling of a binary endpoint.³²

Statistical Analysis

Continuous variables are summarised as mean ± standard deviation if approximately normal, or median (interquartile range) otherwise; categorical variables as counts and percentages. Between-group comparisons used Student’s t-test or Mann–Whitney U-test for continuous variables and χ² or Fisher’s exact test for categorical variables, the latter when expected cell counts were sparse.

Predictor selection. On the training set, LASSO logistic regression with 10-fold cross-validation selected the penalty by the one-standard-error rule (λ_1se) to favour parsimony. Variables with non-zero coefficients at λ_1se entered multivariable LR.

Benchmark models. Using the same LASSO-selected predictors and training data, we trained six machine-learning algorithms for comparison with logistic regression: RandomForest, gradient boosting machine (GBM), LightGBM, multilayer perceptron (MLP), support vector machine with a radial basis function kernel (SVM-RBF), and naïve Bayes. Logistic regression was estimated by maximum likelihood without additional penalisation after LASSO screening. For RandomForest, MLP, and naïve Bayes, key settings were fixed a priori to limit overfitting. For SVM-RBF, cost and gamma were selected by exhaustive grid search with 5-fold cross-validation. For GBM, the number of boosting iterations was selected by internal 5-fold cross-validation. For LightGBM, the number of boosting rounds was determined by early stopping on the internal test set using the AUC metric. The temporal validation cohort was not used for training or hyperparameter selection. Full search spaces, final hyperparameter values, software packages, and random-seed controls are reported in Supplementary Table S1. Comparative performance claims emphasise the internal test set and the temporal validation cohort.

Performance metrics. On training, test, and temporal validation sets we assessed: Discrimination: AUC and area under the precision–recall curve (AUPRC) with 95% confidence intervals; pairwise AUC differences versus LR tested with DeLong’s method; Calibration: calibration plots and Brier score; Clinical utility: decision-curve analysis (DCA) and standardised net benefit (SNB) across threshold probabilities; Classification metrics: F1 score at model-appropriate operating points.

Score development. Because LR offered stable performance, full interpretability, and bedside feasibility, β coefficients were converted to integer points by dividing each coefficient by the smallest absolute coefficient retained for scoring and rounding to the nearest integer. The Youden index on the training ROC defined the primary LTC cut-off. Additional thresholds targeting approximately 1% miss rate (high sensitivity) and 1% false-positive rate among screen-positives at the extreme high end (high specificity) were used to construct four CETS-aligned tiers.

Explainability. The Shapley Additive Explanations (SHAP) values were computed for the LR model to visualise direction and magnitude of feature contributions, ensuring importance narratives matched the recommended model. A higher SHAP value indicates that the feature has a greater positive impact on LTCs, while a lower value indicates that the feature reduces the likelihood of LTCs.

Analyses used SPSS 26.0 and R 4.4.2. Two-sided P < 0.05 was considered statistically significant.

Ethical Considerations

See Declarations. All analyses used anonymised data stored on access-controlled institutional systems.

Results

Cohort Flow and Baseline Characteristics

Participant flow is shown in Figure 1. Among 2,005 internal-cohort patients, 895 (44.6%) were aged > 65 years, 1,152 (57.5%) were male, 689 (34.4%) arrived by 120 ambulance, 899 (44.8%) met the LTC definition, and 34 (1.7%) died during the ED stay. Training (n = 1,391) and test (n = 614) sets were well balanced on demographics, arrival mode, comorbidities, vital signs, MEWS, SI, LTC prevalence (44.2% vs 46.3%, P = 0.425), and mortality (Table 1).

Table 1.

Baseline Characteristics of the Internal Development Cohort (Training vs Test)

Variables	Level	Total (n=2, 005)	Training (n=1, 391)	Test (n=614)	P Value
Age	≤65	1110 (55.4)	771 (55.4)	339 (55.2)	0.967
	>65	895 (44.6)	620 (44.6)	275 (44.8)
Sex	Male	1152 (57.5)	799 (57.4)	353 (57.5)	0.999
	Female	853 (42.5)	592 (42.6)	261 (42.5)
Presentation time	20:00pm-8:00am	802 (40.0)	553 (39.8)	249 (40.6)	0.774
	8:00am-20:00pm	1203 (60.0)	838 (60.2)	365 (59.4)
Transportation	Personal vehicles	1316 (65.6)	915 (65.8)	401 (65.3)	0.878
	120 ambulance	689 (34.4)	476 (34.2)	213 (34.7)
Onset time, n (%)	≤6	1343 (67.0)	943 (67.8)	400 (65.1)	0.267
	>6	662 (33.0)	448 (32.2)	214 (34.9)
Complaints
CCP	No	1423 (71.0)	980 (70.5)	443 (72.1)	0.473
	Yes	582 (29.0)	411 (29.5)	171 (27.9)
NCCP	No	1960 (97.8)	1367 (98.3)	593 (96.6)	0.058
	Yes	45 (2.2)	24 (1.7)	21 (3.4)
Chest tightness	No	1433 (71.5)	1019 (73.3)	414 (67.4)	0.059
	Yes	572 (28.5)	372 (26.7)	200 (32.6)
Palpitations	No	1713 (85.4)	1186 (85.3)	527 (85.8)	0.792
	Yes	292 (14.6)	205 (14.7)	87 (14.2)
Syncope	No	1936 (96.6)	1341 (96.4)	595 (96.9)	0.665
	Yes	69 (3.4)	50 (3.6)	19 (3.1)
Vertigo	No	1974 (98.5)	1368 (98.3)	606 (98.7)	0.696
	Yes	31 (1.5)	23 (1.7)	8 (1.3)
Malaise	No	1993 (99.4)	1382 (99.4)	611 (99.5)	0.913
	Yes	12 (0.6)	9 (0.6)	3 (0.5)
Comorbidities
HTN	No	1178 (58.8)	817 (58.7)	361 (58.8)	0.999
	Yes	827 (41.2)	574 (41.3)	253 (41.2)
Diabetes	No	1748 (87.2)	1215 (87.3)	533 (86.8)	0.794
	Yes	257 (12.8)	176 (12.7)	81 (13.2)
CAD	No	1562 (77.9)	1084 (77.9)	478 (77.9)	0.999
	Yes	443 (22.1)	307 (22.1)	136 (22.1)
CKD	No	1908 (95.2)	1334 (95.9)	574 (93.5)	0.057
	Yes	97 (4.8)	57 (4.1)	40 (6.5)
Chronic lung disease	No	1894 (94.5)	1323 (95.1)	571 (93.0)	0.071
	Yes	111 (5.5)	68 (4.9)	43 (7.0)
CVD	No	1864 (93.0)	1296 (93.2)	568 (92.5)	0.660
	Yes	141 (7.0)	95 (6.8)	46 (7.5)
PCI	No	1852 (92.4)	1285 (92.4)	567 (92.3)	0.999
	Yes	153 (7.6)	106 (7.6)	47 (7.7)
Tumour	No	1952 (97.4)	1351 (97.1)	601 (97.9)	0.410
	Yes	53 (2.6)	40 (2.9)	13 (2.1)
Chronic liver disease	No	1995 (99.5)	1382 (99.4)	613 (99.8)	0.283
	Yes	10 (0.5)	9 (0.6)	1 (0.2)
Arrhythmias	No	1943 (96.9)	1344 (96.6)	599 (97.6)	0.329
	Yes	62 (3.1)	47 (3.4)	15 (2.4)
CTD	No	1991 (99.3)	1380 (99.2)	611 (99.5)	0.647
	Yes	14 (0.7)	11 (0.8)	3 (0.5)
Organic non-ischemic heart disease	No	1979 (98.7)	1376 (98.9)	603 (98.2)	0.277
	Yes	26 (1.3)	15 (1.1)	11 (1.8)
LTCs	No	1106 (55.2)	776 (55.8)	330 (53.7)	0.425
	Yes	899 (44.8)	615 (44.2)	284 (46.3)
Temperature (°C)	Mean ± SD	36.39 ± 0.29	36.39 ± 0.29	36.40 ± 0.28	0.325
Heart rate (/min)	Mean ± SD	88.72 ± 29.90	88.90 ± 30.16	88.32 ± 29.32	0.685
Respiratory (/min)	Mean ± SD	19.93± 2.92	19.89 ± 2.93	20.03 ± 2.91	0.311
Systolic pressure (mmHg)	Mean ± SD	143.01± 34.22	142.98 ± 34.63	143.09 ± 33.32	0.943
Diastolic pressure (mmHg)	Mean ± SD	83.58± 20.42	83.58 ± 20.78	83.57 ± 19.61	0.987
Saturation (%)	Mean ± SD	95.51± 11.02	95.61 ± 11.02	95.28 ± 11.04	0.533
Mental status	Alert	1968 (98.2)	1368 (98.3)	600 (97.7)	0.243
	Voice	7 (0.3)	3 (0.2)	4 (0.7)
	Pain	3 (0.1)	1 (0.1)	2 (0.3)
	Unresponsive	27 (1.3)	19 (1.4)	8 (1.3)
MEWS	Median (IQR)	1.50 (1.00, 2.00)	1.50 (1.00, 2.00)	1.46 (1.00, 2.00)	0.427
SI	Median (IQR)	0.65 (0.48, 0.73)	0.65 (0.48, 0.73)	0.64 (0.47, 0.72)	0.599
ED mortality	No	1971 (98.3)	1365 (98.1)	606 (98.7)	0.455
	Yes	34 (1.7)	26 (1.9)	8(1.3)

Notes: Values are n (%) unless stated.

Abbreviations: CCP, cardiogenic chest pain; HTN, hypertension; CAD, coronary artery disease; CKD, chronic kidney disease; CVD, cerebrovascular disease; LTC, life-threatening condition; MEWS, Modified Early Warning Score; SI, shock index.

Feature Selection

Candidate predictors entered LASSO on the training set (Figure 2A). At λ_1se = 0.005251 (10-fold cross-validation), 13 variables with non-zero coefficients were retained: age, sex, transportation, onset time, CCP, chest tightness, hypertension, diabetes, CAD, CKD, CVD, SI, and MEWS (Figure 2B).

Figure 2.

Cross-validation curve and coefficient shrinkage path of variables selected by LASSO regression. (A) Ten-fold cross-validation curve; the shaded band denotes ±1 standard error. (B) Coefficient shrinkage path of the candidate variables. Vertical dashed lines indicate lambda.min and lambda.1se. The upper axis shows the number of non-zero coefficients.

Model Performance and Benchmarking

Seven models were compared (Figure 3 and Table 2). On the training set, LightGBM achieved the highest AUC (0.909, 95% CI: 0.893–0.924) and RandomForest (0.904, 95% CI: 0.888–0.920) also fit strongly, consistent with flexible learners capturing training-set structure (Figure 3A). On the test set, AUCs converged: LR 0.876 (0.849–0.904), GBM 0.877 (0.850–0.904), SVM-RBF 0.876 (0.848–0.903), LightGBM 0.874 (0.847–0.901), and RandomForest 0.874 (0.847–0.901) (Figure 3B). On validation, GBM led numerically (AUC 0.878, 0.846–0.910), with LR at 0.860 (0.825–0.896) (Figure 3C). DeLong tests showed no statistically significant AUC differences between LR and competing ML models on the test or temporal validation sets (P > 0.05).

Figure 3.

The performance and comparison of 7 different predictive models. (A) ROC Curves for the Training Set; (B) ROC Curves for the Test Set; (C) ROC Curves for the Validation Set; (D) Calibration Curve for the Test Set; (E) Calibration Curve for the Validation Set; (F) Decision Curve Analysis for the Test Set; (G) Decision Curve Analysis for the Validation Set; (H) Precision-Recall Curves for the Test Set; (I) Precision-Recall Curves for the Validation Set.

Abbreviations: LR, Logistic Regression; GBM, Gradient Boosting Machine; MLP, Multilayer Perceptron, SVM, Support Vector Machine.

Table 2.

Detailed Performance Metrics of Various Models for Predicting Life-Threatening Conditions Risk in Patients with Chest Pain Across Training, Test, and Validation Sets

Dataset	Model	AUC (95% CI)	AUPRC (95% CI)	Brier (95% CI)	F1 (95% CI)	SNB (95% CI)
Training	Logistic	0.864 (0.844–0.883)	0.837 (0.805–0.862)	0.148 (0.137–0.158)	0.774 (0.748–0.797)	0.677 (0.645–0.706)
	RandomForest	0.904 (0.888–0.920)	0.895 (0.875–0.912)	0.124 (0.113–0.136)	0.815 (0.789–0.837)	0.727 (0.696–0.755)
	SVM_RBF	0.865 (0.846–0.884)	0.839 (0.808–0.864)	0.146 (0.135–0.158)	0.783 (0.758–0.805)	0.680 (0.647–0.709)
	NaiveBayes	0.829 (0.808–0.850)	0.788 (0.756–0.820)	0.194 (0.178–0.210)	0.742 (0.715–0.767)	0.611 (0.574–0.648)
	MLP	0.890 (0.873–0.908)	0.865 (0.837–0.888)	0.130 (0.119–0.140)	0.804 (0.780–0.827)	0.713 (0.685–0.741)
	GBM	0.883 (0.866–0.901)	0.870 (0.847–0.890)	0.140 (0.132–0.150)	0.793 (0.767–0.818)	0.690 (0.658–0.717)
	LightGBM	0.909 (0.893–0.924)	0.897 (0.877–0.914)	0.123 (0.114–0.132)	0.816 (0.793–0.838)	0.730 (0.701–0.757)
Test	Logistic	0.876 (0.849–0.904)	0.858 (0.817–0.892)	0.142 (0.128–0.158)	0.792 (0.756–0.825)	0.690 (0.644–0.732)
	RandomForest	0.874 (0.847–0.901)	0.847 (0.805–0.888)	0.147 (0.132–0.166)	0.764 (0.720–0.805)	0.678 (0.629–0.722)
	SVM_RBF	0.876 (0.848–0.903)	0.851 (0.803–0.893)	0.142 (0.129–0.158)	0.764 (0.722–0.800)	0.682 (0.637–0.726)
	NaiveBayes	0.861 (0.832–0.890)	0.826 (0.783–0.871)	0.172 (0.147–0.196)	0.773 (0.734–0.809)	0.632 (0.578–0.687)
	MLP	0.869 (0.841–0.897)	0.846 (0.805–0.884)	0.146 (0.130–0.162)	0.769 (0.728–0.808)	0.684 (0.637–0.727)
	GBM	0.877 (0.850–0.904)	0.841 (0.795–0.890)	0.147 (0.134–0.161)	0.784 (0.747–0.822)	0.686 (0.640–0.730)
	LightGBM	0.874 (0.847–0.901)	0.839 (0.791–0.886)	0.145 (0.131–0.160)	0.768 (0.727–0.808)	0.683 (0.638–0.727)
Validation	Logistic	0.860 (0.825–0.896)	0.770 (0.707–0.834)	0.139 (0.121–0.157)	0.733 (0.683–0.784)	0.577 (0.498–0.649)
	RandomForest	0.873 (0.841–0.906)	0.787 (0.729–0.847)	0.132 (0.112–0.149)	0.731 (0.677–0.777)	0.600 (0.528–0.659)
	SVM_RBF	0.866 (0.831–0.901)	0.789 (0.733–0.844)	0.135 (0.117–0.154)	0.734 (0.680–0.782)	0.588 (0.512–0.654)
	NaiveBayes	0.850 (0.815–0.885)	0.743 (0.672–0.807)	0.166 (0.139–0.195)	0.703 (0.654–0.754)	0.551 (0.479–0.626)
	MLP	0.858 (0.823–0.892)	0.756 (0.687–0.821)	0.143 (0.124–0.163)	0.732 (0.679–0.778)	0.574 (0.494–0.642)
	GBM	0.878 (0.846–0.910)	0.803 (0.746–0.856)	0.134 (0.119–0.148)	0.735 (0.685–0.785)	0.584 (0.505–0.647)
	LightGBM	0.873 (0.842–0.905)	0.799 (0.744–0.850)	0.136 (0.120–0.152)	0.709 (0.656–0.757)	0.578 (0.497–0.643)

Notes: DeLong tests: LR vs other models, test and validation, all P > 0.05.

Abbreviations: LR, Logistic Regression; GBM, Gradient Boosting Machine; MLP, Multilayer Perceptron, SVM, Support Vector Machine, AUC, area under the curve; AUPRC, area under the precision–recall curve; SNB, standardised net benefit.

Beyond AUC, LR was competitive or superior on clinically salient metrics: highest test-set F1 (0.792), and SNB (0.690), and lowest test-set Brier score (0.142) (Table 2). Calibration plots demonstrated close agreement between predicted and observed LTC risk for LR on test and validation data (Figure 3D and E). DCA indicated positive net benefit across a wide range of threshold probabilities (Figure 3F and G). On the PR curves, LR exhibits strong performance: highest AUPRC values of 0.858 (95% CI 0.817–0.892) on the test set (Figure 3H), and 0.770 (95% CI 0.707–0.834) on the validation set (Figure 3I), indicating exceptional predictive accuracy for positive samples. Collectively, complex ML did not deliver a statistically or clinically decisive advantage over LR for this triage task, while exhibiting larger train-to-test attenuation suggestive of overfitting.

Predictor Contributions (LR-Aligned)

Mean SHAP ranked transportation (0.676), CCP (0.662), MEWS (0.449), SI (0.292), sex (0.250), and age (0.199) among the leading contributors (Figure 4A and B), aligned with LR effect sizes rather than discarded tree-ensemble importances (Figure 4C). Individual prediction explanations for the LR model showed higher LTC risk associated with 120 arrival, CCP, higher MEWS, more extreme SI, male sex, age ≥ 65 years, CAD, hypertension, diabetes, chest tightness, CKD, CVD, and longer symptom onset category; private arrival, female sex, and younger age (< 65) reduced predicted risk (Figure 4D).

Figure 4.

SHAP-based interpretation of the logistic regression model. (A) SHAP summary plot; each point is one patient, coloured by feature value. (B) Global feature importance ranked by mean absolute SHAP value. (C) Odds ratios with 95% confidence intervals; red indicates P < 0.05. (D) Individual prediction explanations for one high-risk and one low-risk patient.

Abbreviations: CCP, cardiac chest pain; MEWS, Modified Early Warning Score; SI, Shock Index; HTN, hypertension; CAD, coronary artery disease; CVD, cerebrovascular disease; Coef, Coefficient; SE, Standard Error; OR, Odds Ratio; CI, Confidence Interval.

Score Construction and Triage Mapping

Integer points derived from LR coefficients (score = round (β/β_min), β_min = 0.147814) are shown in Table 3. Total scores range from −3 to 49 (Table 3). On the training ROC, the Youden-optimal threshold was 18.8 (Youden index 0.587). A score of 6.4 achieved approximately 99% sensitivity (≈ 1% miss rate among LTCs), whereas 32.1 achieved approximately 99% specificity at the high extreme. These anchors defined four operational tiers aligned to CETS (Table 4).

Table 3.

LASSO and Logistic Regression Results for LTCs Prediction

Variable	LASSO Coef (λ.1se)	Logistic Coef	SE	OR (95% CI)	P Value	Points
CCP	1.8578	2.1564	0.1807	8.640 (6.092–12.378)	<0.001	10
Transportation	1.4434	1.5610	0.1481	4.764 (3.573–6.388)	<0.001	8
MEWS	0.8551	0.5435	0.0627	1.722 (1.529–1.954)	<0.001	3
SI	0.4356	1.7878	0.2843	5.976 (3.453–10.537)	<0.001	9
Sex	−0.5783	−0.6865	0.1444	0.503 (0.379–0.667)	<0.001	−3
Age	0.3144	0.5213	0.1564	1.684 (1.241–2.291)	<0.001	3
CKD	0.5000	0.8849	0.3189	2.423 (1.303–4.565)	<0.001	4
CAD	0.4056	0.5596	0.1822	1.750 (1.226–2.505)	<0.001	3
HTN	0.3451	0.4584	0.1527	1.582 (1.173–2.135)	<0.001	2
Diabetes	0.3603	0.4409	0.2103	1.554 (1.031–2.353)	0.005	2
CVD	0.1058	0.5280	0.2586	1.696 (1.024–2.826)	0.012	3
Chest tightness	0.0679	0.2750	0.1779	1.317 (0.928–1.866)	0.051	1
Onset time	0.0284	0.2056	0.1504	1.228 (0.915–1.650)	0.248	1
Intercept	–	−3.2398	0.3312	–	<0.001

Notes: LASSO regression: 10-fold cross-validation, Lambda.1se = 0.005251. Scoring Method: Variable Score = β/β_min (β_min = 0.147814); Female sex contributes −3 points relative to male sex.

Abbreviations: CCP, cardiac chest pain; MEWS, Modified Early Warning Score; SI, Shock Index; HTN, hypertension; CAD, coronary artery disease; CVD, cerebrovascular disease; Coef, Coefficient; SE, Standard Error; OR, Odds Ratio; CI, Confidence Interval.

Table 4.

Triage Mapping Aligned to CETS

CETS Level	Clinical Label	Score Range	Operational Intent
IV	Non-urgent	< 6.4	Very low predicted LTC risk; delayed pathway with continuous reassessment
III	Urgent	6.4–18.8	Intermediate risk; timely physician evaluation;
II	Critical	18.8–32.1	High risk; expedited assessment and monitoring
I	Resuscitation	> 32.1	Very high risk; immediate resuscitation resources

Notes: Thresholds are decision-support guides derived from this cohort’s ROC operating points; they are not empirical measurements of door-to-electrocardiogram or treatment intervals in this study.

Discussion

Principal Findings

In this single-centre retrospective cohort of adults presenting to the ED with chest pain (Figure 1), we developed a laboratory-free integer risk score from multivariable logistic regression after LASSO selection (Figure 2A and B) and benchmarked it against six machine-learning algorithms (Figure 3A–I). Using only information available at nurse triage, the LR model discriminated final physician-assigned CETS Level I/II status—the predefined LTC endpoint—with AUCs of 0.864 (95% CI 0.844–0.883) in the training set, 0.876 (0.849–0.904) in the internal test set and 0.860 (0.825–0.896) in the temporal validation cohort. Calibration plots (Figure 3D and E) and a test-set Brier score of 0.142 indicated acceptable agreement between predicted and observed risk, decision-curve analysis showed positive SNB across a wide range of threshold probabilities (Figure 3F and G), and PR curves demonstrated strong precision-recall balance (Figure 3H and I). An integer score ranging from −3 to 49, with Youden-optimal and extreme-sensitivity/specificity anchors at 18.8, 6.4 and 32.1, was mapped onto four CETS-aligned tiers to support bedside triage nursing. Importantly, more flexible ML models did not achieve statistically significant AUC gains over LR on the test or temporal validation sets (DeLong P > 0.05), while several showed larger train-to-test attenuation consistent with overfitting.

What the Score Predicts, and What it Does Not

A recurring concern in triage prediction research is circularity: if the label is simply the initial nurse grade, models may reproduce human judgement rather than forecast clinical need.^2,13 We therefore defined LTC as final CETS Level I or II assigned by the treating emergency physician after complete ED evaluation, and we excluded the initial nurse triage category from the predictor set. This endpoint is clinically meaningful for nursing workflow because it is expressed in the same four-level language used nationally in China and has demonstrated reliability in multicentre assessment.¹² It answers a pragmatic question at first contact: which patients are likely to require Level I/II resources before laboratory results are available?

The construct must nonetheless be interpreted with restraint. In the internal cohort, LTC prevalence was 44.8%, whereas ED mortality was only 1.7% (Table 1). This discordance shows that “life-threatening” in the present study denotes high acuity and urgent resource need, not near-certain death. Similar gaps between triage-critical labels and short-term mortality are recognised when systems intentionally favour sensitivity to avoid missed time-critical disease.^2,12 In addition, physiological predictors that contribute to MEWS and SI may later influence physicians’ final acuity judgements, so residual incorporation bias cannot be excluded even though outcome adjudication was separated in time and professional role from first-contact data collection.^13,26 Hard endpoints, confirmed acute coronary syndrome, acute aortic syndrome, pulmonary embolism, major adverse cardiac events or 30-day mortality, were not co-primary outcomes and should be added in external validation studies.

Comparison with Existing Chest-Pain Risk Tools

Direct numerical comparison across studies is imperfect because endpoints, sampling frames and the timing of predictors differ. Orientation against the literature is still useful for nursing readers. The HEART score typically yields AUCs in the low-to-mid 0.80s for short-term major adverse cardiac events when electrocardiography and troponin are included.^15,33 GRACE and related tools perform well for prognosis in acute coronary syndromes but depend on laboratory and electrocardiographic inputs unavailable during pure nurse triage.¹⁶ EDACS-based accelerated diagnostic protocols improve disposition once high-sensitivity troponin is known, yet they operate downstream of first-contact prioritisation.¹⁸ HEAR can support identification of very-low-risk patients in some ED settings, but performance is cut-off dependent and prehospital generalisability is limited.¹⁹ Prehospital models using ambulance data provide complementary early risk estimates for undifferentiated chest pain, with discrimination that is often useful but not uniformly superior to simple clinical scores.^22,34

Against this landscape, AUCs of approximately 0.86–0.88 without laboratory tests, obtained at nurse first contact and expressed as CETS tiers, are competitive for the narrower task of early prioritisation rather than definitive rule-out of acute coronary syndrome. The contribution of the present work is therefore best framed as complementarity: a transparent front-end nursing score that can later hand off to biomarker-enabled pathways (eg, HEART or EDACS), not as a claim of universal superiority over those instruments.^18,19,33 Few published tools are purpose-built for CETS-aligned adult chest-pain triage in Chinese EDs, where EMS utilisation and care-seeking patterns differ from many high-income settings.^12,24

Why Logistic Regression Was Retained Over More Complex ML

A distinctive feature of this study is not the pursuit of maximal algorithmic complexity, but a deliberate, evidence-based choice of the simplest model that preserved predictive performance and nursing usability.

LightGBM and RandomForest achieved the highest training AUCs (0.909 and 0.904, respectively) (Figure 3A) but attenuated on the test set (both 0.874) (Figure 3B) and did not significantly outperform LR on temporal validation (both 0.873) (Figure 3C). In contrast, LR showed less attenuation and more coherent performance across stages (AUC 0.864, 0.876 and 0.860; F1 0.774, 0.792 and 0.733) (Table 2). DeLong tests detected no significant AUC differences between LR and any ML comparator on test or temporal validation (P > 0.05) (Table 2). For nurse triage, where case-mix, crowding and staffing vary continuously, out-of-sample stability is a more relevant success criterion than peak training AUC.^26,35

Clinical usefulness depends on net benefit and calibration, not AUC in isolation. LR achieved the highest test-set SNB (0.690) and the lowest Brier score (0.142), together with acceptable agreement in calibration curves (Figure 3D and E). On temporal validation, the SNB of LR (0.577) remained close to the best ML model (random forest 0.600), decision curves were nearly superimposed across triage-relevant thresholds (Figure 3F and G), and PR curves confirmed superior positive predictive performance (Figure 3H and I). Complex ML therefore conferred no clear decision-analytic increment.

For triage nursing, interpretability is an implementation requirement rather than an aesthetic preference. An integer score can be calculated rapidly, embedded in the electronic triage form with auditable weights, taught to novice nurses and reviewed in clinical governance meetings.^13,35 Applying Occam’s razor, when discrimination, calibration and net benefit are comparable, a parsimonious LR-based instrument is preferable for high-reliability ED environments. ML benchmarking remains valuable as a stress test; in this study it reassured rather than displaced LR.

Interpretation of Key Predictors

Predictor contributions are discussed using LR odds ratios and SHAP values for the recommended model (Figure 4A–D), not feature importances from discarded ensemble algorithms. Feature ranking by mean absolute SHAP values highlight transportation, CCP, and vital-sign indices as dominant risk drivers (Figure 4A and B), mirroring the multivariable odds ratios of the logistic regression model (Figure 4C). Furthermore, individual prediction explanations demonstrate how specific clinical profiles shift the predicted risk in individual patients (Figure 4D).

Suspected cardiogenic chest pain. CCP carried the largest regression weight (OR 8.640; 10 points) (Table 3) and the second-highest mean SHAP value (0.662) (Figure 4A–C). Recognition of typical ischaemic features at triage remains central to contemporary chest-pain guidelines and supports continued investment in nurse education on both classic and sex- or age-atypical presentations.^1,27

Transportation. In published STEMI cohorts, ambulance utilisation has been reported to be much lower in China than in England and Wales (approximately 14% vs 81%).^24,25 In the present unselected chest-pain cohort, 120 arrival accounted for 34.4% of patients in the internal cohort and was strongly associated with LTC (OR 4.764; 8 points; highest mean SHAP 0.676) (Figure 4A–C). Mode of arrival is best read as a risk marker and workflow cue, not as evidence that increasing ambulance use causally improves outcomes. Reverse causation and confounding by indication are plausible: patients or bystanders who perceive a catastrophic threat are more likely to activate EMS.^22,34 Policy recommendations regarding EMS uptake require interventional or quasi-experimental designs beyond the present observational analysis.

Shock index and MEWS. SI (OR 5.976 per modelled unit; 9 points) and MEWS (OR 1.722 per point; 3 points) (Table 3) quantify physiological instability already familiar to emergency nurses. SHAP dependence patterns suggested higher MEWS and more extreme SI were associated with a higher predicted risk (Figure 4D), consistent with both hypoperfusion and extreme physiological states. Their retention after LASSO implies incremental value beyond symptom labels alone, and supports integrating early-warning physiology into complaint-specific pathways rather than running disconnected parallel scores.^30,36

Comorbidities. Among comorbidities, CKD showed the largest adjusted odds ratio (OR 2.423; 4 points) (Table 3), aligning with the established cardiovascular hazard of reduced kidney function.^37,38 CAD (OR 1.750; 3 points), hypertension (OR 1.582; 2 points), diabetes (OR 1.554; 2 points) and CVD (OR 1.696; 3 points) contributed smaller but coherent increments (Table 3). Overall, CKD is influential on the odds scale among comorbidities, whereas transportation and CCP dominate overall prediction.

Demographics. Male sex and age >65 years increased predicted risk (Figure 4C and D), consistent with epidemiological patterns in European and multicentre chest-pain cohorts.^39,40 Lower average predicted risk among women must not encourage under-triage when symptoms are atypical; the score is an adjunct to, not a substitute for, sex-aware clinical assessment.^1,8,27

Nonspecific symptoms. Most nonspecific complaints did not dominate the final model, yet they should not be dismissed clinically. Large administrative studies show nontrivial readmission and cardiovascular risk after nonspecific chest pain.^29,41 Because our endpoint was limited to ED acuity, residual post-discharge hazard is outside the scope of the score and warrants safety-netting and follow-up pathways.

Score Thresholds and Nursing Application

Operational tiers were derived from training-set ROC anchors: scores <6.4 corresponded to a very high-sensitivity boundary (approximately 1% missed LTCs), 18.8 maximised the Youden index and scores >32.1 corresponded to a high-specificity extreme. Mapped to CETS, these bands offer a shared language for escalation, senior review and documentation.¹² They are decision-support thresholds, not empirical measurements of door-to-electrocardiogram or treatment intervals; time targets cited in triage policy documents should not be attributed to this observational analysis.

For nursing practice, the score can (i) standardise first-pass prioritisation when experience varies,¹³ (ii) scaffold teaching for novice triage nurses, (iii) trigger immediate senior review in Level I/II bands and (iv) enable electronic calculation without waiting for troponin. Mandatory reassessment remains essential: scores are cross-sectional snapshots, and trajectories change with treatment and evolving physiology.^12,13 Immediate Level I care is still required for cardiac arrest, refractory shock, critical respiratory failure or unresponsiveness, regardless of the calculated total.

Strengths and Limitations

Strengths include an event-rich sample; restriction to triage-available predictors; LASSO-based parsimony; comprehensive LR-versus-ML comparison with DeLong testing, calibration and decision-curve analysis; SHAP explanation aligned to the deployed model; explicit CETS mapping; and reporting oriented to TRIPOD+AI.²⁶

Several limitations warrant emphasis. First, this is a single-centre study; the hold-out set labelled “validation” is a temporal split within one hospital, which is stronger than a purely random internal split but weaker than multicentre external validation.^26,32 Second, 59.5% of registered encounters were excluded, largely because of incomplete vital signs, raising selection bias and limiting generalisability; comparisons of included versus excluded patients should accompany prospective deployment. Third, the outcome construct reflects disposition-stage acuity rather than hard clinical events, and residual incorporation bias is possible. Fourth, hyperparameter monitoring on the test set may mildly inflate test performance, although temporal validation offers a partial safeguard. Fifth, data collection (July 2020–March 2021) overlapped coronavirus disease 2019–related changes in ED case-mix and care-seeking, which may affect transport patterns and acuity distributions. Sixth, due to the economic disparities between low- and middle-income countries (LMICs) and developed countries, transferability to other triage taxonomies (eg, Emergency Severity Index, Manchester Triage System) and to health systems with different EMS ecology is unproven and will require recalibration.⁴² Seventh, chest tightness and onset time received integer points despite borderline multivariable P values because they were retained by LASSO (Figure 2A, B and Figure 4C); their incremental value should be re-tested externally. Finally, this study cannot estimate effects on waiting times, electrocardiogram timing, nurse workload or patient-centred outcomes; those endpoints need prospective implementation evaluation.

Implications for Future Research

Priorities include multicentre external validation across Chinese regions and internationally; parallel modelling of hard clinical endpoints; fairness assessment across sex and age strata; silent-mode then active-mode implementation studies with process and safety outcomes; and evaluation of sequenced pathways that combine the present first-contact score with subsequent HEART- or EDACS-based strategies.^18,19,33 Open sharing of analytic code, consistent with TRIPOD+AI open-science recommendations, would further strengthen reproducibility.

Conclusion

We developed and temporally validated a laboratory-free, logistic regression–based integer risk score for adults presenting to the ED with chest pain. Using 13 variables routinely captured at nurse triage, the score discriminated final CETS Level I/II acuity with AUCs of 0.864, 0.876 and 0.860 in the training, test and temporal validation sets, respectively, showed acceptable calibration and decision-analytic net benefit, and performed comparably to more complex machine-learning algorithms while remaining interpretable for nursing practice. Thresholds mapped to CETS Levels I–IV are proposed as decision-support anchors rather than autonomous orders.

These findings constitute preliminary single-centre evidence. The score should support, not replace, clinical judgement. Multicentre external validation, assessment of hard clinical outcomes and prospective implementation studies are required before routine clinical adoption or large-scale electronic health record deployment.

Funding Statement

The authors received no specific grant from any funding agency in the public, commercial or not-for-profit sectors.

Data Sharing Statement

The data that support the findings of this study are available from the corresponding author upon reasonable request, subject to institutional and ethical approvals.

Ethics Approval

This study was approved by the Institutional Review Board of The First Affiliated Hospital of USTC (approval number: 2022-ky020). The requirement for informed consent was waived because the study used de-identified data extracted retrospectively from routine clinical records, with no direct patient contact or intervention. The study was conducted in accordance with the Declaration of Helsinki.

Author contributions

All authors made a substantial contribution to the work reported, whether in the conception, study design, execution, acquisition of data, analysis and interpretation, or in all these areas; took part in drafting, revising or critically reviewing the article; gave final approval of the version to be published; have agreed on the journal to which the article has been submitted; and agree to be accountable for all aspects of the work. Specifically: Dongsheng Xu contributed to investigation, data curation, formal analysis, methodology, validation and drafting of the original manuscript; Ping Gao contributed to investigation, data curation and supervision; Qiang Xu contributed to formal analysis and methodology; Huanhuan Nie contributed to data curation, formal analysis, methodology, validation, drafting of the original manuscript and correspondence.

Disclosure

The authors declare that they have no competing interests.