Development and Internal Validation of a Nomogram for Predicting Early Muscular Calf Venous Thrombosis Following Hip Fracture

Summary

This retrospective study developed and internally validated a six-variable nomogram to estimate the risk of early muscular calf vein thrombosis within 48 h of hip fracture. The model accounts for nonlinear triglyceride effects and may support risk awareness pending external validation.

Abstract

Early muscular calf vein thrombosis (MCVT) detected within 48 h after injury may affect perioperative management in patients with hip fractures. This study aimed to develop and internally validate a nomogram for estimating early MCVT risk in this population. This study retrospectively screened hip fracture patients admitted between 2021 and 2024. Patients were categorized into MCVT and non-MCVT groups based on an admission ultrasound performed within 48 h after injury. Least Absolute Shrinkage and Selection Operator (LASSO) regression was used for feature selection, and restricted cubic splines (RCS) were used to evaluate nonlinear associations. A multivariable logistic regression model was constructed and visualized as a nomogram. Model performance was assessed by discrimination (area under the receiver operating characteristic curve, AUC), calibration (Brier score), and clinical utility (decision curve analysis, DCA), with internal validation using 1,000 bootstrap resamples, following Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis Plus Artificial Intelligence (TRIPOD + AI) guidance. Among 665 patients included, 44 (6.6%) had early MCVT. Six variables were retained in the final prediction model: time from injury to admission (TFIA), triglycerides (TG), total cholesterol (TC), activated partial thromboplastin time (APTT), neutrophil-to-lymphocyte ratio (NLR), and platelet-to-lymphocyte ratio (PLR). RCS analysis revealed a significant nonlinear association between TG and MCVT risk (p-value for non-linearity <0.001). The model showed good discrimination (AUC: 0.894; 95% CI: 0.829–0.941) and acceptable calibration (Brier score: 0.041). The optimism-corrected C-index was 0.884. DCA suggested potential clinical net benefit across relevant threshold probabilities. In conclusion, the developed nomogram may assist early risk stratification, but external validation in independent cohorts is required before clinical application.

Introduction

Hip fractures in older adults represent a significant global public health challenge and are frequently associated with high morbidity, mortality, and substantial healthcare expenditure¹. Current clinical guidelines strongly advocate accelerated surgical intervention, ideally within 48 h of admission, to alleviate pain, facilitate early mobilization, and mitigate life-threatening complications^2,3. However, preoperative deep vein thrombosis (DVT) remains a formidable barrier to timely surgery. Muscular calf vein thrombosis (MCVT), a specific subtype of distal DVT, involves thrombus formation within the venous plexuses of the calf muscles, primarily the soleus and gastrocnemius muscles⁴. Although often asymptomatic and historically perceived as less clinically significant than proximal DVT, emerging evidence suggests that MCVT can propagate cephalad into proximal veins or trigger pulmonary embolism (PE), thereby compromising patient safety^5,6. Consequently, MCVT detected early after injury may necessitate surgical postponement, prolong hospital stays, and complicate perioperative management by requiring anticoagulant bridging or inferior vena cava filter placement. Therefore, a robust tool to identify hip fracture patients at high risk for early MCVT, as detected on admission ultrasound, is clinically important for optimizing surgical pathways and improving patient outcomes.

Despite routine ultrasound screening, early risk stratification for acute MCVT in hip fracture patients remains clinically challenging. Previous investigations have attempted to identify risk factors for perioperative thrombosis; however, most have focused on isolated demographic characteristics or traditional coagulation parameters^4,7. These studies are often constrained by the inherent limitations of linear regression models, which may fail to capture the intricate, non-linear dynamics and the complex interplay of the trauma-induced thrombo-inflammatory response. From a pathophysiological standpoint, acute fractures trigger a massive release of inflammatory cytokines, which subsequently activate the coagulation cascade^8,9. While novel composite inflammatory indices—such as the neutrophil-to-lymphocyte ratio (NLR) and platelet-to-lymphocyte ratio (PLR)—have demonstrated significant prognostic value in cardiovascular and surgical domains, their utility in predicting early-phase MCVT remains largely underexplored^10,11.

To bridge this research gap, a multidimensional and individualized predictive tool is needed. This study aimed to investigate predictors of early MCVT in a well-defined cohort of hip fracture patients admitted within 48 h of injury. By employing Least Absolute Shrinkage and Selection Operator (LASSO) regression for feature selection and restricted cubic splines (RCS) to characterize complex nonlinear associations, this study sought to develop and internally validate a clinically interpretable nomogram. This study hypothesized that integrating systemic inflammatory markers and metabolic profiles would improve early risk stratification and support surveillance planning during the acute phase of hip fracture care.

Protocol

The study protocol was approved by the hospital's Institutional Review Board (IRB) ([2025B], IIT Ethics Approval No. 0407).

Study design and participants

This single-center retrospective cohort study was conducted at the First Affiliated Hospital, Zhejiang University School of Medicine. This study consecutively screened the clinical records of patients admitted with hip fractures between January 2021 and December 2024. As this was a retrospective study, a waiver of informed consent from study participants was requested and granted. Based on findings from routine color Doppler ultrasound performed at emergency admission, patients were categorized into the MCVT or non-MCVT groups.

Hip fractures were defined as fractures of the proximal femur, including the femoral neck, intertrochanteric, and subtrochanteric regions, while excluding pelvic fractures. To ensure the purity of the study population, the inclusion criteria were strictly defined as follows: (1) primary closed hip fracture confirmed by X-ray or computed tomography; (2) time from injury to admission (TFIA) within 48 h; (3) completion of bilateral lower extremity deep venous color Doppler ultrasound upon admission (generally within 2 h after radiographic confirmation of hip fracture and before initiation of anticoagulant therapy); and (4) planned surgical intervention. Patients were excluded based on the following criteria: (1) multiple traumas (e.g., concurrent craniocerebral, thoracic, or abdominal injuries); (2) pathological fractures (e.g., bone metastasis); (3) a history of DVT or PE, or anticoagulant/antiplatelet therapy ≥ 30 days prior to admission; (4) severe hepatic or renal dysfunction significantly impairing coagulation factor synthesis or metabolism; (5) concurrent malignant tumors; (6) substantial missing data regarding primary clinical or laboratory variables; and (7) proximal DVT detected on admission ultrasound (e.g., thrombosis involving the popliteal, femoral, common femoral, or iliac veins).

Data collection and candidate predictors

All candidate predictors were retrospectively extracted from the institutional electronic medical record (EMR) system. To ensure data accuracy and integrity, the extraction process was independently performed by two researchers, with any discrepancies resolved through consensus or consultation with a senior physician. The collected data were categorized into the following domains: (1) Demographic Characteristics: Including age, gender, and body mass index (BMI). (2) Clinical and Injury-related Variables: The TFIA (measured in hours) was precisely documented. Additionally, baseline comorbidities were quantified using the Charlson Comorbidity Index (CCI). (3) Laboratory Biomarkers: Laboratory parameters were obtained from the initial venous blood samples collected within 24 h of admission. These encompassed the complete blood count (CBC), liver and renal function tests, and coagulation profiles. Additionally, composite systemic inflammatory indices, specifically the NLR and PLR, were calculated from the admission CBC results.

Outcome definition and diagnostic criteria

The primary endpoint of this study was the occurrence of early MCVT, defined as thrombosis detected during the initial assessment upon admission. All included patients underwent bilateral lower-extremity Doppler ultrasound on admission, prior to initiation of any anticoagulant therapy. The examination was conducted using EPIQ-5 with eL18-4. Patients were examined in the supine and/or lateral position as tolerated. Diagnosis was established via routine bilateral lower extremity color Doppler ultrasound. According to established sonographic criteria for the acute phase, a positive diagnosis of MCVT was confirmed based on the following manifestations¹²: (1) significant dilation and tortuosity of the calf muscular venous lumens; (2) the presence of intraluminal flocculent or uniform hypoechoic echoes, appearing as multiple circular or oval masses on transverse sections; (3) complete non-compressibility of the vein under transducer pressure; (4) absence of spontaneous or phasic blood flow signals, with no detectable flow even during distal limb augmentation (manual compression); and (5) a distinct demarcation between the thrombosed segment and the adjacent muscular tissue. Bilateral lower-extremity Doppler ultrasound was performed upon admission by certified ultrasound physicians using the standardized institutional protocol. Ambiguous cases were reviewed by a senior ultrasound physician.

Statistical analysis

All statistical analyses were performed using R software (version 4.5.2). The normality of continuous variables was assessed using the Shapiro-Wilk test. Continuous data were presented as means ± standard deviations (SD) or medians with interquartile ranges (IQR), as appropriate. Categorical variables were expressed as frequencies and percentages.

Variables with a p-value < 0.01 in univariable analysis and variables considered clinically important were entered into LASSO regression. Candidate predictors were selected based on clinical relevance and availability at admission. Continuous variables were standardized before LASSO regression. Eleven candidate features were evaluated in the LASSO procedure, comprising nine continuous variables and two dummy variables for fracture type. LASSO logistic regression with 10-fold cross-validation was used to reduce dimensionality and screen variables for the final multivariable model. The lambda.min and lambda.1se values were examined, and final variable retention was determined using penalized selection together with clinical relevance and subsequent RCS assessment. Associations between predictors and MCVT were quantified using odds ratios (ORs) and corresponding 95% confidence intervals (CIs).

Based on the predictors identified in the multivariable logistic regression, a clinical nomogram was constructed to facilitate the individualized estimation of early MCVT risk. The predictive performance of the nomogram was comprehensively evaluated across three dimensions: discrimination, calibration, and clinical utility. Discrimination was quantified using the area under the receiver operating characteristic curve (AUC). Calibration was assessed using calibration plots with 1,000 bootstrap samples to evaluate agreement between predicted probabilities and observed outcomes; the Brier score was also calculated to measure overall predictive accuracy. Finally, Decision Curve Analysis (DCA) was performed to determine the model's clinical net benefit across a wide range of threshold probabilities. All statistical analyses were two-sided, and p-values < 0.05 were considered statistically significant.

Consistent with the retrospective design, the sample size was determined by the availability of consecutive patient data over the four-year study period, rather than an a priori power calculation. Given the relatively low incidence of positive MCVT events (n = 44), the entire dataset was utilized for model development without partitioning into separate training and validation sets to maintain statistical power. To mitigate the risk of overfitting associated with this event rate, LASSO regression was employed for feature selection and shrinkage regularization. The final model included six predictors but seven regression parameters, excluding the intercept, because TG was represented by two spline terms; therefore, the parameter-based EPV was approximately 6.3. The model did not meet the traditional events per candidate (EPV) ≥10 heuristic, which is now regarded as a rough guideline rather than an absolute criterion. Therefore, this study additionally evaluated overfitting using bootstrap-corrected performance estimates and calibration slope. Internal validation was performed using 1,000 bootstrap resamples to estimate and adjust for optimistic bias. This study was conducted in accordance with the TRIPOD+AI reporting guidance¹³.

To facilitate independent validation and implementation, the complete model equation, including the intercept, regression coefficients, triglycerides (TG) spline coefficients, and RCS knot locations, is provided in Table 1. The predicted probability of early MCVT was calculated as P = 1 / [1 + exp(-LP)], where LP is the linear predictor.

Results

Demographic data

A total of 665 patients with hip fractures met the inclusion criteria and were enrolled in this study (Figure 1). Among them, 44 patients (6.6%) were diagnosed with early MCVT detected on admission ultrasound within 48 h after injury. A detailed comparison of the demographic and clinical characteristics between the MCVT and non-MCVT groups is summarized in Table 2.

LASSO regression and RCS outcomes

Candidate variables identified as statistically significant (p-value < 0.01) in univariate analysis, along with clinically relevant parameters, were further screened using LASSO regression (Figure 2). Eleven candidate features were evaluated, including nine continuous variables and two fracture-type dummy variables. The final prediction model retained TFIA, total cholesterol (TC), Activated Partial Thromboplastin Time (APTT), NLR, PLR, and TG. PLR showed borderline statistical significance in the multivariable model (p-value = 0.068) but was retained based on LASSO screening and clinical plausibility. To relax the linearity assumption, RCS analysis was performed for the six continuous predictors, adjusting for the other identified variables (Table 3). A significant nonlinear dose-response relationship was observed between TG and the risk of MCVT (overall p-value < 0.001). As illustrated in Figure 3, MCVT risk increased steeply across lower-to-middle TG values and then rose more gradually, with wider uncertainty at higher TG levels. RCS analysis was performed using 3 knots. For TG, knots were placed at 0.62 mmol/L, 1.03 mmol/L, and 1.64 mmol/L, corresponding to the 10th, 50th, and 90th percentiles of TG distribution. The apparent reference point around 1.03-1.04 mmol/L should be considered exploratory rather than a validated clinical cutoff.

Multivariable logistic regression analysis

Incorporating the variables identified through LASSO and RCS analyses, a final multivariable logistic regression model was constructed. In this model, TFIA, TC, APTT, NLR, and PLR were included as linear terms, while TG was incorporated as a restricted cubic spline to account for its nonlinear effect (Table 4). Six variables were retained in the final prediction model: TFIA (OR = 1.026, 95% CI: 1.004–1.049, p-value = 0.021), TC (OR = 0.244, 95% CI: 0.146–0.406, p-value < 0.001), APTT (OR = 0.856, 95% CI: 0.764–0.959, p-value = 0.007), NLR (OR = 1.189, 95% CI: 1.103–1.280, p-value <0.001), PLR (OR = 1.004, 95% CI: 1.000–1.008, p-value = 0.068), and TG modeled with RCS. The overall effect of TG in the model was highly significant (p-value < 0.001). Based on these variables, a clinical nomogram was developed to facilitate individualized risk estimation (Figure 4). The full regression coefficients, model intercept, TG spline coefficients, and knot locations are provided in Table 1. Predictor definitions are summarized in Table 5.

Model performance

The model's discriminative performance was evaluated using receiver operating characteristic (ROC) analysis (Figure 5), yielding an AUC of 0.894 (95% CI: 0.829–0.941). The calibration curve showed overall agreement between predicted probabilities and observed outcomes (Figure 6), supported by a Brier score of 0.041 (bootstrap 95% CI: 0.031–0.053). DCA suggested that using the model may provide higher net benefit than both the "treat-all" and "treat-none" strategies across a range of threshold probabilities (Figure 7). Additional calibration metrics were calculated. The apparent calibration intercept was approximately 0, the apparent calibration slope was 1.000, and the calibration-in-the-large was approximately 0. The bootstrap-corrected calibration slope was 0.896, suggesting modest optimism after internal validation. Internal validation with 1,000 bootstrap resamples showed limited optimism, with an optimism-corrected C-index of 0.884 (bootstrap 95% CI: 0.837-0.946) compared with the apparent AUC of 0.894. The bias-corrected calibration curve approximated the ideal line, with a mean absolute error of 0.013. Bootstrap 95% confidence intervals for DCA net benefit estimates are shown in Figure 7. The Hosmer-Lemeshow goodness-of-fit test showed χ² = 21.77, df = 8, p-value = 0.005, suggesting some discrepancy between predicted and observed probabilities across risk deciles. Therefore, calibration should be interpreted cautiously and reassessed in external cohorts.

Data Availability:

The de-identified individual participant dataset, the completed TRIPOD+AI checklist, and the full R statistical analysis code used for data processing, model development, validation, and performance evaluation are provided as Supplementary Files 1–3. These materials contain all data and analytical resources necessary to reproduce the findings reported in this study.

Figure 1: Participant flow diagram. This flow diagram summarizes patient screening, inclusion and exclusion criteria, and the final study cohort used for model development. Please click here to view a larger version of this figure.

Figure 2: LASSO regression for predictor selection. The plot shows cross-validated binomial deviance as a function of log(lambda) in 10-fold cross-validation. Eleven candidate features were evaluated, comprising nine continuous variables and two fracture-type dummy variables. The top numbers indicate the number of nonzero coefficients at each lambda value. Error bars indicate standard errors. The vertical dashed lines represent lambda.min (0.00526) and lambda.1se (0.02558). Abbreviations: CCI = Charlson Comorbidity Index; TFIA = Time from injury to admission; WBC = White blood cell; N = Neutrophil; L = Lymphocyte; M = Monocyte; HB = Hemoglobin; HCT = Hematocrit; PLT = Platelet; NLR = Neutrophil-to-lymphocyte ratio; PLR = Platelet-to-lymphocyte ratio; TP = Total protein; Alb = Albumin; Tbil = Total bilirubin; Crea = Creatinine; TC = Total cholesterol; GFR = Glomerular filtration rate; IP = Inorganic phosphate; INR = International normalized ratio; Fbg = Fibrinogen; APTT = Activated partial thromboplastin time; TT = Thrombin Time; PT = Prothrombin Time. Please click here to view a larger version of this figure.

Figure 3: RCS analysis of TG and early MCVT. The curve shows adjusted ORs for early MCVT according to TG level (mmol/L). The model was adjusted for TC, NLR, TFIA, APTT, and PLR and used 3 knots located at 0.62 mmol/L, 1.03 mmol/L, and 1.64 mmol/L. The reference value was the median TG level (1.03 mmol/L). The shaded area represents 95% CI. TG was modeled as a continuous nonlinear predictor, and the curve should not be interpreted as defining a validated clinical cutoff. Please click here to view a larger version of this figure.

Figure 4: Nomogram for predicting early MCVT detected within 48 h after injury. The nomogram estimates the predicted probability of early MCVT using TFIA (h), TG (mmol/L), TC (mmol/L), APTT (seconds), NLR, and PLR. Predictor definitions and measurement units are provided in Table 5. Please click here to view a larger version of this figure.

Figure 5: ROC curve of the final prediction model. The ROC curve shows the discrimination ability of the prediction model for early MCVT. The AUC was 0.894 (95% CI: 0.829-0.941). Please click here to view a larger version of this figure.

Figure 6: Calibration curve of the final prediction model. The calibration plot compares predicted and observed probabilities of early MCVT. The ideal line represents perfect calibration; the apparent curve represents model performance on the development dataset; and the bias-corrected curve represents bootstrap-corrected calibration after 1,000 resamples. The Brier score was 0.041, the apparent calibration intercept was approximately 0, the apparent calibration slope was 1.000, and the bootstrap-corrected calibration slope was 0.896. Please click here to view a larger version of this figure.

Figure 7: DCA of the final prediction model. DCA shows the net benefit of the nomogram across threshold probabilities. The model curve represents the net benefit of using the nomogram; the treat-all line assumes that all patients develop MCVT; and the treat-none line assumes that no patients develop MCVT. The curve should be interpreted as evidence of potential clinical utility only during internal validation, because no validated treatment threshold was established. Please click here to view a larger version of this figure.

Predictor/term	Transformation / definition	Coefficient beta
Intercept	Model intercept	-1.621145541
TFIA	Linear; hours from injury to admission	0.025977241
NLR	Linear; neutrophil-to-lymphocyte ratio	0.172703853
PLR	Linear; platelet-to-lymphocyte ratio	0.004034446
TG_linear	RCS linear term; TG in mmol/L	6.24150855
TG_rcs1	RCS nonlinear term; knots at 0.62, 1.03, 1.64 mmol/L	-4.814103739
TC	Linear; total cholesterol in mmol/L	-1.411933879
APTT	Linear; seconds	-0.155260603
TG RCS knots	0.62, 1.03, 1.64 mmol/L
Predicted probability	P = 1 / (1 + exp(-LP))

TABLE 1: Full specification of the prediction model. This table presents the complete prediction model, including the intercept, regression coefficients, restricted cubic spline coefficients for TG, and knot locations required for model implementation and independent validation.

Variable	MCVT group	Non-MCVT group	p value
Mean ± SD or n (%)	(n = 44)	(n = 621)
Gender (male/female)	15/29 (34.1/65.9)	224/397 (36.1/63.9)	0.872a
Age	79.68 ± 12.03	75.66 ± 15.65	0.041b
BMI	22.05 ± 2.74	21.94 ± 3.12	0.796b
Hypertension (yes/no)	15/29 (34.1/65.9)	196/425 (31.6/68.4)	0.739a
Diabetes (yes/no)	6/38 (13.6/86.4)	91/530 (14.7/85.3)	1.000a
Coronary disease (yes/no)	1/43 (2.3/97.3)	29/592 (4.7/95.3)	0.713a
Fracture type (femoral neck/intertrochanteric/subtrochanteric)	22/19/3 (50.0/43.2/6.8)	462/153/6 (74.4/24.6/1.0)	<0.001c
CCI	2.55 ± 2.12	2.31 ± 2.09	0.472b
TFIA (hours)	25.58 ± 18.89	16.22 ± 14.60	0.002b
WBC count (x 10⁹/L)	9.19 ± 2.92	8.90 ± 3.18	0.529b
N count (*10⁹/L)	7.48 ± 2.99	4.20 ± 3.11	<0.001b
L count (x 10⁹/L)	0.99 (0.78-1.21)	1.15 (0.92-1.52)	0.003d
M count (*10⁹/L)	0.57 (0.44-0.71)	0.59 (0.43-0.75)	0.941d
HB count (g/L)	105.13 ± 22.33	110.02 ± 21.16	0.165b
HCT	32.07 ± 6.46	33.62 ± 6.09	0.127b
PLT count (x 10⁹/L)	195.06 ± 69.21	165.82 ± 65.07	0.009b
NLR	6.16 (4.30-12.05)	3.21 (1.56-5.20)	<0.001d
PLR	181.51 (126.19-260.71)	131.76 (96.00-176.23)	<0.001d
TP (g/L)	61.79 ± 7.25	62.33 ± 7.31	0.637b
Alb (g/L)	35.77 ± 5.90	36.58 ± 4.91	0.381b
Tbil (μmol/L)	11.20 (9.23-14.38)	12.70 (9.00-16.70)	0.117d
Crea (μmol/L)	70.50 (58.80-89.50)	67.00 (57.47-82.00)	0.280d
Urea (mmol/L)	6.57 (5.28-9.15)	6.18 (4.92-8.14)	0.121d
TG (mmol/L)	1.24 (1.03-1.62)	1.01 (0.77-1.27)	<0.001d
TC (mmol/L)	3.11 ± 0.76	3.84 ± 0.95	<0.001b
GFR (ml/min)	70.1 ± 25.32	78.04 ± 24.45	0.050b
IP (mmol/L)	1.04 ± 0.24	1.06 ± 0.24	0.787b
INR	1.04 (0.99-1.07)	1.05 (1.01-1.10)	0.094d
Fbg (g/L)	3.7 ± 1.01	4.04 ± 1.06	0.034b
APTT (second)	28.16 ± 3.23	29.48 ± 3.88	0.012b
TT (second)	15.95 (15.30-16.95)	16.10 (15.60-16.80)	0.533d
PT (second)	11.90 (11.40-12.50)	12.27 (11.70-12.80)	0.064d
D-Dimer (μg/L)	7186.00 (4,447.25-17,346.75)	4780.90 (1,980.00-12,140.00)	0.017d
Data are presented as mean ± SD, median (IQR), or n (%), as appropriate. aWelch's t-test; bWilcoxon rank-sum test; cChi-square test; dFiisher's exact test (Fisher-Freeman-Halton exact test was used for multi-category variables when expected cell counts were small).

TABLE 2: Baseline characteristics of patients with and without MCVT. Continuous variables are presented as mean ± standard deviation or median with interquartile range, as appropriate. Categorical variables are presented as numbers and percentages. The final analytic cohort contained no missing values for candidate predictors or outcome variables; 12 screened patients were excluded because of missing key clinical or laboratory data.

Variable	Knots (10th, 50th, 90th percentiles)	Nonlinear p-value	Overall p-value
TC	2.61, 3.77, 4.92	0.369	<0.001
NLR	0.708, 3.35, 8.68	0.058	<0.001
TG	0.62, 1.03, 1.64	<0.001	<0.001
TFIA	2, 12, 48	0.327	0.072
APTT	25.5, 28.9, 33.9	0.554	0.003
PLR	74, 134, 250	0.923	0.163

TABLE 3: Nonlinear associations assessed by RCS analysis. The table presents knot locations, overall p-values, and nonlinear p-values for selected predictors. RCS models used 3 knots placed at the 10th, 50th, and 90th percentiles of each predictor distribution.

Variables	OR	CI-lower	CI-upper	p-value
TFIA	1.026	1.004	1.049	0.021
TC	0.244	0.146	0.406	<0.001
APTT	0.856	0.764	0.959	0.007
NLR	1.189	1.103	1.28	<0.001
PLR	1.004	1	1.008	0.068
TG (modeled with RCS)	-	-	-	<0.001

TABLE 4: Multivariable logistic regression model for early MCVT. ORs, 95% CIs, and p-values are shown for linear terms. Because TG was modeled using RCS, a single OR and 95% CI were not reported. The adjusted dose-response relationship is shown in Figure 3, and the spline coefficients and knot locations are provided in Table 1.

Predictor	Definition	Unit	Timing
TFIA	Time from injury to admission	hours	Admission blood test
TG	Triglycerides	mmol/L	Admission blood test
TC	Total cholesterol	mmol/L	Admission blood test
APTT	Activated partial thromboplastin time	seconds	Admission blood test
NLR	Neutrophil count / lymphocyte count	ratio	Admission blood test
PLR	Platelet count / lymphocyte count	ratio	Admission blood test

TABLE 5: Definitions of prediction model variables. This table summarizes the predictors included in the final model, along with their definitions, measurement units, and, where applicable, clinical interpretations.

Supplementary File 1: Completed TRIPOD+AI reporting checklist. This file contains the completed Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis Plus Artificial Intelligence (TRIPOD+AI) checklist, documenting compliance with the reporting recommendations for prediction model studies. Please click here to download this file.

Supplementary File 2: Statistical analysis outputs for prediction model development and validation. This workbook contains the statistical analysis outputs generated during model development and validation, including univariate analyses, LASSO regression results, restricted cubic spline analyses, multivariable logistic regression coefficients, model performance metrics, calibration analyses, Hosmer-Lemeshow goodness-of-fit test results, decision curve analysis, and related summary tables. Please click here to download this file.

Supplementary File 3: R script for data processing and statistical analysis. This file contains the complete R script used for data preprocessing, variable selection, model development, internal validation, calibration, discrimination analysis, decision curve analysis, and generation of the statistical outputs reported in this study.Please click here to download this file.

Discussion

Following a hip fracture, restricted lower limb mobility can reduce gastrocnemius pump function and venous blood flow velocity. Due to their relatively small diameters and paucity of valves, the muscular calf veins are susceptible to blood stasis and may become an early site of thrombus detection. A previous study by Zhao et al. reported a preoperative MCVT incidence of 23.5% in hip fracture patients¹⁴. In contrast, the cohort incidence was lower at 6.6%. This discrepancy may be attributable to stringent inclusion and exclusion criteria designed to focus on the acute post-traumatic window. Specifically, this study excluded patients with confounding pre-trauma factors, such as a prior venous thromboembolism or long-term use of anticoagulant and antiplatelet medications, thereby providing a more focused assessment of MCVT detected early after injury.

This study identified a significant synergistic interplay between TFIA, systemic inflammatory indices (NLR, PLR), and lipid metabolism (TG, TC) in predicting early MCVT. Prolonged TFIA not only exacerbates the physical risk of venous stasis due to immobilization but also extends the acute stress window during the critical post-traumatic phase^15,16. Univariate analysis revealed a bidirectional alteration characterized by elevated N counts and a depleted L count in the MCVT group, collectively driving the significant increase in NLR. The recruitment of N count reflects the activation of the acute inflammatory cascade following trauma; specifically, the release of neutrophil extracellular traps (NETs) has been shown to provide a prothrombotic scaffold that directly facilitates fibrin deposition and clot formation¹⁷. Conversely, the observed lymphopenia under continuous traumatic stress underscores the intricate interplay between systemic inflammatory response syndrome (SIRS) and compensatory anti-inflammatory response syndrome (CARS). This phenomenon is primarily driven by trauma-induced activation of the hypothalamic-pituitary-adrenal (HPA) axis, in which a surge in endogenous glucocorticoids accelerates L cell apoptosis. By integrating both "inflammatory intensity" and "immune dysfunction," the NLR was robustly selected by LASSO regression as a potent independent predictor (OR = 1.189, p-value < 0.001). Furthermore, the inclusion of the PLR in the model reflects the synergistic imbalance between platelet hyperreactivity and lymphopenia under stress. Although PLR did not reach conventional statistical significance in the multivariable model, it was retained as a potential predictive variable because it was selected by LASSO regression and reflects platelet-related inflammatory activity. Therefore, PLR should be interpreted as a component of the prediction model rather than as an independently significant risk factor.

Furthermore, this thrombo-inflammatory process is exacerbated by trauma-induced dysregulation of lipid metabolism. The inverse association between TC and early MCVT deserves careful interpretation. Although elevated cholesterol is traditionally associated with arterial atherosclerotic disease, venous thrombosis after acute trauma involves distinct mechanisms, including inflammation, venous stasis, endothelial injury, and trauma-induced coagulation activation. In the present cohort, the MCVT group showed lower TC but higher TG levels, suggesting an acute-phase lipid response rather than a conventional chronic dyslipidemia pattern. Previous evidence indicates that inflammation and infection can reduce TC, LDL-C, and HDL-C levels while increasing TG levels, and that the magnitude of these lipid alterations may correlate with disease severity¹⁸. Therefore, lower TC may reflect greater systemic inflammatory stress, catabolic state, or frailty in elderly hip fracture patients, rather than directly causing thrombosis. This finding should be regarded as hypothesis-generating and requires external validation. The observed elevation in TG and the concomitant reduction in TC within the MCVT group characterize a distinct "metabolic remodeling" profile. Hypertriglyceridemia not only triggers systemic inflammatory cascades and oxidative stress but also stimulates endothelial cells to secrete procoagulant factors, thereby inducing profound endothelial dysfunction¹⁹. Conversely, the decline in TC is frequently correlated with a surge in pro-inflammatory cytokines such as Interleukin-6, signaling impaired cell membrane stability and accelerated metabolic exhaustion. The intricate interplay between systemic inflammation and metabolic derangement drives a critical pathological transition of the vascular endothelium from a "thrombo-resistant phenotype" to a "procoagulant phenotype." This endothelial phenotypic shift constitutes the fundamental biological substrate for early-phase MCVT formation following hip fractures²⁰. The apparent TG reference point around 1.03 mmol/L should be interpreted cautiously. Because this value was derived from the current dataset and may be influenced by spline specification, knot placement, and the distribution of TG values, it should not be considered a definitive clinical cutoff. Further sensitivity analyses and external validation are needed before any TG threshold can be recommended for clinical decision-making. Although Fibrinogen (Fbg) levels were lower in patients with MCVT in the univariable analysis, Fbg was not retained in the final predictive model. This suggests that its association with MCVT may be partly explained by other inflammatory, coagulation, and metabolic variables included in the model. While a reduction in Fbg might traditionally be interpreted as a bleeding diathesis, in the clinical context of acute orthopedic trauma, it more likely signifies the onset of consumption coagulopathy²¹. During the hyperacute phase following a hip fracture, the systemic release of procoagulant factors triggers the exogenous coagulation cascade. As the fundamental substrate for fibrin cross-linking, Fbg is consumed in large quantities as it is converted into insoluble fibrin polymers that deposit within damaged or stagnant venous lumens^22,23. This phenomenon aligns with previous research on trauma-induced coagulopathy (TIC), where a localized decrease in circulating Fbg serves as a surrogate marker for extensive focal thrombotic activity²¹. Furthermore, the slight shortening of the APTT observed in the model corroborates the pre-activation of the endogenous coagulation pathway, reflecting a systemic transition from homeostatic stability to a prothrombotic state. The concomitant shortening of APTT and depletion of Fbg establish a cohesive hypercoagulation-consumption logic: endogenous pathway activation accelerates fibrin generation, necessitating a compensatory decline in circulating Fbg reserves. These findings suggest that monitoring the dynamic downward trend of coagulation parameters may provide superior predictive value for MCVT than relying solely on conventional hypercoagulable thresholds during the early post-traumatic window.

In recent years, predictive models for venous thrombosis in patients with hip fractures have been explored. Jiang et al. developed an MCVT nomogram model using data from 388 elderly patients, achieving an AUC of 0.805²⁴. Similarly, Pan et al. identified gender, TFIA, ASA grade, C-reactive protein, and D-dimer as predictors in a cohort of 419 patients, with a model AUC of 0.794²⁵. Another recent nomogram sought to predict preoperative DVT by integrating five risk factors across the domains of venous stasis, coagulation, and immune-inflammation²⁶. While these models focused on coagulation parameters, demographics, and nutritional status, their predictive windows were often broadly defined as the preoperative period. In clinical practice, however, surgery may occur within 48 h of trauma or be significantly delayed due to comorbidities. This study specifically focused on MCVT detected within 48 h after injury, which may be more relevant to the acute decision-making window. The Caprini risk assessment scale is widely used for VTE screening²⁷, but its application in the emergency setting for elderly hip fracture patients can be challenging because it incorporates numerous factors that rely on detailed medical history. In contrast, the present nomogram uses routinely available admission variables. However, the model should be interpreted as an internally validated risk-estimation tool rather than a definitive guide for treatment. The RCS analysis suggested a nonlinear association between TG and MCVT risk, with a change in slope around the median TG level. This pattern should be regarded as exploratory because it may depend on spline specification, knot placement, and the distribution of TG values in the present dataset. External validation is needed before any TG threshold can be recommended for clinical decision-making.

Smoking history was excluded from the initial data collection primarily due to considerations of data reliability and specific population characteristics. Given the demographic profile of the cohort, which was predominantly elderly females with an exceedingly low prevalence of active smoking, this variable would lack sufficient statistical variance to contribute meaningfully to the predictive model. Consequently, to support the objectivity and clinical utility of the nomogram, this study prioritized standardized laboratory biomarkers obtained upon admission. These objective indicators may more accurately reflect the acute pathophysiological shift toward a prothrombotic state within the 48 h post-injury window compared with self-reported chronic lifestyle factors. Some clinically important variables, including age, sex, fracture type, comorbidities, D-dimer, and Fbg, were not retained in the final model. This does not necessarily indicate that these factors are clinically irrelevant. Rather, after penalized selection, they did not add sufficient independent predictive information beyond TFIA, TG, TC, APTT, NLR, and PLR in this dataset. In particular, D-dimer and Fbg may partly overlap with other coagulation and inflammation-related variables. Given the limited number of MCVT events, a parsimonious model was favored to reduce overfitting.

Regarding model construction, the limited number of events (n = 44) yielded a conventional predictor-based EPV of approximately 7.3 and a parameter-based EPV of approximately 6.3 when the TG spline terms were counted separately. Although the final model did not meet the traditional EPV ≥10 heuristic, this rule is not an absolute criterion. This study therefore used bootstrap internal validation to assess optimism and calibration. Nevertheless, the limited number of events may still increase the risk of overfitting, and external validation is warranted. First, rather than relying solely on traditional stepwise selection, LASSO regression was used to support variable screening and reduce model complexity. Second, prior methodological work suggests that the rule of ten can sometimes be relaxed in predictive modeling²⁸, provided that model performance and calibration are carefully evaluated. Third, the 1,000-repetition bootstrap internal validation showed limited optimism in discrimination and acceptable overall calibration metrics, although the significant Hosmer-Lemeshow test indicates that calibration should be interpreted cautiously and reassessed in external cohorts.

The model is intended for physicians and nursing professionals involved in acute hip-fracture care. Users should enter six routinely available admission variables with correct units; no statistical expertise is required once the nomogram or calculator is implemented, but clinical interpretation should be made by trained clinicians. To apply the model, clinicians should collect TFIA, TG, TC, APTT, NLR, and PLR at admission. The linear predictor is calculated using the full regression equation, and the predicted probability is obtained as P = 1/[1+exp(-LP)]. The complete model specification, including the intercept, regression coefficients, TG spline coefficients, and RCS knot locations, is provided in Table 1. The detailed predictor definitions are provided in Table 5. The predicted probability should be interpreted as an estimated risk only. No validated intervention threshold is currently available. For patients predicted to be at high risk, the nomogram may have a potential value in prompting heightened clinical vigilance rather than directly dictating treatment. Potential management strategies include early or repeated Doppler ultrasound assessment when clinically indicated, careful reassessment of both thrombotic and bleeding risks, avoidance of unnecessary delay to surgery, early mobilization, adequate hydration, and timely initiation of guideline-based VTE prophylaxis when not contraindicated. Because no validated risk categories or treatment thresholds were established in this study, the model should be regarded as a tool for risk awareness and surveillance planning rather than a stand-alone basis for anticoagulation decisions. Although the calibration curve and Brier score suggested acceptable overall calibration, the Hosmer-Lemeshow test was statistically significant. This finding may partly reflect the test's grouping sensitivity and the low event rate, but it also indicates that calibration should be interpreted cautiously and requires external validation.

Several limitations should be acknowledged. First, this was a single-center retrospective study with only bootstrap internal validation; therefore, the generalizability and clinical applicability of the nomogram require prospective multicenter external validation. Subgroup-specific performance and fairness could not be formally assessed because of the limited number of MCVT events. Second, because pre-injury ultrasound data were unavailable, the outcome should be interpreted as MCVT detected within 48 h after injury rather than as definitively new-onset thrombosis. Third, no validated risk categories or intervention thresholds were established; thus, the nomogram should support risk awareness, closer monitoring, and timely ultrasound evaluation, rather than serve as a stand-alone basis for anticoagulation decisions. Fourth, inter-observer reliability for ultrasound diagnosis was not formally assessed, which may have introduced diagnostic variability. Finally, the use of univariable screening before model development may have excluded clinically relevant variables, such as age, D-dimer, Fbg, fracture type, and comorbidities. Future larger, externally validated studies should address these limitations.

Conclusion:

This study developed and internally validated a nomogram incorporating six readily available biomarkers to estimate MCVT risk early in patients with hip fractures. The model demonstrated good discriminative performance, acceptable overall calibration following internal validation, and potential clinical net benefit in internal bootstrap validation. In addition, the non-linear association between TG levels and MCVT risk provides additional insight into risk modeling in this population. This internally validated model may assist in early risk stratification of patients with hip fractures. However, given the retrospective single-center design and the absence of external validation, the findings should be considered exploratory. Further multicenter, prospective studies are required to evaluate the model’s generalizability and determine its potential clinical utility before clinical implementation.

Materials

List of materials used in this article

Name	Company	Catalog Number	Comments
Color Doppler ultrasound system	Philips	EPIQ-5 with eL18-4	Detection and diagnosis of MCVT
Coagulation analyzer	Stago	STAR MAX (https://www.stago-cn.com)	APTT measurement
Hematology analyzer	Cobas	8000 (https://www.cobase.com/)	Blood cell counts, NLR, PLR
Laboratory biochemical analyzer	Sysmex	BC7500 (https://www.sysmex.com.cn/)	TG and TC measurement
Microsoft Excel	Microsoft	2023 (https://www.microsoft.com)	Data organization
R software	R Foundation	https://www.r-project.org/	Statistical analysis
R package glmnet	CRAN	https://cran.r-project.org/web/packages/glmnet/index.html	LASSO regression
R package rms	CRAN	https://cran.r-project.org/web/packages/rms/index.html	RCS, nomogram, calibration
R package pROC	CRAN	https://cran.r-project.org/web/packages/pROC/index.html	ROC analysis
R package rmda / dcurves	CRAN	https://cran.r-project.org/web/packages/dcurves/index.html	Decision curve analysis