Risk prediction models for re-fractures following hip fracture in older adults: A methodological evaluation

Abstract

BACKGROUND

Hip fracture in elderly patients is often termed “the last fracture in life“. The incidence of hip re-fractures is high, but prognosis is poor, significantly impairing patients’ quality of life and imposing a substantial burden on healthcare systems. A reliable prediction model for re-fractures could play a crucial role in guiding preventive strategies.

AIM

To conduct a critical appraisal of existing prediction models for re-fractures in hip fracture patients.

METHODS

A systematic search was conducted across five databases, PubMed, EMBASE, the Cochrane Library, Web of Science, and the China National Knowledge Infrastructure, from inception to January 1, 2025. Eligible studies included those that developed prediction models for postoperative re-fractures following hip fractures. The Prediction Model Risk of Bias Assessment Tool was employed to evaluate the risk of bias and clinical applicability. A narrative synthesis summarised the characteristics, methodological quality, and predictive performance of the identified studies. Additionally, a meta-analysis was performed to quantitatively assess the performance of prediction models.

RESULTS

Of the 6056 studies retrieved, 8 studies with 28 predictive models were ultimately identified. Internal and external validation was performed for five (62.5%, internal), two (25.0%, external), and one (12.5%, internal and external) models. The number of predictors per model ranged from four to nineteen. The most frequently included predictors were age, rehabilitation exercise, osteoporosis, heart disease, and Alzheimer's disease. The models demonstrated area under the curve values of 0.69-0.98 in internal validation and 0.76-0.98 in external validation. Pooled analysis of the area under the curves yielded values of 0.970 (95%CI: 0.960-0.980) and 0.932 (95%CI: 0.907-0.959) for the model development and validation, respectively. All included models had a high risk of bias, while only two (25.0%) showed low concerns regarding applicability.

CONCLUSION

Current risk prediction models for postoperative re-fracture after hip fractures surgery lack robust validation and comprehensive evaluation. Future studies should prioritise refining model development, improving generalizability, and assessing clinical utility. Collaborative initiatives involving researchers, clinicians, and policymakers are crucial to transforming these models into effective tools for mitigating the burden of re-fractures in elderly populations.

Key Words: Older adults; Prediction model; Hip re-fracture; Meta-analysis; Systematic review

Core Tip: This study systematically evaluates existing models designed to predict re-fractures after hip fracture in older adults. By critically examining their development, validation, and methodological rigor, the review highlights substantial limitations in bias control, generalizability, and clinical applicability. Despite showing moderate to high predictive performance, most models lack robust external validation. The findings underscore the urgent need for improved model design and collaborative efforts to create reliable, clinically useful tools that can better guide prevention and reduce the burden of re-fractures in elderly populations.

INTRODUCTION

Hip fractures represent a major public health concern, particularly among the elderly, due to their impact on morbidity, mortality, and healthcare costs[1]. With ageing populations, the incidence of hip fractures is projected to rise, increasing pressure on healthcare systems[2,3]. Re-fractures, which occur in 2%-20% of cases[4,5], are associated with compromised quality of life and further complications, including impaired gait, balance issues, and falls[6,7]. These re-fractures lead to extended hospital stays and greater financial burden[8]. Despite advancements in surgical techniques and postoperative care, re-fractures remain a significant clinical challenge. The identification of high-risk patients is critical for implementing targeted prevention strategies. Recent research has increasingly focused on constructing predictive models for re-fracture risk, integrating demographic profiles, bone mineral density measurements, comorbidity burdens, and operative variables to enhance risk stratification[9,10]. While these models can guide rehabilitation and preventive measures, there is considerable variability in their development and validation. A systematic evaluation of existing models is needed to assess their validity, limitations, and clinical applicability. This study aims to synthesise current evidence, evaluate model quality, support evidence-based clinical decision-making, and identify research gaps to guide future model development.

MATERIALS AND METHODS

This systematic review was conducted following the Checklist for Critical Appraisal and Data Extraction for Systematic Reviews of Prediction Modelling Studies checklist[11] and the Preferred Reporting Items for Systematic Reviews and Meta-Analyses guidelines[12]. The protocol was prospectively registered with the International Prospective Register of Systematic Reviews before commencement, ensuring transparency and minimizing potential bias in the review process (CRD: 42024540749).

Data sources

A comprehensive literature search was conducted across five major databases: PubMed, EMBASE, Cochrane Library, Web of Science, and China National Knowledge Infrastructure (CNKI), covering all available records from the databases’ inception to January 1, 2025. The search strategy incorporated a combination of Medical Subject Headings terms and free-text keywords related to three key concepts: (1) Hip fractures; (2) Re-fractures; and (3) Prediction models (Supplementary Tables 1-4). Data extraction was performed in duplicate, categorized into: (1) Study characteristics of the included studies [e.g., author, year, country, design (period), model design, age, incidence of re-fracture (%), sites of re-fracture]; and (2) Overview of the included prediction models (e.g., predictors, type of validation, performance, model presentation).

Study selection

Inclusion criteria: (1) Patients aged ≥ 60 years with radiologically confirmed hip re-fractures (femoral neck, intertrochanteric, or subtrochanteric); (2) The development and/or validation of predictive models for postoperative re-fractures; (3) The reporting of at least one performance measure (sensitivity or specificity); and (4) Study designs included prospective cohorts, retrospective cohorts, and case-control studies.

Exclusion criteria: (1) Studies focusing solely on risk factors without predictive modeling; (2) Non-original research (abstracts, reviews, editorials); (3) Animal studies; and (4) Studies with unavailable full-text.

Two independent reviewers (Qian M and Chen YX) conducted the study selection following a standardized approach using the Checklist for Critical Appraisal and Data Extraction for Systematic Reviews of Prediction Modelling Studies checklist for study appraisal: (1) Executing the electronic database search using predefined search terms; (2) Systematically removing duplicate records; (3) Screening titles and abstracts against eligibility criteria; (4) Performing full-text reviews of potentially relevant studies; and (5) Hand-searching the reference lists of included articles. Discrepancies in full-text eligibility were resolved through consensus or arbitration by a third-reviewer (Li GQ).

Quality assessment

Risk of bias was assessed independently by two investigators (Qian M and Chen YX) using the Prediction model Risk of Bias Assessment (PROBAST) tool[13]. The tool includes 20 questions across four domains: (1) Participant selection; (2) Predictor variables; (3) Outcome measures; and (4) Analytical approaches. Each question was rated as ’yes’, ’probably yes’, ’no’, ’probably no’, or ’no information available’. Studies were classified as high risk of bias if any domain showed high risk. Applicability was evaluated across participants, predictors, and outcomes to ensure alignment with research objectives[14]. Discrepancies in the assessments were resolved through discussions and arbitration by a senior researcher (Li GQ).

Statistical analysis

Key study characteristics, including country of origin, data sources, participant demographics, re-fracture incidence, and sites of re-fracture, were extracted and synthesized. Model performance was evaluated using the concordance index (C-index), with values ranging from 0.5 (no discrimination) to 1.0 (perfect discrimination). Meta-analyses were performed using Comprehensive Meta-Analysis software to assess re-fracture prediction in elderly populations, employing fixed-effect models for low heterogeneity (I² < 50%) and random-effects models for high heterogeneity (I² ≥ 50%). We also performed a sensitivity analysis by sequentially excluding each study and rerunning the meta-analysis to assess how the pooled effect size and heterogeneity were affected. All effect measures were reported as pooled area under the curve (AUC) with 95%CI, and statistical significance was set at P < 0.05.

RESULTS

Overall search results

Our systematic search initially identified 6056 records from indexed databases. After removing 1741 duplicates, we screened 4315 unique records based on titles and abstracts. Seventy-five full-text articles were assessed for eligibility, and 68 studies were excluded for the following reasons: (1) 43 did not address prediction models for hip re-fractures; (2) 2 were conference abstracts without full-text availability; and (3) 23 focused solely on risk factor analysis without predictive modeling. One additional relevant study was identified through manual searches. This process resulted in 8 studies meeting our inclusion criteria (Figure 1), encompassing 28 prediction models[9,10,15-20].

Figure 1 Flow diagram of the selection process. CNKI: China National Knowledge Infrastructure.

Studies characteristics

Our review identified 8 studies with 28 prediction models for hip re-fracture risk in elderly patients (Table 1)[9,10,15-20]. The reported re-fracture incidence varied from 6.4% to 21.2%. The studies included two prospective cohort designs, one case-control study, and five retrospective cohort analyses, with sample sizes ranging from 601 participants to 40357 participants. Studies originated from three countries: (1) Six from China; (2) One from Korea; and (3) One from Spain. All models incorporated clinical and demographic variables, although specific components varied.

Table 1 Study characteristics of the included studies, mean ± SD.

Ref.	Country	Design (period)	Model design	Age (years), re-fracture vs normal	Incidence of re-fracture (%)	Sites of re-fracture
Wen et al[19], 2024	China	Prospective cohort (2018-2020)	D + V	75.51 ± 6.7 vs 72.01 ± 5.89	51/601 (8.49)	Contralateral
Guo[20], 2022	China	Case-control (2014-2020)	D + V	77.4	130/1645 (7.9)	Contralateral
Wu et al[9], 2025	China	Prospective cohort (2018-2020)	D + V	72.11 ± 8.978 vs 68.80 ± 8.57	158/1350 (11.7)	Ipsilateral
Kim et al[15], 2024	Korea	Retrospective cohort (2004-2020)	D + V	77.9 ± 8.4 vs 74.6 ± 13.9	135/1480 (9.1)	Ipsilateral
Liang et al[16], 2023	China	Retrospective cohort (2016-2020)	D + V	82.23 ± 6.52 vs 79 (70, 83)	52/734 (7.08)	Contralateral
Huang et al[17], 2024	China	Retrospective cohort (2009-2020)	D + V	66.01 (61.74-70.22) vs 65.06 (61.73-70.25)	8553/40357 (21.2)	Ipsilateral
Larrainzar-Garijo et al[10], 2024	Spain	Retrospective cohort (2011-2019)	D + V	82.77 ± 6.31 vs 83.88 ± 6.99	124/1960 (6.4)	Contralateral
Wang et al[18], 2025	China	Retrospective cohort (2018-2023)	D + V	79	53/629 (8.4)	Ipsilateral

D: Development; V: Validation.

Overview of prediction models

Of the eight models, five (62.5%) underwent internal validation, two (25.0%) included external validation, and one (12.5%) underwent both internal and external validation. The models ranged from four to nineteen predictors, with the most commonly included being advanced age, rehabilitation exercise, osteoporosis, heart disease, and Alzheimer’s disease. Figure 2 illustrates the frequency distribution of predictors. Table 2 summarizes the overview of the prediction models[9,10,15-20]. Six studies used logistic regression for model development. The models demonstrated strong discriminative ability, with AUC values ranging from 0.69 to 0.98 of internal validation. External validation showed predictive accuracy, with AUC values between 0.76 and 0.98.

Figure 2  Frequency of predictors in the final model.

Table 2 Overview of the included prediction models.

Ref.	Cases/sample	Methods	Number of predictors	Predictors	Type of validation	Performance		Presentation
						Development	Validation
Wen et al[19], 2024	A: 35/421 (8.31%); B: 16/180 (8.89%)	LR	4	Age, female, OP, comorbidity	IV	Sensitivity: 0.826; specificity: 0.804; AUC: 0.876; C-index: 0.810; 95%CI: 0.711-0.912	Sensitivity: 0.788; specificity: 0.781; AUC: 0.830; C-index: 0.832; 95%CI: 0.720-0.928	Nomogram
Guo[20], 2022	A: 90/1097 (8.2%); B: 40/540 (7.3%)	LR	7	Age, Harris score, AD, PD, visual impairment, heart disease, exercise	IV	Sensitivity: 0.722; specificity: 0.736; AUC: 0.776; 95%CI: 0.727-0.825	Sensitivity: 0.750; specificity: 0.701; AUC: 0.815; 95%CI: 0.743-0.887	Nomogram
Wu et al[9], 2025	NR	LR	5	Age, DM, OP, exercise, preoperative total protein	EV	Precision: 0.906-0.926; accuracy: 0.806-0.877; AUC: 0.912-0.976	Precision: 0.901-0.921; accuracy: 0.818-0.903; AUC: 0.893-0.976	Weighted Ensemble-L2, XGBoost, NeuralNetTorch, LightGBM, CatBoost, LightGBMXT, RandomForestEntr, RandomForestGini, LightGBMLarge, NeuralNetFastAl
Kim et al[15], 2024	A: 113/1012 (11.17%); B: 22/468 (4.7%)	XGBoost Algorithm	3	Two-dimensional frontal, lateral, and axial digitally reconstructed radiographs	IV	C-index: 0.59-0.73; AUC: 0.55-0.77		Deep learning-based convolutional neural network model
Liang et al[16], 2023	A: 31/513 (6%); B: 21/221 (9.5%)	LR	8	Age, hemoglobin, heart disease, neurovascular disease, PD, AD, COPD, CKD	IV	AUC: 0.906; 95%CI: 0.845-0.967	AUC: 0.956; 95%CI: 0.927-0.985	Nomogram
Huang et al[17], 2024	NR	Machine learning	19	Age, male, height, weight, LOS, smoking, CKD, arteriosclerosis, epilepsy, DM, liver disease, dyslipidaemia, pisphosphonates, PD, raloxifene and alendronate intake, single, activities, incomes	IV + EV	Sensitivity: 0.83-0.95; specificity: 0.82-0.95; AUC: 0.91-0.98		Generalised linear models, random forests, stochastic gradient boosting machines, generalized additive models, support vector machines, Naive Bayes
Larrainzar-Garijo et al[10], 2024	NR	NR	16	Age, female, OP, dementia, heart disease, COPD, asthma, renal failure, hypothyroidism, stroke, malnutrition, visual deficit, anaemia, walking assistance, pertrochanteric fracture	IV	AUC: 0.69; C-index: 0.58		Fine and Gray sub-distribution hazard competing risk model
Wang et al[18], 2025	NR	LR	5	Harris score, sunshine time, exercise, AD, LOS	EV	Sensitivity: 0.670; specificity: 0.774; AUC: 0.778	C-index: 0.761; accuracy: 0.789	Nomogram

A: Re-fracture; AD: Alzheimer’s disease; AUC: Area under the curve; B: Normal; CKD: Chronic kidney disease; COPD: Chronic obstructive pulmonary disease; DM: Diabetes mellitus; EV: External validation; IV: Internal validation; LOS: Length of hospital stay; LR: Logistic regression; NR: Not report; OP: Osteoporosis; PD: Parkinson’s disease.

Quality assessment of prediction models

The methodological quality of the included studies was evaluated using the Prediction model Risk of Bias Assessment tool, focusing on risk of bias and applicability (Table 3)[9,10,15-20]. Six of the eight studies were found to have a high overall risk of bias, primarily due to issues with analytical methods and predictor selection. Applicability concerns were also identified in these studies, mostly due to participant selection criteria not aligning with our target population. Two studies[15,17] had either low or unclear risk of bias, with fewer applicability concerns.

Table 3 Prediction model Risk of Bias Assessment results of included studies.

Ref.	ROB				Applicability			Overall
	Participants	Predictors	Outcome	Analysis	Participants	Predictors	Outcome	ROB	Applicability
Wen et al[19], 2024	1	1	1	3	1	1	1	3	1
Guo[20], 2022	1	2	1	1	1	1	1	2	1
Wu et al[9], 2025	1	1	2	1	1	1	1	2	1
Kim et al[15], 2024	1	2	2	2	1	3	1	2	3
Liang et al[16], 2023	1	2	1	1	1	1	1	2	1
Huang et al[17], 2024	3	2	1	2	1	1	2	2	2
Larrainzar-Garijo et al[10], 2024	1	2	2	2	1	1	1	2	1
Wang et al[18], 2025	1	2	1	3	1	1	1	3	1

1: Low risk of bias (ROB) or low concern regarding applicability. 2: Indicates high ROB or high concern regarding applicability. 3: Indicates unclear ROB or unclear concern regarding applicability. ROB: Risk of bias.

Meta-analysis of prediction models

Due to the limited number of studies on hip re-fracture prediction models, only four model development studies and three model validation studies met the criteria for meta-analysis. The pooled AUC values were calculated using both random-effects and fixed-effects models, yielding 0.970 (95%CI: 0.960-0.980; Figure 3A)[15-17,20] and 0.932 (95%CI: 0.907-0.959; Figure 3B)[15,16,20] for the model development and validation, respectively. Heterogeneity was high in model development (I² = 80.35%) but low in model validation (I² = 20.11%). Furthermore, sensitivity analyses demonstrated the results remained consistently robust. These results indicate the high discriminative ability of the models but also highlight the considerable variability across studies, suggesting caution in their interpretation.

Figure 3 Forest plot. A: Forest plot of the random effects meta-analysis of pooled area under the curve estimates for 4 models-development; B: Forest plot of the fixed effects meta-analysis of pooled area under the curve estimates for 3 models-validation.

DISCUSSION

This systematic review critically evaluates existing prediction models for postoperative hip re-fracture risk in elderly patients, identifying key challenges such as methodological flaws in predictor selection, insufficient external validation, and variability in clinical applicability. Notwithstanding a strong pooled AUC of 0.970, significant heterogeneity (I² = 80.35%) and potential biases warrant caution in interpreting these models. The analysis highlights the need for standardized predictors, rigorous external validation, and greater reporting transparency to improve clinical utility. Moreover, the findings underscore the importance of dynamic risk prediction models that incorporate both modifiable and non-modifiable factors, as well as serial assessments to better tailor interventions and reduce re-fracture risk.

This review demonstrates that prediction models for hip re-fractures consistently achieve AUC values above 0.7, indicating strong predictive ability. Key predictors include non-modifiable factors (age > 85 years, female sex) and modifiable factors (diabetes, stage 3+ chronic kidney disease, uncorrected visual impairment), aligning with both domestic and international findings. The re-fracture risk peaks within the first postoperative year, with 60%-70% of fractures occurring during this period, supporting the “imminent fracture risk window“ concept from the International Osteoporosis Foundation[21]. This critical period requires interventions such as anti-osteoporosis therapy and fall prevention, with dynamic factors like mobility recovery and treatment adherence further influencing risk[22,23]. For patients meeting at least two of the high-risk criteria, immediate implementation of a Fracture Liaison Service model is recommended, involving bone health assessments, early anti-osteoporosis therapy, fall prevention, and regular monitoring. These findings are consistent with established evidence and emphasize the need for a multidisciplinary approach to reduce fracture risk[24,25]. Future studies should refine predictive tools and evaluate integrated interventions to enhance outcomes.

Our PROBAST-based assessment revealed significant methodological concerns in the included studies, despite their clinical applicability. Five studies[10,15-18] relied on retrospective data, introducing potential biases from incomplete documentation, unmeasured confounders, and non-standardized outcomes. Predictor selection was often based on univariate screening or clinician consensus, increasing the risk of overfitting and missing important interactions. Only 37.5% of the models underwent external validation[9,17,18], and none assessed clinical utility or cost-effectiveness. Additionally, none included time-varying predictors, despite their known impact on refracture risk. Future research should focus on developing dynamic algorithms that incorporate serial measurements of bone turnover markers, functional mobility scores, and treatment adherence. Models should be validated across diverse healthcare settings and ethnic groups, and the cost-effectiveness of Fracture Liaison Service interventions should be evaluated. Incorporating advanced imaging biomarkers, such as trabecular bone score and finite element analysis, is essential. These findings highlight the need for a shift from static risk assessment to dynamic, multidisciplinary approaches that consider the complex interplay of skeletal, metabolic, and functional factors.

Prediction models for postoperative hip re-fracture risk in elderly patients offer the potential for risk-stratified care, improving outcomes through personalized interventions. Early identification of high-risk individuals enables targeted approaches, such as intensified anti-osteoporosis therapy, structured fall prevention programs, and enhanced postoperative monitoring. Key considerations for clinical adoption include predictive accuracy (AUC > 0.8), reliability, and calibration to prevent over-treatment/under-treatment. Integration into clinical workflows, such as electronic health record-embedded decision support and mobile/web applications, is crucial. Provider education on model limitations and interpretability is necessary for effective use. Future directions include dynamic risk prediction updates based on post-discharge data and cost-effectiveness studies comparing model-guided care to standard practice. Successful implementation relies on combining high-performing models, seamless integration, and clinician engagement to improve patient outcomes.

This is the first systematic review and critical appraisal of prediction models for re-fractures in elderly patients with hip fractures, highlighting both the strengths and limitations. Key strengths include a comprehensive search strategy across multiple databases, adherence to Preferred Reporting Items for Systematic Reviews and Meta-Analyses guidelines for transparent reporting, and rigorous quality assessment using PROBAST. However, several limitations should be acknowledged. Given the language restrictions, geographic imbalance, and limited eligible evidence, the predominance of Chinese cohorts may reduce the generalizability of our results to other healthcare systems and ethnic groups. These constraints further weakened meta-analytic power. Methodological heterogeneity, including variability in predictor selection and modeling techniques, further hindered cross-study comparisons. Future research should expand language inclusion, prioritize multinational cohorts, and adopt standardized reporting. High risk of bias such as retrospective designs and predictor selection strategies, should be cautioned in interpreting the results. Large-scale prospective validations in diverse settings are necessary to improve model generalizability and clinical integration. While this review consolidates the evidence on re-fracture prediction models, its findings must be interpreted considering geographical and methodological constraints, with future efforts focusing on optimizing post-hip-fracture care globally.

CONCLUSION

This systematic review evaluated eight studies with 28 prediction models for postoperative hip re-fractures in elderly patients, finding moderate-to-high discriminative ability but notable methodological limitations, including high risk of bias, retrospective designs, and limited external validation. Applicability concerns were raised due to non-representative cohorts and the absence of dynamic predictors. Key barriers to clinical implementation include limited robustness, geographical bias, and the lack of time-varying factors such as post-operative mobility and treatment adherence. To move the field forward, future research should focus on prospective model development, external validation, and clinical integration through pragmatic trials and electronic health record-embedded tools. Multidisciplinary collaboration and policy advocacy for high-quality validation are essential. Despite promising discrimination, the real-world utility of these models remains limited, underscoring the need for a coordinated effort to enhance methodology, validation, and implementation to reduce re-fractures in ageing populations.