Skip to main content
NCBI home page
Bone & Joint Research logoLink to Bone & Joint Research
. 2026 May 13;15(5):508–518. doi: 10.1302/2046-3758.155.BJR-2025-0385.R1

MRI-based radiomics explainable model for predicting recurrence of limb chronic osteomyelitis in limb bones treated by Masquelet technique

Yi Cao 1,2, Xuesen Zhao 1, Ruofei Wang 1, Fu Hao 1, Zekun Zhang 3, Tao Zhang 4, Liqiang Li 3, Dong Ma 1,✉
PMCID: PMC13169451  PMID: 42125887

Abstract

Aims

This study aimed to develop a machine learning model for predicting chronic osteomyelitis recurrence (COR) following the Masquelet technique (MT). The model integrated MRI-based radiomics with clinical characteristics to guide the definitive treatment of bone defects in limb chronic osteomyelitis (LCO).

Methods

We retrospectively analyzed data from patients with chronic osteomyelitis who underwent debridement and the Masquelet technique (MT) as definitive treatment at two medical centres between 2015 and 2023. The dataset included demographics, MRI scans, clinical characteristics, and infection recurrence, with a two-year follow-up period. Radiomics features were extracted from MRI-images using PyRadiomics. Clinical features were identified by logistic regression analyses. A COR predictive model was developed using machine learning algorithms and SHapley Additive exPlanations (SHAP). Additionally, a web-based application was also constructed to support the model.

Results

Among 279 patients (mean age, 43.19 years (SD 11.87); 195 men and 84 women), 59 (21.15%) patients had COR, involving the hip, lower limb, foot, and upper limb. The radiomic feature “Radscore” was constructed by eight image features. Four significant clinical features (age, surgery times, duration of infection, ESR) were selected, and combined with radiomics feature to construct a predictive model using machine learning algorithms. This integrated model exhibited a superior performance (area under the curve (AUC) = 0.901, 0.952, and 0.910 in training, validation, and external validation cohort, respectively) than only clinical (AUC = 0.862) or radiomics (AUC = 0.684) model. Lastly, a web-based application was developed and validated to predict COR risk in patients with MT treatment.

Conclusion

A web-based application integrating radiomics and clinical factors was developed to predict risk of COR in patients after MT treatment, allowing for the implementation of preventive interventions and targeted management.

Cite this article: Bone Joint Res 2026;15(5):508–518.

Keywords: Chronic osteomyelitis, Radiomics, Recurrence, Magnetic resonance imaging, Machine learning

Keywords: chronic osteomyelitis, Masquelet technique, MRI scans, infections, logistic regression model, surgical debridement, limb bones, hip, bone defects, lower limb

Article focus

  • Masquelet technique is widely used in treatment of chronic osteomyelitis; however, their long-term outcomes and predictors of recurrence remain unclear.

  • This study aimed to develop an interpretable machine learning-based MRI radiomics model to predict the chronic osteomyelitis recurrence (COR) following Masquelet technique (MT) surgery. The model integrates clinical and MRI-based radiomics features within web-based application to effectively identify high-risk patients.

Key messages

  • The data demonstrate that patients with risk of COR following MT surgery can be predicted using a model incorporating clinical factors (e.g., hospitalization frequency, surgical history, infection duration) and radiomic features (extraction from MRI images).

  • The web-based application that hosts an explainable machine learning model can effectively predict the risk of COR after MT surgery, thereby facilitating preventive interventions and guiding targeted patient management.

Strengths and limitations

  • The strength of this study lies in the development of an explainable machine learning-based model incorporated clinical and radiological factors for predicting COR following MT surgery, the effectiveness of which was demonstrated at two representative hospitals.

  • This was a retrospective study, and not all potential risk factors could be fully assessed within the scope of this investigation, which may limit the generalizability of our findings to some extent the generalizability to a certain limitation.

Introduction

Limb chronic osteomyelitis (LCO), a severe bone infection typically caused by bacteria (e.g. Staphylococcus aureus) or fungi (e.g. Candida albicans), with a recurrence rate of around 20% to 30%, resulting in prolonged antibiotic therapy and repeated surgical interventions, represents a significant burden on healthcare systems worldwide.1 Despite Masquelet technique (MT) notabaly improving LCO treatment in patients with chronic osteomyelitis during the last few decades, the recurrence remains challenge.2 Inability to predict chronic osteomyelitis recurrence (COR) increases patient morbidity, emphasizing the compelling demands for reliable prediction tools.3 Traditional evaluation approaches including biochemical indices, cultures, and imaging are often unavailable or unreliable due to nonspecific biochemical indices, misleading non-bone cultures, and limitations in sensitivity and specificity in imaging, also manual interpretation of radiographs and MRIs is prone to variability and key signs lost.4 Moreover, inflammatory biomarkers, such as CRP and ESR, in serum also have limitations of delayed sensitivity, invasive sampling, and poor specificity in distinguishing recurrence from post-MT treatment changes.5

Accumulating evidence show that radiomics provide considerable help by extracting high-dimensional quantitative features from medical images and integrating them with clinical data,6 such as oncology,6 musculoskeletal diseases,7,8 bones, and joints.9,10 However, current MRI-based radiomics for predicting COR in patients with MT-treatment are limited by small sample sizes or insufficient external validation.11-13 Recent studies show that models combined with radiomics can significantly raising predictive performance of various disease outcomes. For example, radiomic features from CT and MRI scans have effectively predicted cancer treatment responses.14 Preliminary investigation in patients with osteomyelitis also suggests that texture analysis of MRI scans can identify early signs of recurrence.15 It is worth noting that, the acquisition of multicentre datasets and a significant improvement in model interpretability are urgently required, and current radiomics are young in patients with MT-treatment due to lack of suitable features and machine learning algorithms.16

Herein, we retrospectively analyzed chronic osteomyelitis patients after MT surgery from Hebei Province Hospital of Chinese Medicine and The Third Hospital of Hebei Province to develop a comprehensive prediction model integrating clinical and MRI-derived radiomic features using machine learning algorithms. Additionally, we further developed a web-based application for enhancing the feasibility and usability of COR prediction.

Methods

Study design, patient selection, and data collection

We conducted a retrospective analysis of totally 1,368 patients with limb osteomyelitis. To stringently evaluate model generalizability, a center-wise validation strategy was employed. Consecutively enrolled patients from Hebei Province Hospital of Chinese Medicine (January 2019 to May 2023) were randomly divided into a training set and an internal hold-out validation set (7:3 ratio) for model development and tuning.17 The Synthetic Minority Over-sampling Technique (SMOTE) was applied to address class imbalance in the training set. The second independent cohort from the Third Hospital of Hebei Province (January 2015 to May 2023) was assembled under identical protocols, served as an external test set. We developed a new evaluation framework, which provides a robust assessment of model performance and its prospective generalizability to unseen clinical environments. This framework provides a robust assessment of model performance and its prospective generalizability to unseen clinical environments. The inclusion criteria were: age ≥ 18 years, histopathological or microbiological confirmation of osteomyelitis, complete surgical debridement records, pretreatment MRI scans, and at least two years of follow-up data. The exclusion criteria were: concurrent malignancies, immunocompromised states (defined as the use of immunosuppressants, diabetes with haemoglobin (Hb)A1c > 6.5%, chronic kidney disease, chronic inflammatory disorders, or malnutrition), incomplete records, or treatment outside our institutions.

Patient clinical information was extracted from electronic medical records, including demographics (age, sex, smoking and alcohol histories, metabolic diseases, and education), laboratory markers (ESR, procalcitonin (PCT), and CRP), identified microbiological results, and disease-specific features (aetiology, infection location, symptom duration, prior surgeries, and hospitalization history). MRI data were retrieved from the picture archiving and communication system (PACS) system with quality control to ensure imaging completeness. All laboratory assessments and pretreatment MRI scans were obtained prior to the initiation of any antimicrobial or surgical intervention for osteomyelitis. A flowchart illustrating the study steps is presented in Figure 1.

Fig. 1.

A schematic showing an image based and clinical data machine learning workflow, including image acquisition, segmentation, feature extraction, model training, clinical data integration, and combined model performance evaluation. The figure is a schematic diagram showing a parallel workflow that integrates imaging data and clinical data for machine learning based prediction. In the upper pathway, the process begins with image acquisition using a medical imaging scanner, followed by image segmentation where a target region is delineated within the scan. Feature extraction and selection are then performed, illustrated by a segmented anatomical structure and a bar chart of quantitative imaging features. These features are used to train multiple machine learning models, including support vector machine (SVM), extreme gradient boosting (XGBoost), random forest, and logistic regression. In the lower pathway, clinical data collection is shown, including patient consultation, electronic medical record review, blood sample collection, and pathological or physical examination findings. These clinical variables undergo feature selection using univariate and multivariate logistic regression analysis. Arrows from both the imaging based workflow and the clinical data workflow converge into a combined model evaluation section on the right. This section contains multiple performance plots, including receiver operating characteristic curves, calibration curves, and decision curve analysis graphs, which assess predictive accuracy, agreement, and clinical utility. The entire diagram relies on labeled icons, arrows, charts, and spatial organization to convey data flow, integration, and analysis steps without dependence on color to communicate meaning.

Open in a new tab

Flowchart of the study steps. SVM, support vector machine; XGboost, eXtreme gradient boosting.

Masquelet technique protocol

MT is a well-recognized surgical protocol for managing chronic osteomyelitis.18 It consists of two distinct stages.2 Initially, thorough debridement of the infected bone and surrounding tissues is conducted, followed by the placement of a calcium sulfate spacer. This spacer promotes the development of a well-vascularized, biologically active membrane over a period of 4 to 8 weeks. In the second stage, the spacer is removed, and an autologous bone graft is implanted into the defect site. The presence of the membrane enhances graft integration and supports osseous regeneration. This protocol has demonstrated efficacy in eradicating infection, facilitating bone reconstruction, and minimizing the risk of recurrence, making it a valuable approach for treating complex osteomyelitis.

Recurrence assessment and follow-up

COR was diagnosed according to the BACH criteria,19 requiring fulfillment of at least one of the following definitive criteria: 1) B - Bone Culture: Isolation of a pathogenic microorganism from bone tissue obtained during surgery or percutaneous biopsy; 2) A - Antibiogram: Antimicrobial susceptibility testing performed on the isolated pathogen to guide targeted therapy; 3) C - Clinical Signs: The presence of classic clinical features such as a draining sinus tract, localized pain, erythema, swelling, or fever attributable to the site of infection; 4) H - Histology: Histopathological examination of bone tissue revealing neutrophilic infiltration, osteonecrosis, or other acute inflammatory changes consistent with osteomyelitis. All patients underwent standardized clinical evaluations at one, three, six, and 12 months following post-MT treatment. Infection ‘cured’ was defined as infection-free after treatment in the two-year follow-up.20 Each follow-up visit comprised a physical examination, serological tests (CRP and ESR), and an radiograph to confirm any suspected recurrence.

Image processed and radiomics feature extraction

MRI examinations were performed on a 1.5 T scanner (MAGNETOM Avanto, Siemens, Germany) with a musculoskeletal coil. T2-weighted imaging (repetition time (TR)/echo time (TE) = 3000/88 ms) was used due to its high contrast resolution for detecting osteomyelitis inflammation. The protocol included 4 mm slices with a 0.4 mm gap, acquired in coronal, sagittal, and axial planes as needed. Two experienced musculoskeletal radiologists (each with > 10 years of experience) manually segmented inflammatory regions using 3DSlicer software (v5.3.0; Brigham and Women's Hospital, USA), carefully delineating regions of interest on consecutive coronal T2-weighted images while maintaining anatomical boundaries.

For radiomics feature extraction, the “PyRadiomics” toolbox (v3.0.1; Dana-Farber Cancer Institute and Brigham and Women's Hospital, USA) was used, following Image Biomarker Standardization Initiative guidelines. Preprocessing involved original T2-weighted images, Laplacian of Gaussian filtered images (σ = 1.0, 3.0, 5.0 mm), and wavelet-transformed images. This yielded 1274 quantitative features, including first-order intensity statistics, 3D-shape descriptors, and second-order texture features (gray-level co-occurrence matrix, run length matrix, size zone matrix, and dependence matrix). We normalized all features using a z-score and conducted quality assessment before analysis. All features were normalized using z-score and underwent quality assessment before analysis.

Feature screening and models establishment

Clinical features were initially selected through univariate and multivariate logistic regression analyses to identify independent risk factors for osteomyelitis recurrence. Radiomics features were selected using the minimum redundancy maximum relevance (mRMR) algorithm, which optimizes the balance between feature relevance and redundancy to enhance model generalizability.21 The stability of the mRMR ranking was rigorously assessed through 100 iterative runs, confirming the reproducibility of the top feature subset. The optimal dimensionality of the radiomics signature was subsequently determined by identifying the inflection point on the cross-validated performance curve, thereby balancing model complexity with predictive power. The optimal feature count was determined by plotting model performance curves, and key features were selected to form a radiomics score (Rad-score, RS) using weighted linear regression. The RS-recurrence relationship was validated through correlation analysis and violin scatter plots. Combined recurrence prediction models integrating clinical and radiomic features were evaluated using logistic regression (LR), eXtreme gradient boosting (XGBoost), random forest (RF), and support vector machine (SVM).

Evaluation and mobile application

Model performance was comprehensively assessed using area under the curve (AUC), positive predictive value, F1 score, sensitivity, and specificity, with DeLong’s test for AUC comparison. Predictive consistency was assessed using calibration curves, while clinical utility was evaluated with decision curves. The prediction process was elucidated through Shapley additive explanation (SHAP) analysis. The model’s generalizability was confirmed through external validation cohorts. A R-Shiny application (version 4.3.1, R Studio, USA) was developed to facilitate clinical translation, enabling real-time visualization of individualized recurrence risk probabilities based on patient-specific parameters and imaging features.

Statistical analysis

Statistical analyses performed using Python (version3.1.2, Python Software Foundation, USA) and R (version 4.3.1, R Foundation for Statistical Computing, Austria). Categorical variables were presented as counts (%) and analyzed by χ²/Fisher’s test; non-normally distributed continuous variables as median (IQR) using Mann-Whitney U test. Clinical features with a significance level of p < 0.1 from the univariate analysis were incorporated into the multivariate logistic regression model. The use of a p < 0.1 threshold for inclusion in the multivariate model was not based on strict statistical significance but was employed as a conservative strategy to reduce the risk of omitting potential predictors with possible adjusted associations. Receiver operating characteristic (ROC) curves (compared using the DeLong test) and decision curves were employed, with statistical significance set at p < 0.05 (two-tailed).

Results

Baseline characteristics of study cohorts

Upon inclusion and exclusion criteria, ultimately, a total of 279 eligible chronic osteomyelitis patients with MT treatment were included in this study. 279 patients were enrolled (mean age 43.19 years; 69.9% male) across training (n = 93), validation (n = 40), and external validation cohorts (n = 146), with recurrence rates of 23.7% (22/93), 20.0% (8/40), and 19.9% (29/146), respectively (Figure 2). The cohorts were predominantly male, with most patients having high rates of metabolic diseases. Recurrent cases exhibited longer hospitalization durations and higher levels of inflammatory markers, and Staphylococcus aureus infection was the most frequently pathogen (Table I).

Fig. 2.

A flowchart showing patient selection for an osteomyelitis recurrence study, detailing inclusion and exclusion from two medical centers, resulting cohorts for training, validation, and external validation. The figure is a flowchart describing patient selection and cohort allocation for a study predicting osteomyelitis recurrence. At the top, a box states that 1368 patients were diagnosed with osteomyelitis. A downward arrow leads to a box indicating that 568 of these patients underwent Masquelet technique therapy. The flow then splits into two dashed boxes representing two medical centers. The left box shows Medical Center 1 with data collected from January 2019 to May 2023, including 278 patients. Within this box, exclusion criteria are listed: 37 patients younger than 18 years, 6 with concurrent malignant disease, 20 treated at outside institutions, 56 with incomplete magnetic resonance imaging (MRI) or clinical data, and 26 lost to follow up. After exclusions, 133 eligible patients remain. These are further divided into a training cohort of 93 patients and a validation cohort of 40 patients. The right dashed box represents Medical Center 2 with data collected from January 2015 to May 2023, including 290 patients. Exclusions include 41 patients younger than 18 years, 4 with concurrent malignant disease, 12 treated outside the institution, 68 with incomplete MRI or clinical data, and 19 lost to follow up, leaving 146 eligible patients. These 146 patients form an external validation cohort. At the bottom of the diagram, all cohorts feed into a final box labeled model construction to predict osteomyelitis recurrence. The flowchart uses labeled boxes and arrows to present eligibility filtering and cohort assignment without reliance on color.

Open in a new tab

Flowchart of patient inclusion and exclusion.

Table I.

Characteristics in two cohorts with or without recurrence.

Characteristic Training cohort (n = 93） Validation cohort (n = 40） External-validation cohort (n = 146）
No recurrence
(n = 71）		 Recurrence
(n = 22）		 No recurrence
(n = 32）		 Recurrence
(n = 8）		 No recurrence
(n = 117）		 Recurrence
(n = 29）
Mean age, yrs (SD) 45.0 (32.0 to 54.0) 52.0 (40.8 to 61.0) 38.0 (25.8 to 48.0) 59.5 (52.0 to 64.0) 42.2 (36.0 to 50.0) 55.1 (39.8 to 58.0)
Median hospitalization duration, days (IQR) 21.0 (11.0 to 32.5) 24.5 (16.5 to 41.5) 20.5 (13.8 to 42.0) 24.0 (19.3 to 32.5) 19.9 (12.8 to 30.6) 19.3 (11.1 to 29.8)
Sex, n (%)
Male 52 (73) 15 (68) 23 (72) 7 (87) 75 (64) 23 (79)
Female 19 (27) 7 (32) 9 (28) 1 (13) 42 (36) 6 (21)
Smoker, n (%)
No 46 (65) 16 (73) 18 (56) 6 (75) 86 (74) 20 (69)
Yes 25 (35) 6 (27) 14 (44) 2 (25) 31 (26) 9 (31)
Alcohol, n (%)
No 55 (77) 18 (82) 24 (75) 7 (88) 96 (82) 18 (62)
Yes 16 (23) 4 (18) 8 (25) 1 (12) 21 (18) 11 (38)
Metabolic diseases, n (%)
No 58 (82) 12 (55) 26 (81) 5 (62) 105 (90) 26 (90)
Yes 13 (18) 10 (45) 6 (19) 3 (38) 12 (10) 3 (10)
Educational standard, n (%)
Higher 25 (35) 7 (32) 12 (38) 3 (38) 36 (31) 7 (24)
Primary/secondary 46 (65) 15 (68) 20 (62) 5 (62) 81 (69) 22 (76)
ESR, n (%)
Normal 45 (63) 7 (32) 18 (56) 4 (50) 66 (56) 9 (31)
Elevated 26 (37) 15 (68) 14 (44) 4 (50) 51 (44) 20 (69)
CRP, n (%)
Normal 42 (59) 13 (59) 21 (66) 8 (100) 68 (58) 16 (55)
Elevated 29 (41) 9 (41) 11 (34) 0 (0) 49 (42) 13 (45)
PCT, n (%)
Normal 61 (86) 19 (86) 29 (91) 8 (100) 104 (89) 23 (79)
Elevated 10 (14) 3 (14) 3 (9) 0 (0) 13 (11) 6 (21)
Gram-negative bacilli, n (%)
No 50 (70) 18 (82) 25 (78) 6 (75) 86 (74) 15 (52)
Yes 21 (30) 4 (18) 7 (22) 2 (25) 31 (26) 14 (48)
Infection site, n (%)
Hip 11 (16) 3 (14) 4 (14) 1 (12) 11 (9) 3 (10)
Lower limb 37 (52) 12 (54) 13 (41) 4 (50) 63 (54) 14 (48)
Foot 18 (25) 7 (32) 14 (42) 3 (38) 38 (33) 10 (35)
Upper limb 5 (7) 0 (0) 1 (3) 0 (0) 5 (4) 2 (7)
Infection duration, n (%)
≤ 3 months 33 (46) 0 (0) 14 (44) 0 (0) 52 (44) 14 (48)
3 to 18 months 26 (37) 15 (68) 12 (37) 7 (88) 49 (42) 12 (42)
≥ 18 months 12 (17) 7 (32) 6 (19) 1 (12) 16 (14) 3 (10)
Previous surgery, n (%)
< 2 57 (80) 19 (86) 25 (78) 7 (88) 71 (61) 17 (59)
≥ 2 14 (20) 3 (14) 7 (22) 1 (12) 46 (39) 12 (41)
Aetiology, n (%)
Traumatic 51 (72) 16 (73) 25 (78) 6 (75) 84 (72) 23 (79)
Non-traumatic 20 (28) 6 (27) 7 (22) 2 (25) 33 (28) 6 (21)
Bacterial species, n (%)
Staphylococcus aureus 27 (38) 11 (50) 7 (22) 2 (26) 45 (39) 8 (27)
Other 15 (21) 5 (23) 15 (47) 3 (37) 34 (29) 9 (32)
Multiple 23 (32) 4 (18) 7 (22) 3 (37) 34 (29) 8 (27)
Not detected 6 (9) 2 (9) 3 (9) 0 (0) 4 (3) 4 (14)
Open in a new tab

PCT, procalcitonin.

Feature selection and radiomics signature construction

Clinically, the results of univariate regression analyses identified that age, metabolic diseases, ESR, previous surgeries, and infection duration were significantly related to recurrence, while multivariate regression analysis determined that age (OR 1.060; 95% CI 1.016 to 1.107; p = 0.007), ESR (OR 3.359; 95% CI 1.148 to 9.820; p = 0.026), previous surgeries (OR 7.194; 95% CI 2.445 to 21.162; p = 0.001), and infection duration (OR: 2.460, 95% CI 1.132 to 5.340; p = 0.022, Wald test) were independent predictors of recurrence (Table II).

Table II.

Univariate and multivariate analysis of infection recurrence.

Variable Univariate regression analysis Multivariate regression analysis
OR CI-low CI-high p-value* OR CI-low CI-high p-value*
Age 1.067 1.030 1.105 0.001 1.060 1.016 1.107 0.007
Metabolic diseases 3.381 1.406 8.128 0.006 2.933 0.958 8.979 0.059
ESR 2.720 1.172 6.312 0.020 3.359 1.148 9.820 0.026
Infection duration 3.003 1.613 5.590 0.001 2.460 1.132 5.340 0.022
Previous surgery 6.240 2.581 15.084 0.001 7.194 2.445 21.162 0.001
Open in a new tab
*

Wald test.

OR, odds ratio.

Upon 1274 MRI radiomic variables acquired by region of interest (ROI) and the mRMR algorithm, 8 features (including wavelet-based texture metrics and log-sigma transformed gray-level dependence) were selected according to performance variation curve (Figures 3a and 3b). Subsequently, a new RS integrated eight features was produced using the following formula: RS = (-0.0757* wavelet_glrlm_wavelet-HHL-RunEntropy)+(-0.8912*wavelet_firstorder_wavelet-HLL-Skewness)+(-0.9233*wavelet_glrlm_wavelet-HLH-LongRunLowGrayLevelEmphasis)+(-0.0967*wavelet_glcm_wavelet-LLL-Correlation)+(0.0397*wavelet_gldm_wavelet-LHH-LargeDependenceLowGrayLevelEmphasis)+(0.0003*original_glcm_ClusterShade) + (0.0086* wavelet_glcm_wavelet-LHH-Imc1)+(-0.0155*wavelet_glszm_wavelet-HLL-SmallAreaLowGrayLevelEmphasis). Ultimately, we found that the patients with recurrence had obviously higher rad-scores than those without (p = 0.002, r = 0.31) (Figure 3c).

Fig. 3.

A three panel figure showing feature selection performance, feature importance rankings from radiomics analysis, and a comparison of radiomics score distributions between recurrence and no recurrence groups. The figure contains three panels labeled a, b, and c. Panel a shows a line plot of average cross validation accuracy plotted against the number of selected features, ranging from eight to fifteen on the horizontal axis. Each point represents a model iteration, with accuracy values fluctuating across feature counts, indicating a peak accuracy around twelve and fourteen features. Panel b presents a horizontal bar chart labeled Feature importance, ranking radiomics features derived from wavelet and texture analysis. The features include wavelet based run length, gray level dependence, gray level co occurrence matrix, and first order statistics, with original gray level co occurrence matrix cluster shade shown as one of the most influential features. The horizontal axis represents relative importance values, while the vertical axis lists individual feature names. Panel c displays two violin plots comparing radiomics score, abbreviated as Rad score (radiomics score), between patients with no recurrence and those with recurrence. Each violin plot includes individual data points, a central box indicating median and interquartile range, and distribution shape showing score spread. A correlation coefficient of r equal to zero point three one and a p value of zero point zero zero two are displayed above the plots, indicating a statistically significant association. The figure uses axes, shapes, and numeric labels rather than color to communicate analytical results.

Open in a new tab

a) Performance variation curve assessment of models with different feature counts. b) Importance chart of imaging features. c) Violin plot of Rad-score (RS) in the recurrence group and non-recurrence group using the Mann-Whitney U test.

Model performance comparison and interpretation

Next, four predictive models integrated RS and clinical features were developed according to machine learning algorithms with combined models (C) – SVM-C, RF-C, XGBoost-C, and LR-C – in which the LR-combined model stood out because of the superior performance and robustness demonstrated through the radar chart (Figure 4a) and the ROC curves (Figure 4b). This model achieved impressive AUC scores of 0.901 (95% CI 0.819 to 0.963) in the training cohort, 0.952 (95% CI 0.843 to 1.000) in the validation cohort, and 0.910 (95% CI 0.859 to 0.954) in the external validation cohort (all p-values from DeLong’s test < 0.05, Table III). Moreover, the model outperformed the single clinical or radiomics model in terms of predicted probabilities (Figures 5a to 5c) and decision curve analysis (Figures 5d to 5f) with all p-values less than 0.01, indicating its potential advantage in clinical decision-making. Furthermore, SHAP analysis revealed that surgery times and age significantly boosting the COR risk in patients after MT treatment (Figure 6) while ESR, duration of infection, and RS exhibit minimal influence. These results indicated that LR-combined model developed is meaningful for understanding and enhancing the clinical utility.

Fig. 4.

A multi panel figure comparing model performance across training, validation, and external validation cohorts using radar charts for classification metrics and receiver operating characteristic curves for clinical, radiomics, and combined models. The figure contains two main rows labeled a and b, each presenting model performance across three cohorts: training, validation, and external validation. In row a, three radar charts summarize classification metrics for four models labeled LR-C for logistic regression clinical model, SVM-C for support vector machine clinical model, RF-C for random forest clinical model, and XGBoost-C for extreme gradient boosting clinical model. Each radar chart displays sensitivity, precision, specificity, F1 score, and area under the curve, abbreviated as AUC, arranged around a circular axis. The training cohort radar chart shows generally higher and more uniform performance across metrics, while the validation and external validation charts show slight reductions and variability among models. In row b, three receiver operating characteristic curves are shown, one for each cohort. The horizontal axis represents false positive rate and the vertical axis represents true positive rate. Each panel compares three models: a clinical model, a radiomics model, and a combined logistic regression model. The training cohort shows area under the curve values of approximately 0.86 for the clinical model, 0.68 for the radiomics model, and 0.90 for the combined model. The validation cohort shows similar trends with area under the curve values around 0.91, 0.66, and 0.95 respectively. The external validation cohort again demonstrates higher performance for the combined model compared with either individual component. Across all panels, axes, labels, and curve shapes convey performance differences without reliance on color.

Open in a new tab

a) Radar charts analysis for the predictive performance of models. b) ROC curve evaluation of clinical, radiomics, and LR-C models in three cohorts. AUC, area under the curve; LR-C, logistic regression combined model; RF-C, random forest combined model; SVM-C, support vector machine combined model; XGBoost-C, eXtreme Gradient Boosting combined model.

Table III.

Performance metrics of clinical, radiomics, and LR-combined models.

Model AUC (95% CI) Sensitivity Specificity F1 Score PPV
Clinical
Training 0.862 (0.749 to 0.961) 0.800 0.855 0.708 0.635
Validation 0.912 (0.771 to 1.000) 0.759 0.519 0.666 0.595
External-validation 0.877 (0.810 to 0.933) 0.562 0.942 0.673 0.835
Radiomics
Training 0.684 (0.561 to 0.804) 0.700 0.618 0.776 0.870
Validation 0.661 (0.479 to 0.847) 0.690 0.740 0.664 0.640
External-validation 0.677 (0.593 to 0.768) 0.649 0.719 0.680 0.719
LR-Combined
Training 0.901 (0.819 to 0.963) 0.850 0.882 0.862 0.875
Validation 0.952 (0.843 to 1.000) 0.801 0.910 0.866 0.946
External-validation 0.910 (0.859 to 0.954) 0.600 0.952 0.662 0.739
Open in a new tab

AUC, area under the curve; PPV, positive predictive value.

Fig. 5.

A six panel figure showing decision curve analysis and calibration curves comparing radiomics, clinical, and combined predictive models across training, validation, and external validation cohorts. The figure consists of six panels labeled a through f that evaluate predictive model performance using decision curve analysis and calibration curves across three cohorts. Panels a, b, and c present decision curve analysis plots for the training cohort, validation cohort, and external validation cohort respectively. In each plot, the horizontal axis represents threshold probability and the vertical axis represents net benefit. Three prediction models are shown: a radiomics model, a clinical model, and a combined model. Reference lines labeled treat all and treat none indicate baseline strategies. Across all three cohorts, the combined model shows a higher net benefit over a wider range of threshold probabilities compared with either the radiomics or clinical model alone. Panels d, e, and f display calibration curves for the training cohort, validation cohort, and external validation cohort respectively. In these plots, predicted probability is shown on the horizontal axis and observed probability on the vertical axis. Each panel includes curves for the radiomics model, the clinical model, and the combined model, along with a diagonal reference

Open in a new tab

a) to c) Calibration curves and d) to f) decision curves of the radiomics, clinical, and combined models.

Fig. 6.

A Shapley value summary plot showing the importance and distribution of five features, including surgery times, age, erythrocyte sedimentation rate, infection duration, and radiomics score, in a predictive model. The figure is a Shapley additive explanations plot summarizing feature importance and effect direction for a predictive model. The horizontal axis represents Shapley values, indicating the contribution of each feature to the model output, with negative values reducing and positive values increasing predicted risk. Five features are listed vertically from top to bottom: surgery times, age, erythrocyte sedimentation rate (ESR), duration of infection, and radiomics score. For each feature, a horizontal swarm of points forms a violin shaped density distribution centered around zero, showing how feature contributions vary across patients. Individual points represent observations, and vertical spread reflects distribution density rather than sample count. A vertical reference line at zero Shapley value indicates no contribution. A color gradient bar on the right represents feature value magnitude, with low values at one end and high values at the other, though interpretation does not rely on color. Surgery times and age show the widest spread of Shapley values, indicating stronger influence on predictions, while radiomics score shows a narrower distribution. ESR and duration of infection display intermediate influence. The plot communicates both relative importance and directional impact of clinical and radiomics features on model predictions using spatial distribution, labels, and numeric axes rather than color alone.

Open in a new tab

SHapley Additive exPlanations (SHAP) dual-axis visualization: unravelling feature contributions and importance. Rad Score, radiomics score.

Application of a mobile application for COR prediction

Finally, we developed a web-based mobile application leveraging LR-combined model to prognosticate COR in patients with MT treatment, anticipating for practicality in clinic. This application demonstrated precision in estimating a high risk of 82.09% of recurrence for a patient who subsequently experienced an event (Figures 7a to 7e); conversely, a low risk of 11.95% for a patient with no recurrence (Figures 7f to 7j), both consistent with actual outcomes. These results indicated that the established web-based application could serve as a valuable clinical tool for clinical decision-making, postoperative COR risk assessment, and precise preventive interventions.

Fig. 7.

A multi panel figure showing representative clinical photographs, magnetic resonance imaging scans, X ray images, radiomics score visualization, and an osteomyelitis recurrence prediction calculator for two example patients. The figure presents two representative patient cases arranged in two rows labeled a through e for the first case and f through j for the second case. In the top row, panel a shows a clinical photograph of a lower limb with a surgical incision and surrounding skin changes. Panel b shows a magnetic resonance imaging scan of the affected limb demonstrating internal tissue structure. Panel c shows an X ray image of the limb with visible bone architecture and surgical fixation. Panel d shows a healed postoperative photograph of the same limb. Panel e displays an osteomyelitis recurrence prediction calculator interface, listing input variables including duration of infection, surgery times, erythrocyte sedimentation rate (ESR), age, and radiomics score, followed by a predicted recurrence probability. Below these images, a horizontal scale labeled Rad Score visualizes the radiomics score relative to low and high contribution thresholds, with individual feature contributions listed beneath, including duration of infection, surgery times, age, and ESR. In the bottom row, panel f shows a clinical photograph of another lower limb with postoperative changes. Panel g presents a magnetic resonance imaging scan of the affected bone. Panel h shows an X ray image highlighting bone structure. Panel i shows a postoperative clinical photograph with healed incision. Panel j displays the prediction calculator output for this second case with the same input variables and a corresponding recurrence probability. Across both cases, the figure integrates clinical imaging, radiomics analysis, and predictive modeling using structured layouts, numerical labels, and clear panel organization rather than color to convey information.

Open in a new tab

Visualization of COR predictive mode through developed application. Figures a-e illustrate case 1 (recurrence) and figures f-j illustrate case 2 (non-recurrence) with accurately predicted COR recurrence status. a) Pretreatment sinus tract in the distal left femur of the thigh. b) Preoperative MRI image. c) Follow-up radiograph at 7 months post-surgery reveals signs of recurrence. d) The recurrence site corresponds to the previous sinus tract location, with surrounding skin exhibiting signs of exudation and ulceration. e) For Patient 1, inputting the relevant clinical data into the application yielded a predicted recurrence rate of 82.09%, which accurately aligned with the patient’s actual clinical outcome. f) Pre-treatment sinus tract in the proximal-mid right tibia of the lower leg. g) Preoperative MRI image. h) Follow-up radiograph at 24 months post-surgery shows no signs of recurrence. i) The wound has healed well. j) For Patient 2, application generated a predicted recurrence rate of 11.95% based on the input data, which was also consistent with the patient’s actual clinical situation. RS, radiomics score.

Discussion

This study developed a practical model that integrated clinical and radiomic features in the form of a web-based application for predicting COR in patients undergoing standardized MT treatment. Age, ESR, surgery times, and infection duration were identified as independent risk factors for COR. Interestingly, while metabolic disease patients showed significant risk in univariate analysis, this association did not hold in multivariate analysis. The discrepancy may be due to the confounding effect of age, as patients with metabolic diseases tend to be older, a factor that independently increases the risk of osteomyelitis recurrence (OR = 1.060). Age-related immunosenescence,22 with its associated immune dysfunction and comorbid metabolic disorders, can exacerbate inflammation and hinder tissue repair.23 Furthermore, age-related vascular insufficiency can compromise tissue perfusion and repair mechanisms,24 collectively contributing to the elevated risk of recurrence. These insights underscore the complex interplay between age, inflammation, and tissue repair in the context of osteomyelitis, highlighting the need for tailored treatment strategies that address these multifactorial challenges. Notably, ESR (OR = 3.359) was identified as a significant predictor, whereas no statistical significance for CRP, which is often recognized as an inflammatory indicator.25 This unusual result may be attributed to the distinct pathophysiology and temporal dynamics under context of the chronic osteomyelitis. Specifically, CRP, as a rapid-response acute-phase reactant, may have normalized by the time of long-term prognosis assessment and thus fails to reflect the cumulative inflammatory burden.26 In contrast, the ESR changes more gradually, and its sustained elevation provides a more relevant indicator of the inflammatory load over weeks to months. Elevated ESR level is closely linked to hyperfibrinogenemia and persistent humoral immune activity in chronic osteomyelitis, which may impair local defenses and facilitate bacterial infection.27 Both surgery times (OR = 7.194) and infection duration (OR = 2.460) disrupt tissue integrity and immune homeostasis, creating a conducive environment for recurrent infections.28,29

Radiomics provided additional predictive power by capturing subtle tissue changes invisible through clinical assessment alone. The selected MRI features, encapsulated in the RS, revealed that texture and gray-level dependencies could indicate underlying pathological states associated with COR in patients with MT treatment. This finding complements the work of Ma et al,30 who demonstrated that radiomic features could detect early signs of tissue changes invisible to the naked eye, thus offering a more comprehensive assessment of disease progression. The developed LR-Combined model’s superior performance (AUC values of 0.901, 0.952, and 0.910 across cohorts) demonstrates the value of integrating clinical and radiomic data, offering a comprehensive viewpoint of a patient’s COR risk after MT surgery.

Notably, the model’s performance on the independent, held-out external test set from Medical centre 2 (AUC = 0.910) is particularly encouraging. This center-wise validation strategy, which is recognized as a stringent test for clinical generalizability,31 underscores that our model captures robust predictors of recurrence that are transferable across different clinical settings, rather than being overfitted to the specifics of the development cohort.

However, several limitations of this study should be acknowledged. First, its retrospective nature may introduce selection bias and inherent limitations in data completeness. Although we employed strict data extraction protocols and objective diagnostic criteria to mitigate these issues, the findings must be interpreted with caution. Second, the study cohort was exclusively composed of patients treated with the Masquelet technique. While this allowed for a homogeneous study population, it may limit the generalizability of our model to patients managed with other surgical methods. Future studies require broader treatment cohorts to validate and extend our findings.

Notwithstanding these limitations, the mobile application’s accurate prediction validates its potential as a clinical decision support tool. This technology can facilitate personalized risk assessment, enabling clinicians to tailor postoperative care and preventive strategies. Future work should focus on expanding the model to include other imaging methods and clinical markers, enhancing its applicability across diverse patient populations and clinical settings. This study underscores the importance of a multidisciplinary approach in managing complex infections like osteomyelitis.

Author contributions

Y. Cao: Data curation, Formal analysis, Writing – original draft

X. Zhao: Data curation, Formal analysis

R. Wang: Data curation, Investigation

F. Hao: Data curation, Investigation

Z. Zhang: Investigation, Resources

T. Zhang: Investigation, Resources

L. Li: Methodology, Supervision, Writing – review & editing

D. Ma: Methodology, Supervision, Writing – review & editing

Funding statement

The author(s) received no financial or material support for the research, authorship, and/or publication of this article.

Data sharing

The data that support the findings for this study are available to other researchers from the corresponding author upon reasonable request.

Acknowledgements

We thank all participants and researchers who contributed to this study.

Ethical review statement

The study was approved by the Institutional Review Board of Hebei Provincial Hospital of Traditional Chinese Medicine (Approval No: HBZY2025-KYLL-006-01), which waived the requirement for informed consent given the retrospective nature of the analysis. Two medical centers mean “Hebei Province Hospital of Chinese Medicine” and “Third Hospital of Hebei Medical University (Approval No: HB/KB-06-06/2.0)”.

Open access funding

The authors report that they received open access funding for their manuscript from the Clinical Medicine Excellent Talents Training Project of Hebei Province, China (No. ZF2026243) and the Scientific Research Project of the Administration of Traditional Chinese Medicine of Hebei Province (No. 2024202).

© 2026 Cao et al. This is an open-access article distributed under the terms of the Creative Commons Attribution Non-Commercial No Derivatives (CC BY-NC-ND 4.0) licence, which permits the copying and redistribution of the work only, and provided the original author and source are credited. See https://creativecommons.org/licenses/by-nc-nd/4.0/

Data Availability

The data that support the findings for this study are available to other researchers from the corresponding author upon reasonable request.

References

  • 1. Bury DC, Rogers TS, Dickman MM. Osteomyelitis: diagnosis and treatment. Am Fam Physician. 2021;104(4):395–402. [PubMed] [Google Scholar]
  • 2. Masquelet A, Kanakaris NK, Obert L, Stafford P, Giannoudis PV. Bone repair using the masquelet technique. J Bone Joint Surg Am. 2019;101(11):1024–1036. doi: 10.2106/JBJS.18.00842. [DOI] [PubMed] [Google Scholar]
  • 3. Zhu H, Gao Y, Wang C, et al. A nomogram for decision-making assistance on surgical treatment of chronic osteomyelitis in long bones: Establishment and validation based on a retrospective multicenter cohort. Int J Surg. 2022;99:106267. doi: 10.1016/j.ijsu.2022.106267. [DOI] [PubMed] [Google Scholar]
  • 4. Debs P, Boutin RD, Smith SE, et al. Chronic nonspinal osteomyelitis in adults: consensus recommendations on percutaneous bone biopsies from the society of academic bone radiologists. Radiology. 2024;311(1):e231348. doi: 10.1148/radiol.231348. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 5. Lew DP, Waldvogel FA. Osteomyelitis. Lancet. 2004;364(9431):369–379. doi: 10.1016/S0140-6736(04)16727-5. [DOI] [PubMed] [Google Scholar]
  • 6. Gillies RJ, Kinahan PE, Hricak H. Radiomics: images are more than pictures, they are data. Radiology. 2016;278(2):563–577. doi: 10.1148/radiol.2015151169. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 7. Chee CG, Yoon MA, Kim KW, et al. Combined radiomics-clinical model to predict malignancy of vertebral compression fractures on CT. Eur Radiol. 2021;31(9):6825–6834. doi: 10.1007/s00330-021-07832-x. [DOI] [PubMed] [Google Scholar]
  • 8. Fritz B, Yi PH, Kijowski R, Fritz J. Radiomics and deep learning for disease detection in musculoskeletal radiology: an overview of novel MRI- and CT-based approaches. Invest Radiol. 2023;58(1):3–13. doi: 10.1097/RLI.0000000000000907. [DOI] [PubMed] [Google Scholar]
  • 9. Wu H, Sun D, Wang S, et al. Incidence and risk factors of recurrence in limb osteomyelitis patients after antibiotic‐loaded cement spacer for definitive bone defect treatment. Bone Joint Res. 2023;12(8):467–475. doi: 10.1302/2046-3758.128.BJR-2022-0413.R2. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 10. Subbiah Ponniah H, Montin E, Namireddy S, Lattanzi R, Logishetty K. Radiomics in FAI: current status and perspectives. Bone Joint Res. 2025;14(10):832–838. doi: 10.1302/2046-3758.1410.BJR-2024-0353.R2. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 11. Jiang T, Lau S-H, Zhang J, et al. Radiomics signature of osteoarthritis: current status and perspective. J Orthop Translat. 2024;45:100–106. doi: 10.1016/j.jot.2023.10.003. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 12. Lu R, Xia S, Deng G, Li W, Huang Z, Ling H. Constructing a nomogram for the recurrence of tibial osteomyelitis after debridement surgery based on platelet to lymphocyte ratio at admission: a dual-center retrospective study from China. BMC Musculoskelet Disord. 2024;25(1):967. doi: 10.1186/s12891-024-08106-5. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 13. Deng Q, Ren S, Zhang N, et al. Unravelling lumbar disc herniation severity beyond MRI: integrated transcriptomic and metabolomic analyses highlight glycerophospholipid metabolism and inform a machine-learning diagnostic model: a pilot study. Bone Joint Res. 2025;14(5):434–447. doi: 10.1302/2046-3758.145.BJR-2024-0071.R1. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 14. Bera K, Braman N, Gupta A, Velcheti V, Madabhushi A. Predicting cancer outcomes with radiomics and artificial intelligence in radiology. Nat Rev Clin Oncol. 2022;19(2):132–146. doi: 10.1038/s41571-021-00560-7. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 15. Fritz J, Tzaribatchev N, Claussen CD, Carrino JA, Horger MS. Chronic recurrent multifocal osteomyelitis: comparison of whole-body MR imaging with radiography and correlation with clinical and laboratory data. Radiology. 2009;252(3):842–851. doi: 10.1148/radiol.2523081335. [DOI] [PubMed] [Google Scholar]
  • 16. Santos CS, Amorim-Lopes M. Externally validated and clinically useful machine learning algorithms to support patient-related decision-making in oncology: a scoping review. BMC Med Res Methodol. 2025;25(1):45. doi: 10.1186/s12874-025-02463-y. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 17. Sivakumar M, Parthasarathy S, Padmapriya T. Trade-off between training and testing ratio in machine learning for medical image processing. PeerJ Comput Sci. 2024;10:e2245. doi: 10.7717/peerj-cs.2245. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 18. Cuvillier M, Meucci JF, Cazorla C, Carricajo A, Neri T, Boyer B. Masquelet’s induced membrane technique associated with Reamer Irrigation Aspiration grafting and intramedullary Nailing (MaRIAN) for chronic diaphyseal osteomyelitis of the lower limb. Orthop Traumatol Surg Res. 2022;108(7):103395. doi: 10.1016/j.otsr.2022.103395. [DOI] [PubMed] [Google Scholar]
  • 19. Hotchen AJ, Dudareva M, Ferguson JY, Sendi P, McNally MA. The BACH classification of long bone osteomyelitis. Bone Joint Res. 2019;8(10):459–468. doi: 10.1302/2046-3758.810.BJR-2019-0050.R1. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 20. Cierny G. Surgical treatment of osteomyelitis. Plast Reconstr Surg. 2011;127 Suppl 1:190S–204S. doi: 10.1097/PRS.0b013e3182025070. [DOI] [PubMed] [Google Scholar]
  • 21. Cascianelli S, Galzerano A, Masseroli M. Supervised relevance-redundancy assessments for feature selection in omics-based classification scenarios. J Biomed Inform. 2023;144:104457. doi: 10.1016/j.jbi.2023.104457. [DOI] [PubMed] [Google Scholar]
  • 22. Subramanyam KN, Mundargi AV, Prabhu MV, Gopakumar KU, Gowda DSA, Reddy DR. Surgical management of chronic osteomyelitis: organisms, recurrence and treatment outcome. Chin J Traumatol. 2023;26(4):228–235. doi: 10.1016/j.cjtee.2023.01.003. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 23. Barbé-Tuana F, Funchal G, Schmitz CRR, Maurmann RM, Bauer ME. The interplay between immunosenescence and age-related diseases. Semin Immunopathol. 2020;42(5):545–557. doi: 10.1007/s00281-020-00806-z. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 24. Yang JP, Lee YN, Son JW, Han SK. The impact of extracorporeal shock wave therapy on microcirculation in diabetic feet: a pilot study. Adv Skin Wound Care. 2019;32(12):563–567. doi: 10.1097/01.ASW.0000604180.54706.b2. [DOI] [PubMed] [Google Scholar]
  • 25. Sproston NR, Ashworth JJ. Role of C-Reactive Protein at Sites of Inflammation and Infection. Front Immunol. 2018;9:754. doi: 10.3389/fimmu.2018.00754. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 26. Lapić I, Padoan A, Bozzato D, Plebani M. Erythrocyte sedimentation rate and C-reactive protein in acute inflammation. Am J Clin Pathol. 2020;153(1):14–29. doi: 10.1093/ajcp/aqz142. [DOI] [PubMed] [Google Scholar]
  • 27. Zhao W, Xu D, Dong Y, Feng W. Diagnostic value of neutrophil to lymphocyte ratio and serum biomarkers in chronic osteomyelitis. Sci Rep. 2025;15(1):21752. doi: 10.1038/s41598-025-05856-7. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 28. Ravi B, Jenkinson R, O’Heireamhoin S, et al. Surgical duration is associated with an increased risk of periprosthetic infection following total knee arthroplasty: a population-based retrospective cohort study. EClinicalMedicine. 2019;16:74–80. doi: 10.1016/j.eclinm.2019.09.015. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 29. Peersman G, Laskin R, Davis J, Peterson MGE, Richart T. Prolonged operative time correlates with increased infection rate after total knee arthroplasty. HSS J. 2006;2(1):70–72. doi: 10.1007/s11420-005-0130-2. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 30. Ma J, Yang F. [Advances in clinical diagnosis and treatment of chronic osteomyelitis in adults] Zhongguo Xiu Fu Chong Jian Wai Ke Za Zhi = Chinese journal of reparative and reconstructive surgery. 2020;34(5):651–655. doi: 10.7507/1002-1892.201909032. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 31. Jiang Y, Zhou K, Sun Z, et al. Non-invasive tumor microenvironment evaluation and treatment response prediction in gastric cancer using deep learning radiomics. Cell Rep Med. 2023;4(8):101146. doi: 10.1016/j.xcrm.2023.101146. [DOI] [PMC free article] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Data Availability Statement

The data that support the findings for this study are available to other researchers from the corresponding author upon reasonable request.

Articles from Bone & Joint Research are provided here courtesy of British Editorial Society of Bone and Joint Surgery

RESOURCES