Introduction

The aging global population is increasing at an unprecedented rate, particularly in Asia and other developing countries. With this demographic shift, the incidence of hip fractures is projected to nearly triple between 2018 and 2050, with over 50% of the cases expected to occur in Asia1,2. In China, the world’s most populous nation, over 366 million people, or more than 26.1% of the population, are projected to be aged 65 years and older by 20503. This demographic change poses significant challenges to healthcare systems, especially in managing the increasing number of elderly patients at a high risk of hip fractures4.

Timely surgical intervention is critical for reducing mortality and improving functional outcomes in elderly patients with hip fractures5,6,7. However, despite the benefits of surgery, elderly patients undergoing hip fracture repair are at high risk of developing perioperative pulmonary complications (POPCs)8,9.These complications are particularly prevalent in elderly patients and substantially increase the risk of perioperative mortality8,10. The incidence of POPCs in elderly patients with hip fractures ranges from 13.9 to 33.3% depending on the underlying disease, preoperative pulmonary dysfunction, and diminished compensatory capacity11,12,13.

Current research on pulmonary complications in hip fracture surgery primarily focuses on postoperative complications (PPCs)14,15,with limited attention paid to the preoperative and intraoperative factors. Furthermore, existing predictive models primarily concentrate on postoperative outcomes, often neglecting intraoperative risk factors, and lacking specificity for elderly patients16,17,18,19. This results in a gap in the accurate identification of high-risk patients throughout the perioperative period, which is crucial for improving clinical decision making and patient outcomes.

To address these gaps, this study aimed to develop and validate a predictive model that incorporates preoperative, intraoperative, and postoperative risk factors to more accurately predict POPCs in elderly patients undergoing hip fracture surgery. By employing a combination of retrospective and prospective cohort designs, this study offers a comprehensive risk-assessment tool specifically tailored to the elderly population. The novelty of this approach lies in its holistic evaluation of perioperative risk factors, addressing the limitations of existing models, enhancing the accuracy of risk prediction, and ultimately, improving patient outcomes.

Methods

Study design overview

This single-center, bidirectional cohort study will include both retrospective and prospective studies. This study was approved by the Institutional Review Board of the Fengdu People’s Hospital of Chongqing (Ethical Approval ID: 2023SC0915-149) and registered with ClinicalTrials.gov (ChiCTR2300071115) on 5 March 2023 to ensure transparency and adherence to best practices. The retrospective phase spans 1 January 2017 to 31 August 2023 and the prospective phase extends from 1 September 2023 to 30 September 2024. Model development and internal validation will take place between 1 September 2023 and 30 September 2024 followed by external validation between 1 October 2024 and 31 December 2024. A total of 3667 participants from the Fengdu People’s Hospital of Chongqing, China will be recruited for the study. During the trial, participants will be evaluated for POPCs and other postoperative complications. Study assessments will be conducted at three key time points: preoperatively (day–n), intraoperatively (day 0), and postoperatively (days 1–30)20 (Fig. 1).

Fig. 1
Fig. 1The alternative text for this image may have been generated using AI.
Full size image

Flow diagram of enrolment and assessment.

Participants

Inclusion criteria

  1. 1.

    Patients aged ≥ 65 years21.

  2. 2.

    Radiographically confirmed hip fracture.

  3. 3.

    Scheduling for surgical treatment of fractures.

Exclusion criteria

  1. 1.

    Refusal to participate in the study.

  2. 2.

    Inability to follow the study protocol due to cancellation of the procedure.

  3. 3.

    Withdrawal was initiated by the patients or their families.

  4. 4.

    Incomplete data or loss to follow-up for other reasons.

  5. 5.

    Polytrauma (multiple fracture sites).

Data collection

Data will be collected in two phases, retrospective and prospective. Retrospective data will be obtained from the hospital records to provide information on patient demographics, preoperative assessments, intraoperative details, and postoperative outcomes. The data will be divided into training and validation subsets at a 7:3 ratio using the R software to facilitate model development22.

For prospective data, the collection will adhere to a standardised protocol to ensure consistency across patients. Monitoring will be performed from hospital admission until 30 days after surgery. All data will be entered into a secure, password-protected Epidata V.4.6 database with restricted access to maintain data integrity. Following data entry, information will be exported for analysis by an independent biostatistician to ensure unbiased and reliable results.

Data sources will include patient medical records, electronic health records (EHRs), and relevant clinical databases. The collection processes will ensure consistency across the retrospective and prospective phases to maintain high data quality. Six individuals, including the BT. J., JT.S., XJ.D., JH.C., L.Q., and YL.T. were included in the data-collection process. To minimize potential bias, preoperative imaging was performed and interpreted using a blinded approach, in which clinicians and radiologists were unaware of the patients’ study group allocation.

Primary objective and reporting guidelines

The primary objective of this study was to develop and validate predictive models for POPCs, including both the internal and external validation stages (Fig. 2). This study will strictly adhere to the reporting requirements of the Transparent Reporting of a multivariable prediction model for Individual Prognosis or Diagnosis (TRIPOD) framework23. To ensure consistency and rigor, Supplementary Table S1 provides detailed information on the reporting criteria and methodology.

Fig. 2
Fig. 2The alternative text for this image may have been generated using AI.
Full size image

Flow chart. LASSO least absolute shrinkage and selection operator, ROC receiver operating characteristic, DCA decision curve analysis.

Predictor variables and outcomes

Predictor variables

This study identified the key predictor variables that were clinically relevant and quantifiable. These predictors have been selected through a comprehensive review of the relevant literature, expert consensus, and their practical applicability in the clinical setting9,10,14,15,24,25,26,27,28,29,30,31. The variables will be assessed before, during, and after the surgery. The perioperative indicators are listed in Table 1.

Table 1 Predictive variables considere in the model.
Full size table

Outcomes

Primary outcomes

This study focused on assessing POPCs from the decision to undergo surgery to 30 days after surgery or until discharge, prioritising the earlier period. It specifically considers patients released within this timeframe.

POPCs including pulmonary infection, respiratory failure, pleural effusion, atelectasis, bronchospasm, pulmonary embolism, aspiration pneumonia, acute respiratory distress syndrome (ARDS), pulmonary edema, and an unexpected need for extended mechanical ventilation8,31,32,33. Each instance of POPCs requires collaborative diagnosis based on detailed patient data, including progress notes, medical records, laboratory tests, and imaging reports. Senior radiologists will assess the perioperative images. All diagnoses will be made based on the combined opinions of the clinician and radiologist. The definitions of the POPCs are listed in Table 29,10,25,34.

Table 2 Perioperative pulmonary complications and their definitions.
Full size table

Secondary outcomes

Our study will evaluate postoperative complications such as heart failure, myocardial infarction, renal failure, deep venous thrombosis, stroke, and sudden death, starting from the decision to operate and continuing until 30 days after surgery or discharge.

Sample size calculation

Retrospective cohort

To ensure robust model parameter estimation, we set the variable count to (k + 1 = 45), where (k) represents the study variable and includes 44 variables. Based on unpublished data(Supplementary Table S2), we estimated the incidence of POPCs (φ) in older patients who underwent surgery for hip fractures to be 21.43%. The sample size was calculated using Eq. (1)35.

$$\:n=\frac{EPV\times\:(k+1)}{{\upphi\:}}=\frac{10\times\:(44+1)}{0.2143}\approx\:2099.86\approx\:2100$$
(1)

Sample size for external validation (prospective cohort)

To ensure the accuracy of the external validation of the model, the required sample size will be calculated based on the O/E ratio (observed/expected). Debray et al. proposed the use of the delta method to compute the standard error (SE) of the O/E ratio using the following Eq. (2)36:

$$\:SE\left(\text{l}\text{n}\right(\frac{O}{E}\left)\right)\approx\:\sqrt{\frac{1-\varnothing\:}{O}}$$
(2)

Recognizing that O = Nϕ, the rearranged Eq. (3) yields the minimum sample size (N) required for external validation37.

$$\:\text{N}=\frac{1-{\varnothing}}{{\varnothing}\left(\text{S}\text{E}\right(\text{I}\text{N}\left(\text{O}\right)/\text{E}\left)\right){)}^{2}}$$
(3)

Assuming that the ϕ is 0.2143 in the population undergoing external validation and that the observed-to-expected (O/E) ratio is 1, the objective is to confirm model calibration by achieving a 95% confidence interval for O/E with a width of 0.2. The corresponding SE on the ln(O/E) scale is approximately 0.051. The sample size for external validation was determined using Eq. (4).

$$\:N=\frac{1-0.2143}{0.2143\times\:0.05{1}^{2}}\approx\:1409.59\approx\:1410$$

We estimated the loss of follow-up rate to be 10%; therefore, we increased the sample size to 1567 to allow for dropouts.

Missing data and outlier handling

Missing data

Prior to statistical analyses, we will review the initial dataset to address missing values. Variables with > 10% missing values after supplementation attempts will be excluded. If the missing value rate decreases to less than 10%, a random test will be conducted to evaluate the missingness pattern and guide subsequent actions38.

To manage different types of missingness, we will employ the following strategies39: Mean imputation will be used for completely random missing data with normal or near-symmetrical distributions; Median imputation will be applied to the right-skewed data or outliers; Multiple imputations will be performed for missing random data considering the relationships within the observed data; For data not missing at random, the‘mincemeat’package in R, which is specifically designed for such cases, will be employed.

Outlier handling

Outliers will be systematically identified and managed to maintain analysis integrity and prevent the exclusion of legitimate data points. Outliers, defined as data points significantly deviating from the overall pattern, are values beyond 1.5 × the interquartile range (IQR) from the first and third quartiles40, or with z-scores exceeding ± 3 in normally distributed data. Box plots will visually inspect outliers using the IQR method, while z-scores will identify data points more than ± 3 standard deviations from the mean41. For skewed data, outliers were defined as values outside [Q1–1.5 × IQR, Q3 + 1.5 × IQR]. The influence of outliers on model performance will be assessed using leverage analysis and Cook’s distance, excluding significant outliers with leverage greater than twice the average. Informative outliers reflecting valid variations will be retained by applying robust statistical methods, such as robust regression or log transformation, to mitigate their impact. The model’s performance will then be re-evaluated to ensure accurate variability accounting, documenting any performance metric changes in the Results section.

Statistical methods and model development

Statistical analysis

Statistical analysis will be performed via R software (version 4.0.4). Dichotomous data will be presented as percentages, and theχ²test will be used for analysis. Continuous data are shown as mean ± standard deviation or median (interquartile range, 25th–75th percentiles) based on their distribution, and t-tests or Mann-Whitney U tests will be used accordingly. A significance threshold will be established at P < 0.05, and an absolute standardised difference (ASD) over 0.118 indicated an imbalance between the development and validation groups.

Model development

  1. 1.

    Data privacy and preliminary handling.

    During the initial phase of the model development, retrospective cohort data comprising 2100 samples will be collected to train the model. The dataset will be divided into a 70% training group (n = 1470) and 30% validation group (n = 630). Critical steps to ensure data quality and privacy protection include de-identification by replacing sensitive attributes, data cleaning to correct errors, assessing outliers, and addressing missing values via established statistical methods. Descriptive statistics will be used to summarise the data distribution, and visual analysis via histograms, scatter plots, and box plots will be used to identify the patterns and anomalies. Finally, multiple imputations or mean/median imputations will be employed to address data loss and ensure the integrity of the dataset.

  2. 2.

    Variable selection.

    Least absolute shrinkage and selection operator (LASSO) and logistic regression analyses will be used to identify the key variables influencing POPCs. Significant variables will be selected based on the coefficients or p-values obtained from the regression analysis. This process is essential for eliminating redundant variables and retaining only those variables that significantly affect the prediction outcomes.

  3. 3.

    Model construction.

    A multivariate logistic regression model will be developed by using the identified variables. The Akaike information criterion (AIC) evaluates the influence of these variables by filtering out less influential factors. A nomogram prediction model will be constructed via the RMS package in the R software.

  4. 4.

    Model validation.

    The remaining 30% of the retrospective dataset, consisting of 630 samples, will be used for internal validation. The model will also be tested in a prospective cohort study to verify its effectiveness in predicting POPCs in new patient groups. The performance of the model will be rigorously evaluated via receiver operating characteristic (ROC) curve analysis, decision curve analysis (DCA), and the Hosmer–Lemeshow test. Key metrics, such as sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and area under the curve (AUC-ROC), will be used to determine the efficacy of the model. The precision and recall metrics will be used to assess the accuracy and reliability of the model in the external validation set. Prospective validation will be employed to assess the model’s performance in a real-world setting, guiding decisions regarding recalibration, replacement, or retirement of the active model based on its applicability and effectiveness.

  5. 5.

    Model updates and maintenance.

    The model will be updated periodically to incorporate new data or significant medical advancements, ensuring its accuracy and relevance by integrating the latest research and accounting for demographic changes. Updates occur annually, or when substantial new evidence emerges. Each update will undergo validation to confirm its efficacy with comprehensive documentation provided for modification and justification. Feedback from clinical practitioners could also be integrated to enhance the practical utility of the model.

Limitations

A multidisciplinary team of anaesthesiology, orthopaedic surgery, and radiology experts will use bidirectional data for model development and internal and external data validation to reduce bias and ensure study quality. Single-centre data collection may limit the generalisability of the model beyond elderly patients with hip fractures as regional characteristics and institution-specific practices could affect its applicability. The relevance of the model to different age groups and clinical conditions remains unclear. Future studies will verify the applicability of this model to multicentre data and explore the differences between the models of patients of different ages.

Ethics and dissemination

The study “Geriatric Hip Fracture Patients: Constructing and Validating a Predictive Model for Perioperative Pulmonary Complications” had been approved by the Institutional Review Board (IRB) of Fengdu People’s Hospital, Chongqing, China (Ethical Approval ID: 2023SC0915-149). The study was registered in the Chinese Clinical Trial Registry (ChiCTR.org.cn) as ChiCTR2300071115 on 5 May 2023.

This study will strictly follow the ethical guidelines, national regulations, and data protection protocols to ensure patient confidentiality and privacy. The requirement for informed consent was waived for retrospective data. In the prospective phase, written informed consent will be obtained from all participants in line with IRB approval. The study results will be published in peer-reviewed journals and presented at scientific conferences until December 2025 with the aim of enhancing the knowledge of POPCs in geriatric hip fracture patients.

Discussion

Recent studies have demonstrated that POPCs not only prolong hospital stay, but also significantly increase postoperative mortality, particularly among elderly patients with hip fractures42,43. Therefore, the timely identification of high-risk individuals is crucial to prevent adverse outcomes and improve postoperative recovery.

This study aimed to develop a comprehensive predictive risk model for POPCs in elderly patients undergoing hip fracture surgery. By incorporating risk factors from the preoperative, intraoperative, and postoperative stages, the model aimed to provide a holistic risk assessment, thereby improving the precision of POPCs prediction across the entire perioperative period.

The model will be constructed and validated using retrospective and prospective data sets. These datasets enable the identification of key risk factors across all stages of the perioperative process. A logistic regression-based approach will be employed to develop the risk model, and the results will be visually represented via a nomogram model. Internal and external data validation ensures the reliability of the model.

Despite its strengths, this model has several limitations that must be acknowledged. This included potential confounders and the single-centre nature of the study, which may have affected the generalisability of the findings. Future studies should focus on validating this model across diverse clinical settings and patient populations, to ensure its relevance and applicability. Furthermore, the incorporation of machine learning techniques and large-scale datasets could further increase the predictive accuracy and clinical utility of the model.

To the best of our knowledge, this study is one of the few to explore the effects of perioperative multistage risk factors on POPCs in elderly patients undergoing hip fracture surgery. The results of this study will have potential implications for the clinical practice of elderly patients with hip fractures.