Development and external validation of a preoperative prediction model for in-hospital major adverse cardiovascular events after hip fracture surgery in older adults: a retrospective two-center study

Abstract

Background

Older adults with hip fracture are vulnerable to postoperative major adverse cardiovascular events (MACE) during hospitalization. However, early risk stratification remains difficult because existing assessments may require complex or late-available information. A simple model based on variables routinely available before surgery may support perioperative decision-making.

Methods

We conducted a retrospective two-center cohort study of patients ≥ 60 years who underwent surgery for traumatic hip fracture at 2 hospitals in China. The development cohort included 935 patients treated between January 2024 and December 2025, and the external validation cohort included 168 patients treated between January 2020 and December 2025. Candidate preoperative predictors were screened using penalized regression. Six candidate prediction models were developed and compared using the same preprocessing pipeline. The final model was calibrated using Platt scaling. Model performance was assessed in an internal test set and an external validation cohort.

Results

In the development cohort, 122 of 935 patients developed postoperative in-hospital MACE. The final model was a Platt-calibrated logistic regression model including age, baseline electrocardiographic abnormality, chronic heart failure, serum potassium, albumin, and hemoglobin. In the internal test set, the model had an area under the receiver operating characteristic curve of 0.814, with a 95% CI of 0.748–0.876, and a Brier score of 0.102. At the prespecified threshold of 0.147, sensitivity was 0.710 and specificity was 0.754. In the external validation cohort, the area under the receiver operating characteristic curve was 0.760, with a 95% CI of 0.657–0.849. The Brier score was 0.141, sensitivity was 0.788, specificity was 0.630, and negative predictive value was 0.924.

Conclusions

A simple preoperative logistic regression model based on routinely available clinical variables showed moderate discrimination in internal testing and provided preliminary external validation evidence for predicting postoperative in-hospital MACE after hip fracture surgery in older adults. The model may support early perioperative risk stratification and postoperative surveillance, although prospective multicenter validation and recalibration are needed before wider clinical use.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s12877-026-07981-y.

Introduction

Hip fracture is a major public health problem in older adults. A Global Burden of Disease analysis showed that the global age-standardized incidence rate of hip fracture in older adults increased from 781.56 per 100,000 in 1990 to 948.81 per 100,000 in 2021, accompanied by substantial increases in both incident and prevalent cases [1]. In addition to its growing incidence, hip fracture is associated with considerable mortality, with one-year mortality in older patients reported to approach 20%–30% [2]. Diseases of the circulatory system, including ischemic heart disease and cerebrovascular disease, are an important cause of early death after hip fracture and accounted for approximately one-fifth of deaths within 30 days in a population-based study [3]. Older patients with hip fracture often have chronic comorbidities and limited physiological reserve. Fracture-related pain, blood loss, immobilization, inflammation, surgery, and anesthesia further increase physiological stress. Together, these factors may worsen underlying disease, reduce tolerance to perioperative stress, and increase susceptibility to cardiovascular complications. Perioperative cardiovascular complications therefore represent an important component of early clinical deterioration after hip fracture [4, 5].

In addition to their potential contribution to early mortality after hip fracture, postoperative in-hospital major adverse cardiovascular events (MACE) impose a substantial clinical burden among patients who survive. Even when nonfatal, these events may require intensive treatment and higher-acuity care, prolong hospitalization, and contribute to persistent functional decline, delayed rehabilitation, greater dependence, and impaired health-related quality of life [4, 6, 7]. However, baseline comorbidities alone do not reliably identify which patients will develop postoperative cardiovascular deterioration [8]. Some patients with extensive pre-existing disease remain stable after surgery, whereas others with less obvious baseline instability develop new or progressive cardiovascular events during hospitalization. Early identification of patients at increased risk of postoperative in-hospital MACE therefore remains a clinical challenge.

General perioperative cardiac risk tools have also been developed, but these tools were derived from broader noncardiac surgical populations and were not specifically designed to predict the composite of postoperative in-hospital MACE examined in older patients with hip fracture [9]. Several recent studies have explored cardiovascular risk prediction in this population using clinical variables, machine learning methods, or additional bedside assessments [10–12]. However, the available evidence remains limited because these studies have addressed different clinical scenarios and often involved broader perioperative information or additional assessments, while independent external validation of models based on routinely available preoperative information remains scarce. Therefore, further development and external validation of parsimonious models for postoperative in-hospital MACE are warranted in this population.

Therefore, we developed and externally validated a simple preoperative model to predict postoperative in-hospital major adverse cardiovascular events in older patients with hip fracture. The model used routinely available comorbidities, baseline electrocardiographic findings, and laboratory tests to support early cardiovascular risk stratification before clinical deterioration.

Methods

Study design and participants

This retrospective two-center cohort study used electronic medical records from two hospitals in China: Foshan Hospital of Traditional Chinese Medicine and the Third Affiliated Hospital of Sun Yat-sen University. This study was reported in accordance with the TRIPOD + AI statement [13]. Patients who had surgery for traumatic hip fracture were consecutively screened. Hip fractures included femoral neck and intertrochanteric fractures. Surgical procedures included internal fixation, hemiarthroplasty, and total hip arthroplasty.

In the primary center, Foshan Hospital of Traditional Chinese Medicine, 969 patients hospitalized between January 2024 and December 2025 were initially screened. After excluding 34 patients who did not meet the eligibility criteria, 935 patients were included in the development cohort. In the external validation center, the Third Affiliated Hospital of Sun Yat-sen University, 200 patients hospitalized between January 2020 and December 2025 were initially screened. Five patients did not meet the age inclusion criterion, leaving 195 eligible patients. After applying the exclusion criteria to these patients, 27 were excluded, and 168 patients were included in the external validation cohort. The same eligibility criteria and predefined study definitions were applied to both cohorts throughout their respective study periods. A flow chart of patient enrollment is shown in Fig. 1. Because this retrospective study included all consecutive eligible patients during the predefined study periods, no a priori sample size calculation was performed, and the cohort sizes were determined by the number of eligible patients available at the two centers.

Fig. 1.

Flow chart of patient enrollment. Patients from Foshan Hospital of Traditional Chinese Medicine were included in the development cohort, and patients from the Third Affiliated Hospital of Sun Yat-sen University were included in the external validation cohort

The inclusion criteria were as follows: (1) age ≥ 60 years; (2) diagnosis of traumatic hip fracture, including femoral neck fracture or intertrochanteric fracture; (3) treatment with surgical procedures during the index hospitalization, including internal fixation, hemiarthroplasty, or total hip arthroplasty; and (4) availability of perioperative medical records for assessment of candidate predictors and clinical outcomes.

The exclusion criteria were as follows: (1) non-acute, non-fracture hip disease treated as the index condition, such as osteonecrosis of the femoral head; (2) pathological fracture related to tumor, infection, or metabolic bone disease; (3) revision surgery or secondary surgery after previous hip surgery; (4) acute MACE requiring immediate clinical intervention at admission, including acute myocardial infarction, acute decompensated heart failure, cardiogenic shock, acute stroke or intracerebral hemorrhage, or hemodynamically unstable severe arrhythmia; (5) concomitant femoral shaft fracture requiring substantially more complex surgical management; and (6) incomplete medical records that precluded ascertainment of key predictors or in-hospital MACE outcomes.

The medical data were extracted from electronic medical record systems and perioperative databases, including demographic characteristics, preoperative comorbidities, laboratory tests, electrocardiographic findings, cardiac biomarkers, surgical information, and perioperative outcomes.

Outcome definition

The primary outcome was postoperative MACE, defined as major adverse cardiovascular events occurring after surgery and before hospital discharge. MACE was defined as a composite outcome including all-cause death, myocardial injury or myocardial infarction, new-onset or worsening heart failure, severe arrhythmia requiring clinical intervention, ischemic stroke or intracerebral hemorrhage, and cardiogenic shock or hemodynamic collapse.

Postoperative MACE included cardiovascular events that were not present before surgery and occurred after surgery, as well as clinically significant postoperative worsening of pre-existing cardiovascular conditions. Stable pre-existing cardiovascular disease and isolated abnormal test results were not counted as outcome events unless they were followed by clinically significant postoperative deterioration or treatment escalation. If more than one MACE component occurred in the same patient, the patient was counted once for the primary composite outcome. Detailed diagnostic criteria for each component are provided in Supplementary Methods S1.

Outcome events were identified from electronic medical records, including progress notes, laboratory results, electrocardiographic and imaging reports, treatment records, and discharge diagnoses. Two investigators independently reviewed all potential events according to predefined criteria. Disagreements were resolved by consensus or by a senior clinician. If the cause of death could not be reliably determined from retrospective records, all-cause death after surgery and before discharge was included in the composite outcome.

Predictors, missing data, and selection

Candidate predictors were selected based on clinical relevance, preoperative availability, data completeness, and standard variable-screening procedures performed within the development cohort. The initial candidate predictor set included 48 variables extracted from electronic medical records and perioperative databases, including demographic characteristics, pre-existing comorbidities, routine laboratory tests, and electrocardiographic findings. All predictors were derived from data obtained after admission and before the index operation. Measurements used to establish postoperative outcome events were not used as predictors for the same event.

Among the electrocardiographic variables, baseline electrocardiographic (ECG) abnormality was defined using the first formal ECG report obtained after admission and before surgery. Baseline ECG abnormalities included clinically relevant rhythm or conduction abnormalities, such as frequent premature atrial or ventricular contractions, atrial fibrillation or flutter, atrioventricular block, and bundle branch block. Nonspecific ST-T changes, isolated T-wave abnormalities, sinus tachycardia, and sinus bradycardia were not classified as baseline ECG abnormalities. ECG abnormalities causing hemodynamic instability or requiring urgent treatment at admission were not treated as baseline predictors because these patients met the exclusion criteria for acute cardiovascular events. Comorbidities were identified based on documented medical history, admission records, consultation records, medication records, and discharge diagnoses.

Variables with more than 30% missingness in the development cohort were excluded before model development. Variables unavailable in the development cohort or considered unreliable for routine retrospective extraction were also excluded. Retained variables with missing values were handled within the modeling pipeline. Continuous variables were imputed using medians, and categorical or binary variables were imputed using the most frequent category. Binary variables were encoded as 0 or 1. For internal testing and external validation, the same preprocessing strategy was applied. Preprocessing parameters were estimated within the model-development workflow and then applied unchanged to the internal test set and external validation cohort.

Predictor selection was performed within the model-development set. Candidate predictors were screened using univariable logistic regression and least absolute shrinkage and selection operator (LASSO) regression. LASSO regression used the one-standard-error criterion to obtain a more parsimonious predictor set. Variables supported by these analyses were then reviewed for routine preoperative availability, potential outcome-related information leakage, cross-center consistency, clinical interpretability, redundancy, and practical applicability. The final predictor set was determined before internal testing and external validation.

Statistical analysis and model development

All analyses were performed using Python 3.12.0. Data preprocessing and model development were implemented mainly using scikit-learn, and LightGBM, XGBoost, and CatBoost were implemented using their corresponding Python packages. Calibration curves, decision curve analysis, nomogram construction, and model performance evaluation were conducted using Python-based statistical and visualization tools.

Continuous variables were expressed as mean ± standard deviation or median with interquartile range, depending on their distribution. Categorical variables were expressed as counts and percentages. Between-group comparisons were performed using the Student’s t test or Mann–Whitney U test for continuous variables and the chi-square test or Fisher’s exact test for categorical variables, as appropriate. All statistical tests were two-sided, and a P value < 0.05 was considered statistically significant. Multicollinearity among the six predictors retained in the final model was assessed using variance inflation factors (VIFs).

Before predictor screening, the development cohort was stratified by outcome and randomly divided into a model-development set and an internal test set at a 75:25 ratio. The model-development set was used for predictor screening, model fitting, calibration, model selection, and threshold determination, whereas the internal test set was reserved for performance evaluation. Within the model-development set, patients were further divided, again with outcome stratification, into a model-fitting set and a calibration set at a 75:25 ratio.

Six candidate models were developed and compared: logistic regression, support vector machine, random forest, LightGBM, XGBoost, and CatBoost. The same preprocessing and model-fitting procedures were used for all candidate models. Random oversampling was applied only within the model-fitting set to achieve a 1:1 event-to-nonevent ratio by resampling the minority class with replacement; no synthetic observations were generated, while the calibration set, internal test set, and external validation cohort retained their original event proportions. Predicted probabilities from each model were calibrated in the calibration set using Platt scaling, and classification thresholds were determined in the same calibration set by maximizing the Youden index. The calibrated models and their prespecified thresholds were evaluated in the internal test set. The final selected model and its fixed threshold were then applied to the external validation cohort without refitting, recalibration, model reselection, or threshold adjustment.

Model performance was assessed using the area under the receiver operating characteristic curve, sensitivity, specificity, accuracy, F1 score, Brier score, calibration plots, and decision curve analysis. The final model was selected before external validation based on overall discrimination, calibration, clinical utility, and interpretability.

External validation and model interpretation

In external validation, the fixed model and threshold were applied to the external cohort without any model updating. The fixed threshold derived from the calibration set was used for threshold-dependent metrics in both the internal test set and external validation cohort.

Model discrimination was assessed using the area under the receiver operating characteristic curve and the area under the precision-recall curve. Calibration was assessed using calibration curves, the Brier score, calibration intercept, calibration slope, and observed-to-expected event ratio. Clinical usefulness was assessed using decision curve analysis. Sensitivity, specificity, positive predictive value, negative predictive value, accuracy, and F1 score were calculated at the fixed threshold. Confusion matrices were used to show classification results at this threshold. Confidence intervals for threshold-dependent proportions were calculated using binomial methods. Bootstrap resampling with 1000 replicates was used to estimate 95% confidence intervals for the area under the receiver operating characteristic curve.

If a regression-based model was selected as the final model, a nomogram was constructed to graphically present the individualized predicted probability of postoperative in-hospital MACE. SHAP values were calculated for the final selected model to describe the contribution and direction of effect of each predictor. SHAP analyses were used only for model interpretation and were not used for predictor selection or model updating.

Results

Patient characteristics

A total of 935 patients admitted to Foshan Hospital of Traditional Chinese Medicine between January 2024 and December 2025 were included in the development cohort, of whom 122 developed postoperative in-hospital MACE. Baseline characteristics of the development cohort according to postoperative MACE status are summarized in Table 1.

Table 1.

Baseline characteristics of the development cohort according to postoperative in-hospital MACE status

Variable	Overall	Non-MACE group	MACE group	P value
Number of patients	n = 935	n = 813	n = 122
Age, years, median (IQR)	80.0 (74.0–86.0)	79.0 (73.0–85.0)	84.0 (79.0–89.0)	< 0.001
Male sex, n (%)	296 (31.7)	251 (30.9)	45 (36.9)	0.183
Diabetes mellitus, n (%)	174 (18.6)	142 (17.5)	32 (26.2)	0.020
Coronary artery disease, n (%)	66 (7.1)	49 (6.0)	17 (13.9)	0.001
Chronic heart failure, n (%)	90 (9.6)	58 (7.1)	32 (26.2)	< 0.001
History of stroke, n (%)	123 (13.2)	98 (12.1)	25 (20.5)	0.010
Baseline ECG abnormality, n (%)	213 (22.8)	161 (19.8)	52 (42.6)	< 0.001
Atrial fibrillation/flutter, n (%)	34 (3.6)	17 (2.1)	17 (13.9)	< 0.001
Frequent premature beats, n (%)	50 (5.3)	35 (4.3)	15 (12.3)	< 0.001
Short-run atrial tachycardia, n (%)	24 (2.6)	14 (1.7)	10 (8.2)	< 0.001
Bundle/fascicular block, n (%)	94 (10.1)	76 (9.3)	18 (14.8)	0.064
Atrioventricular block, n (%)	44 (4.7)	36 (4.4)	8 (6.6)	0.300
Other clinically relevant rhythm/conduction abnormality, n (%)	23 (2.5)	17 (2.1)	6 (4.9)	0.106
Serum potassium, mmol/L, median (IQR)	3.8 (3.5–4.1)	3.8 (3.5-4.0)	3.9 (3.6–4.2)	0.002
Albumin, g/L, median (IQR)	35.6 (33.3–38.0)	35.8 (33.4–38.2)	34.4 (31.6–37.4)	< 0.001
Hemoglobin, g/L, median (IQR)	114.0 (100.0-125.0)	114.0 (102.0-126.0)	108.0 (92.0-119.8)	< 0.001

Values are presented as n (%) or median (interquartile range), unless otherwise indicated. Patients were stratified according to the occurrence of postoperative in-hospital MACE. ECG abnormality subtypes were not mutually exclusive; therefore, the sum of ECG abnormality subtypes may exceed the total number of patients with baseline ECG abnormality

MACE major adverse cardiovascular events, ECG electrocardiography, IQR interquartile range

Compared with patients without postoperative in-hospital MACE, those who developed MACE were older and more frequently had diabetes mellitus, coronary artery disease, chronic heart failure, history of stroke, and baseline ECG abnormality. They also had higher serum potassium levels and lower albumin and hemoglobin concentrations.

Height, body weight, and body mass index were not analyzed because these measurements were not consistently or reliably documented in this older hip fracture population, particularly because of acute pain, limited mobility, and difficulty obtaining accurate anthropometric data after injury.

Components of postoperative in-hospital MACE

The distribution of individual postoperative in-hospital MACE components in the development cohort is shown in Table 2. Heart failure-related events were the most frequent component, followed by myocardial injury or myocardial infarction and severe arrhythmia requiring clinical intervention. Because some patients experienced more than one component event, the individual components were not mutually exclusive. The corresponding distribution in the external validation cohort is provided in Supplementary Table S1.

Table 2.

Components of postoperative in-hospital MACE in the development cohort

Postoperative MACE component	Development cohort (n = 122)
All-cause death	14 (11.5)
Myocardial injury / myocardial infarction	31 (25.4)
Heart failure-related event	68 (55.7)
Severe arrhythmia requiring clinical intervention	22 (18.0)
Ischemic stroke / intracerebral hemorrhage	2 (1.6)
Cardiogenic shock / hemodynamic collapse	2 (1.6)

Values are presented as n (%). Percentages were calculated among patients with postoperative in-hospital MACE in the development cohort. Individual MACE components were not mutually exclusive; therefore, a patient could contribute to more than one component but was counted only once for the primary composite outcome

MACE major adverse cardiovascular events

Model development and internal validation performance

Univariable logistic regression and LASSO regression using the 1-SE criterion were performed in the model-development set. Detailed results of the univariable analyses for all candidate variables are provided in Supplementary Table S2. The LASSO coefficient profiles and cross-validation results are shown in Supplementary Figure S1. Among the 48 initial candidate variables, LASSO selected 18 variables, of which eight overlapped with variables showing P < 0.05 in univariable analysis: age, baseline ECG abnormality, chronic heart failure, serum potassium, albumin, troponin I, hemoglobin concentration, and lactate dehydrogenase. After clinical review considering routine preoperative availability, cross-center consistency, potential outcome-related information leakage, redundancy, interpretability, and practical applicability, troponin I and lactate dehydrogenase were not retained.

The final model was therefore constructed using six routinely available and clinically interpretable preoperative predictors: age, baseline ECG abnormality, chronic heart failure, serum potassium, albumin, and hemoglobin concentration. The VIF values for the six final predictors ranged from 1.05 to 1.42, with all values below 1.5, indicating no substantial multicollinearity. Baseline ECG abnormality and chronic heart failure were treated as binary variables, and continuous variables were entered as continuous predictors after preprocessing.

Using these six predictors, six Platt-calibrated candidate models were developed and evaluated in the internal test set. The hyperparameter-tuning procedure and final model-specific hyperparameter settings are provided in Supplementary Table S3. The internal test performance of these models is summarized in Table 3, and the corresponding ROC curves are shown in Fig. 2. Logistic regression achieved the highest AUC of 0.814 (95% CI: 0.748–0.876), followed closely by support vector machine and random forest. LightGBM showed acceptable but slightly lower discrimination, whereas CatBoost and XGBoost showed lower discrimination in the internal test set.

Table 3.

Internal test performance of the six Platt-calibrated candidate models

Model	AUC (95% CI)	PR-AUC	Brier score	Sensitivity (95% CI)	Specificity (95% CI)	Accuracy (95% CI)	F1 score
Logistic regression	0.814 (0.748–0.876)	0.343	0.102	0.710 (0.520–0.858)	0.754 (0.688–0.811)	0.748 (0.687–0.802)	0.427
SVM	0.813 (0.746–0.875)	0.335	0.102	0.742 (0.554–0.881)	0.744 (0.678–0.802)	0.744 (0.683–0.798)	0.434
Random forest	0.811 (0.745–0.871)	0.313	0.103	0.645 (0.454–0.808)	0.788 (0.725–0.842)	0.769 (0.710–0.822)	0.426
LightGBM	0.790 (0.722–0.858)	0.287	0.104	0.548 (0.360–0.727)	0.833 (0.774–0.881)	0.795 (0.737–0.845)	0.415
CatBoost	0.710 (0.630–0.798)	0.242	0.109	0.581 (0.391–0.755)	0.714 (0.647–0.775)	0.697 (0.633–0.755)	0.336
XGBoost	0.694 (0.592–0.779)	0.256	0.110	0.742 (0.554–0.881)	0.552 (0.481–0.621)	0.577 (0.511–0.641)	0.317

Values are presented as estimates with 95% confidence intervals where applicable. Confidence intervals for AUC were estimated using resampling methods, whereas confidence intervals for threshold-dependent proportions were calculated using binomial methods

AUC area under the receiver operating characteristic curve, PR-AUC area under the precision-recall curve, CI confidence interval, SVM support vector machine, MACE major adverse cardiovascular events

Fig. 2.

Receiver operating characteristic curves of the six Platt-calibrated candidate models in the internal test set. Predicted probabilities were generated by models trained on the model-fitting set and calibrated using the independent calibration set. AUC values are shown in the legend. AUC, area under the receiver operating characteristic curve; ROC, receiver operating characteristic; SVM, support vector machine

Logistic regression was selected as the final model because it provided the most favorable balance of discrimination, calibration, clinical utility, and interpretability. It also achieved the highest AUC in the internal test set. The coefficients, intercept, calibration parameters, and imputation medians are provided in Supplementary Table S4. For the final model, the Youden-derived threshold determined in the independent calibration set was 0.147 and was fixed before internal test evaluation and external validation. At this fixed threshold, the final model achieved a sensitivity of 0.710, specificity of 0.754, accuracy of 0.748, F1 score of 0.427. Additional threshold-dependent metrics and calibration measures for the final model are summarized in Supplementary Table S5.

In a targeted sensitivity analysis, all-cause death was removed as an independent component of the composite outcome, while patients who died but experienced another qualifying postoperative MACE component remained classified as having MACE. Using the original fixed model and threshold without refitting or recalibration, the AUC in the internal test set changed from 0.814 to 0.798. Full results are provided in Supplementary Table S6.

Calibration and decision curve analysis in the internal test set

The calibration performance and clinical utility of the final Platt-calibrated logistic regression model were further evaluated in the internal test set. In the internal test set, the final model had a Brier score of 0.102 and an observed-to-expected event ratio of 1.006, indicating close overall agreement between the observed and predicted numbers of events.

Decision curve analysis was evaluated across threshold probabilities from 0.01 to 0.35. The final model provided greater net benefit than both the treat-all and treat-none strategies across threshold probabilities of approximately 0.05 to 0.30 in the internal test set. Its net benefit was generally comparable to that of the support vector machine and random forest models. The calibration curves and decision curve analyses of the six candidate models are shown in Figs. 3 and 4.

Fig. 3.

Calibration curves of the six Platt-calibrated candidate models in the internal test set. The curves compare mean predicted probabilities with observed event rates across risk groups. The dashed diagonal line represents perfect calibration

Fig. 4.

Decision curve analysis of the six Platt-calibrated candidate models in the independent internal test set. Decision curve analysis was performed to evaluate the net benefit of each model across threshold probabilities from 0.01 to 0.35. The dotted line represents the treat-all strategy, and the dashed horizontal line represents the treat-none strategy

External validation performance

The final Platt-calibrated logistic regression model was externally validated in an independent cohort from the Third Affiliated Hospital of Sun Yat-sen University between January 2020 and December 2025. The cohort included 168 patients; 33 developed postoperative in-hospital MACE, corresponding to an event rate of 19.6%. The same six predictors and the fixed threshold were used without model refitting, recalibration, model reselection, or threshold adjustment.

In external validation, the model achieved an AUC of 0.760 (95% CI, 0.657 to 0.849), a PR-AUC of 0.462, and a Brier score of 0.141. At the fixed threshold of 0.147, sensitivity was 0.788 and specificity was 0.630. Detailed threshold-dependent metrics, confusion matrix results, and calibration measures are summarized in Supplementary Table S5.

The distributions of the six predictors in the development and external validation cohorts and the corresponding standardized mean differences are summarized in Supplementary Table S7. These comparisons were descriptive and were not used for model updating. ROC curves, calibration curves, and decision curve analyses for the final logistic regression model in the internal test set and external validation cohort are shown in Fig. 5.

Fig. 5.

Internal test and external validation performance of the final Platt-calibrated logistic regression model. a Receiver operating characteristic curves. b Calibration curves. c Decision curve analysis. The internal test set and external validation cohort are shown for comparison. The dashed diagonal line in panel a represents the reference line, and the dashed diagonal line in panel b represents perfect calibration. In panel c, the dotted line represents the treat-all strategy, and the dashed horizontal line represents the treat-none strategy. AUC, area under the receiver operating characteristic curve; ROC, receiver operating characteristic

In the corresponding sensitivity analysis, in which all-cause death was removed as an independent component of the composite outcome while patients with another qualifying postoperative MACE component remained classified as having MACE, the AUC in the external validation cohort remained essentially unchanged (0.761 vs. 0.760 in the primary analysis). Other performance measures were also broadly comparable. Full results are provided in Supplementary Table S6.

Nomogram-based model interpretation

A nomogram was constructed based on the final logistic regression model to provide a visual representation of individualized risk estimation. The nomogram showed that chronic heart failure, baseline ECG abnormality, older age, higher serum potassium, lower albumin, and lower hemoglobin concentration were associated with a higher predicted risk of postoperative in-hospital MACE. The nomogram is shown in Fig. 6.

Fig. 6.

Nomogram of the final Platt-calibrated logistic regression model. The nomogram provides a graphical representation of the contribution of age, baseline ECG abnormality, chronic heart failure, serum potassium, albumin, and hemoglobin concentration to individualized risk estimation for postoperative in-hospital MACE. ECG, electrocardiography

Model interpretation using SHAP analysis

SHAP analysis was performed in the internal test cohort to evaluate the contribution and direction of each predictor in the final logistic regression model. The SHAP summary plot showed that older age, baseline ECG abnormality, chronic heart failure, and higher serum potassium tended to shift model predictions toward a higher risk of postoperative in-hospital MACE, whereas higher albumin and hemoglobin levels shifted predictions toward a lower risk. The SHAP summary plot is shown in Fig. 7, and global and individual-level SHAP explanations, dependence plots for selected continuous predictors, and the decision plot are provided in Supplementary Figures S2–S4.

Fig. 7.

SHAP summary plot for model interpretation in the internal test cohort. Each dot represents one patient. The x-axis shows the SHAP value on the calibrated log-odds scale, where positive values indicate an increased predicted risk of postoperative in-hospital MACE and negative values indicate a decreased predicted risk. Color represents the feature value, with red indicating higher values and blue indicating lower values. Variables are ranked by their mean absolute SHAP values

Discussion

In this retrospective two-center study, we developed and externally validated a preoperative model for predicting postoperative in-hospital MACE in older patients with traumatic hip fracture. Using six routinely available predictors, the final Platt-calibrated logistic regression model demonstrated moderate-to-good discrimination in the internal test set and acceptable discrimination in the external validation cohort. These findings suggest that routinely available preoperative clinical information may help identify patients who require closer postoperative cardiac surveillance.

The clinical rationale for this model is that older patients with hip fracture have highly heterogeneous perioperative vulnerability. Hip fracture can act as an acute systemic stressor superimposed on advanced age, pain, immobilization, blood loss, inflammation, and reduced physiological reserve [14, 15]. Chronic comorbidities further increase risk, but comorbidity count alone does not fully capture perioperative vulnerability. Some patients with several chronic diseases remain stable, whereas others with few documented comorbidities may decompensate after minor perioperative stress [8, 14]. Therefore, early risk assessment should not simply label patients as high risk because they have multiple comorbidities. The more clinically relevant question is whether these baseline conditions and vulnerabilities are likely to translate into clinically significant adverse events, such as MACE, during the perioperative course.

This distinction is difficult in routine practice because relevant risk signals are scattered across medical history, electrocardiographic findings, and laboratory tests. Their interpretation often depends on individual clinical experience. The present model integrates these routinely available preoperative data into an individualized estimate of postoperative in-hospital MACE risk. Consequently, it may help clinicians identify patients who need closer surveillance, earlier cardiology consultation, and more proactive perioperative optimization, while supporting more consistent interpretation of low-risk predictions.

Liu et al. developed a nomogram for predicting postoperative cardiac events; however, the study did not compare multiple candidate algorithms within a unified evaluation framework, and the model underwent internal validation only, without independent external validation [10]. The prospective LUSHIP study demonstrated the potential value of preoperative lung ultrasound for MACE risk assessment, but this approach requires an additional bedside examination and trained operators [11]. Biomarker-based approaches, particularly those involving NT-proBNP, have also shown prognostic value, although these tests are generally performed selectively rather than routinely in all patients [16]. Our study developed and compared six candidate models within a unified preprocessing, calibration, and evaluation framework, selected the final model according to overall performance and interpretability, and evaluated it in an independent external cohort. The final model uses six routinely available preoperative variables, does not require additional specialized examinations, and may therefore be more readily incorporated into early perioperative risk assessment.

The clinical plausibility of the final model is supported by the potential pathophysiological relevance of its predictors. Chronic heart failure reflects impaired cardiac reserve and greater susceptibility to decompensation during perioperative blood loss, fluid shifts, inflammation, and sympathetic activation [17], while baseline ECG abnormality may indicate underlying rhythm or conduction disease and increased electrical vulnerability under perioperative stress [9]. Older age is associated with reduced physiological and cardiovascular reserve and a greater burden of frailty [9, 14]. Serum potassium directly influences myocardial excitability and conduction, and disturbed potassium homeostasis may increase susceptibility to perioperative arrhythmia [18]. Lower albumin may reflect systemic inflammation, nutritional depletion, frailty, and reduced physiological reserve [19], whereas lower hemoglobin reduces oxygen-carrying capacity and may aggravate myocardial oxygen supply–demand imbalance during perioperative stress [20]. These predictors should therefore be interpreted primarily as clinically plausible markers of underlying cardiovascular vulnerability and reduced tolerance to perioperative stress, rather than as isolated causal determinants of MACE.

Troponin I and lactate dehydrogenase were not retained despite being supported by statistical screening. Preoperative troponin testing is not uniformly available and is often selectively performed in patients already considered at cardiovascular risk, which may introduce informative missingness and limit cross-center applicability. In addition, assay methods and reference ranges vary across institutions. Lactate dehydrogenase is nonspecific and, in patients with hip fracture, may partly reflect fracture-related tissue injury rather than underlying cardiovascular vulnerability. Their exclusion was therefore intended to preserve the early availability and transportability of the final model.

The external validation provided preliminary evidence of model performance in an independent clinical setting. At the prespecified threshold, the model showed relatively high sensitivity and negative predictive value, suggesting that it may be most useful for identifying patients at lower predicted risk. However, the modest PPV, F1 score, and PR-AUC indicate limited precision in identifying positive cases. Given the limited sample size and number of events in the external cohort, these findings should be interpreted cautiously. Positive predictions should prompt closer clinical review rather than be interpreted as evidence that MACE will occur. Therefore, its main clinical value lies in prompting closer surveillance, earlier consultation, and perioperative optimization when appropriate.

The selection of logistic regression as the final model also supports clinical interpretability and implementation. Although several machine learning algorithms were evaluated using the same training, calibration, and validation framework, logistic regression provided the most favorable balance between performance, calibration, and transparency. It also allowed construction of an interpretable nomogram, which may facilitate communication of risk estimates in clinical practice. This simplicity is important because a model intended for early perioperative risk stratification should be easy to understand, reproduce, and apply.

This study has several strengths. The model was developed and externally validated using data from two independent clinical centers. It was based on six objective and routinely available preoperative predictors, which may facilitate implementation in standard perioperative assessment. In addition, model development, calibration, internal testing, and external validation were kept separate to reduce information leakage and optimistic performance estimation. Model performance was also evaluated from several complementary perspectives, including discrimination, calibration, decision curve analysis, and nomogram-based interpretation.

Several limitations should be acknowledged. First, postoperative in-hospital MACE was assessed only from surgery to discharge; therefore, events occurring after discharge were not captured. Second, the retrospective design relied on routinely documented clinical records and may be affected by incomplete documentation, residual confounding, and missing data. Third, although external validation was performed in an independent cohort, its sample size and number of events were limited. Differences in enrollment period, annual case volume, patient characteristics, and clinical practice between the two centers may also have introduced temporal heterogeneity and center-related selection bias. Broader contemporary multicenter validation is therefore required. Fourth, all-cause death was included as a component of MACE because cause-specific death could not always be reliably adjudicated from retrospective records, which may have introduced outcome heterogeneity. Nevertheless, model discrimination remained broadly similar in a targeted sensitivity analysis in which deaths without another qualifying postoperative MACE component were reclassified as non-events. In addition, heart failure-related events were the most frequent component of the composite outcome. Therefore, the overall model performance may have been influenced more strongly by predictors associated with postoperative heart failure than by less frequent components such as stroke or cardiogenic shock. The model should consequently be interpreted as predicting the prespecified composite outcome rather than each individual MACE component with equal accuracy. Finally, the model used routinely available preoperative variables collected early after admission. This design supports early risk stratification, but the model cannot update risk according to subsequent changes in laboratory results, perioperative physiology, or postoperative clinical trajectories. Therefore, it should be interpreted as an early preoperative risk-stratification tool rather than a dynamic prediction model. Future prospective multicenter studies are needed to confirm its generalizability, evaluate its impact on clinical decision-making, and determine whether local recalibration is required.

In conclusion, we developed and externally validated a preoperative model for predicting postoperative in-hospital MACE in older patients with hip fracture. The final Platt-calibrated logistic regression model incorporated six routinely available variables and demonstrated moderate-to-good internal discrimination and acceptable external discrimination. The model may help identify patients who require closer postoperative surveillance and earlier perioperative optimization. Future prospective multicenter studies are needed to confirm the model’s generalizability, evaluate its impact on clinical decision-making, determine whether local recalibration is required, and standardize the collection of anthropometric measures.

Abbreviations

AUC

Area under the receiver operating characteristic curve

BMI

Body mass index

CI

Confidence interval

DCA

Decision curve analysis

ECG

Electrocardiography

IQR

Interquartile range

LASSO

Least absolute shrinkage and selection operator

LightGBM

Light gradient boosting machine

MACE

Major adverse cardiovascular events

NPV

Negative predictive value

PR-AUC

Area under the precision-recall curve

PPV

Positive predictive value

ROC

Receiver operating characteristic

SHAP

Shapley additive explanations

SVM

Support vector machine

XGBoost

Extreme gradient boosting

Authors' contributions

Ruimin Luo and Jirong Yang contributed equally to this work and share first authorship. They contributed to conceptualization, data curation, formal analysis, methodology, model development, validation, visualization, and writing of the original draft. Jieqiang Du, Siyang Zeng, Tianqiang Sheng, and Shuotao Zhong contributed to investigation, data curation, outcome verification, and manuscript review and editing. Chaojin Chen and Yuyuan You contributed equally as corresponding authors and were responsible for supervision, project administration, methodology, clinical interpretation, and critical revision of the manuscript. All authors reviewed and approved the final manuscript.

Funding

This work was conducted under the Foshan Medical Research Project (Project No. 20250281), with institutional self-funding provided by Foshan Hospital of Traditional Chinese Medicine. The funder had no role in the study design, data collection, data analysis, interpretation of the results, preparation of the manuscript, or decision to submit the manuscript for publication.

Data availability

The datasets generated and analyzed during the current study are not publicly available because they contain de-identified patient-level clinical data subject to institutional and ethical restrictions, but may be available from the corresponding author upon reasonable request and with permission from the participating institutions.

Declarations

Ethics approval and consent to participate

This retrospective two-center study was approved by the Ethics Committee of Foshan Hospital of Traditional Chinese Medicine (approval no. KY [2026]174; approval date: May 12, 2026). All methods were carried out in accordance with relevant institutional guidelines and regulations, including the Declaration of Helsinki. The requirement for informed consent to participate was waived by the Ethics Committee because the study used retrospectively collected clinical data and involved no additional intervention. Data from the external validation center, the Third Affiliated Hospital of Sun Yat-sen University, were used with institutional approval. All data were anonymized before analysis. The study was registered with the Chinese Clinical Trial Registry (registration number: ChiCTR2600125955).

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.