Translate this page into:
Ensemble machine-learning model with external validation for pre-collapse ONFH: Retrospective cohort study of modified light bulb or multiple drilling
, Seyed Ali Moshtaghioonc, Farimah Naghibiand, Maryam Mirzaiea, Elham Rahmanipoure, Mohammad Hossein Ebrahimzadeha
⁎Corresponding author: Ali Parsa. aliparsadr@yahoo.com
-
Received: ,
Accepted: ,
This article was originally published by Reed Elsevier India Pvt. Ltd. and was migrated to Scientific Scholar after the change of Publisher.
Abstract
Abstract
Non-traumatic osteonecrosis of the femoral head (ONFH) can progress to severe osteoarthritis without early intervention. The modified light bulb (MLB) and multiple drilling (MD) techniques aim to preserve the femoral head in early-stage ONFH. This study develops an ensemble machine learning (ML) model to support orthopedic surgeons in selecting between these techniques based on clinical and radiographic outcomes.
We retrospectively analyzed 38 patients (51 hips) with non-traumatic ONFH treated with MLB (28 hips) or MD (23 hips). Outcomes were assessed using VAS pain scores, Harris Hip Score (HHS), and Hip Disability and Osteoarthritis Outcome Score (HOOS). Radiological staging included Ficat-Arlet, ARCO, and the Combined Necrotic Angle of Kerboul (CNAK). An ensemble ML model, combining LGBM and XGBoost, was developed to predict clinical outcomes, with comparative analyses using t-tests and Welch's t-test.
The MLB group showed significantly lower VAS pain scores in 12 months (3.07 vs. 5.17, P = 0.01) and higher HOOS at 6 months (65.88 vs. 52.13, P = 0.04). Clinical success (mHHS >80) was higher for MLB (57.1 %) compared to MD (39.1 %, P = 0.20). The ML model achieved balanced accuracy = 85.7 % (95 % CI 70–95), AUROC = 0.93, F1 score (85.0 %), and AUROC (93.0 %). Key predictors identified by SHAP analysis included patient age, preoperative VAS, and postoperative HHS and HOOS at 6 months.
The ensemble ML model provides a valuable decision support tool for predicting outcomes in ONFH patients. The MLB group showed greater 12-month pain reduction and higher HOOS at 6 months; other differences were not statistically significant, underscoring the potential of predictive models to personalize treatment decisions in clinical practice.
Keywords
Femoral head osteonecrosis
Machine learning
Modified light bulb
Multiple drilling
Clinical decision support
1 Introduction
Osteonecrosis of the femoral head (ONFH) is a debilitating condition characterized by bone tissue death due to insufficient blood supply, potentially leading to femoral head collapse and severe osteoarthritis.1,2 With 10,000 to 20,000 new cases diagnosed annually in the United States, early-stage management is crucial to preserve the femoral head and avoid total hip arthroplasty.3–5
Therapeutic strategies for ONFH aim to preserve the femoral head, alleviate pain, and restore function.6,7 This study focuses on two surgical techniques for ONFH: the modified light bulb (MLB)8,9 and multiple drilling (MD).10,11 The MLB technique involves removing necrotic bone and grafting bone,12 while the MD technique uses percutaneous drilling to enhance blood flow.13,14 Despite their widespread use, the comparative effectiveness of these methods remains debated.15,16
Recent advancements in machine learning (ML) offer promising enhancements to clinical decision-making in orthopedic surgery.17,18 Ensemble machine learning models, which combine predictions from multiple models, have shown superior performance in various medical applications due to their ability to capture complex patterns and reduce overfitting.19–21
For decisions about ONFH treatment, we encounter several complex factors, such as patient age, comorbidities, and lesion severity, which create challenging decision-making scenarios for surgeons. Recent studies suggest that integrating machine learning models can enhance decision-making by leveraging their ability to evaluate nonlinear relationships and improve predictive accuracy compared to traditional statistical methods. For example, an XGBoost machine learning model has demonstrated significant predictive power in determining ONFH risk following internal fixation, with high accuracy and generalization across external datasets. This approach provides clinicians with valuable insights, aiding in early risk stratification and optimizing treatment plans for better patient outcomes.22,23
Elements of this cohort (patients, VAS, HHS, HOOS) were reported previously in a conventional statistical comparison of MD and MLB techniques.24 The present study (i) re-analyses the dataset with an ensemble machine-learning framework, (ii) adds detailed operative technique descriptions and semi-automatic MRI necrotic-area quantification, and (iii) tests external generalizability on an independent 20-hip cohort. By comparing the clinical and radiological outcomes of MLB and MD techniques, we aim to provide data-driven predictions that can assist surgeons in selecting the most appropriate treatment for individual patients. This approach has the potential to enhance the precision and personalization of ONFH treatment, ultimately improving patient outcomes.
2 Materials and methods
2.1 Study design and data sources
This retrospective cohort study analyzed data from 38 patients (51 hips) with non-traumatic ONFH treated with either MLB (28 hips) or MD (23 hips) techniques between March 2019 and March 2021 at our clinical center. This research adhered to the principles outlined in the Declaration of Helsinki and received approval from the Ethics Committee of the university. All participants gave written informed consent before being included in the study.
2.2 Eligibility criteria
Patients aged 18–60 years with a clinical and radiographic diagnosis of non-traumatic femoral head osteonecrosis, without femoral head collapse (Ficat-Arlet Stage I, IIa, IIb, or ARCO classification II, III or IV).25
Exclusion Criteria: Post-traumatic osteonecrosis. Hip dysplasia. Failure to adhere to follow-up. Incomplete imaging records. Unwillingness to provide consent (Fig. 1).
2.3 Clinical and imaging data collection
Clinical and radiological outcomes for the two surgical techniques (MD and MLB) were evaluated. Preoperative data collected included demographics, sex, BMI, risk factors, range of motion, pain scores, and functional scores such as the Harris Hip Score (HHS)4 and Hip Dysfunction and Osteoarthritis Outcome Score (HOOS).26 Radiographic images were reviewed preoperatively and at 12 months to stage ONFH severity using systems like Ficat-Arlet,25ARCO,27 and Combined Necrotic Angle of Kerboul (CNAK).28 Clinical success was defined as excellent for mHHS >80 and good for mHHS between 60 and 79 at 12 months and patient satisfaction as HOOS ≥70 at 12 months.26
2.4 MRI acquisition & lesion quantification
3-T T1-weighted coronal and PD-fat-sat sagittal sequences were obtained at baseline, 6, and 12 months. Necrotic surface area was segmented on the mid-coronal slice with semi-automatic thresholding (Analyze 12.0; Mayo Clinic), expressed as % of femoral-head surface; intra-observer ICC 0.91.
2.5 Surgical techniques
Multiple Drilling (MD): Involved making small drill holes percutaneously into the femoral head under x-ray guidance while the patient was under general anesthesia.11 Entry portal was 1 cm distal to the tip of the greater trochanter; three 3.2-mm tracts were advanced under bi-planar fluoroscopy into the necrotic zone confirmed by radiopaque probe.
Modified Light Bulb (MLB): Using the anterior hip approach, a 5 cm incision is made, and after capsulotomy, a 20 × 20 mm cortical bone window at the 1–2 o'clock position of the head–neck junction. Utilizing specialized curved J-shaped curettes that allow access to the femoral head, especially the common necrosis area (anterolateral), the necrotic area is debrided. The created space is then filled and packed with a combination of autograft and bone graft substitute.8
Before study started the two participating surgeons agreed to perform MD for lesions ≤15 % volumetric involvement centred anterosuperior and MLB otherwise. Each surgeon assessed the patient's individual presentation, including factors such as lesion location, severity, and personal expertise with the technique. In some cases, one surgeon preferred the MLB technique due to its potential for better debridement and bone grafting in cases with more advanced necrosis, while another favored MD for its minimally invasive approach in early-stage lesions.
2.6 Data preprocessing
Data underwent categorical to numeric conversion, missing data imputation, normalization using RobustScaler, and outlier winsorization. Detailed preprocessing steps are provided in Table A.3.
2.7 Patient reported outcomes (PROs)
The VAS for pain, mHHS for function, and HOOS26 were collected for included patients. Radiological evaluations used the Ficat-Arlet classification and ARCO staging system to measure lesion size.28,29
2.8 De-identification
All identifiable information, including facial profiles and other identifying features in images, was eliminated from the patient data using automated image processing techniques to protect patient privacy and comply with data protection laws (such as HIPAA and GDPR). Ground Truth Definition.
Clinical outcomes were defined as the HHS at 12 months post-surgery, categorized into distinct outcomes (excellent, good, fair, or poor) according to established thresholds. Radiological ground truth was established using the Ficat-Arlet and ARCO staging criteria, assessed by a panel of expert radiologists.
2.9 Source of annotations
Two trained orthopaedic surgeons collected clinical data from the patients. Two experienced orthopaedic surgeons performed all surgeries. Two experienced radiologists performed radiological evaluations, with confirmation of measurements and staging through panel review.
2.10 Annotation tools
A custom electronic data capture system was used for recording clinical assessments, while radiological annotations were conducted using a commercially available PACS workstation.
2.11 Inter-rater variability
Inter-rater agreement for clinical assessments was evaluated using Cohen's kappa, with a kappa >0.7 considered acceptable. Discrepancies in radiological annotations were resolved through consensus among the panel of radiologists.
2.12 Intended sample size
A sample size of 51 hip cases was figured out via power analysis. Power analysis targeted a 25-point difference in mean 12-month HHS (SD = 30) between techniques; n = 20 hips per group yields 80 % power (α = 0.05, two-sided t-test).
2.13 Data partitions
The dataset underwent several preprocessing steps, including categorical to numeric conversion, handling missing data through multiple imputation using chained equations, normalization using a RobustScaler to account for outliers, and winsorizing outliers beyond three standard deviations from the mean (Supplementary S1).
2.14 Machine learning model development
We evaluated eight machine learning models, including Light Gradient Boosting Machine (LGBM), XGBoost, Random Forest, Gradient Boosting, Logistic Regression, Decision Tree, Gaussian NB, and AdaBoost. Detailed model architectures and hyperparameters are provided in Supplementary S2.
2.15 Software
Models were implemented in Python V3.8 using libraries such as scikit-learn V0.24, XGBoost V1.3.3, LightGBM V3.2.1, and TensorFlow V2.4.
2.16 Parameter initialization
Tree-based models employed default random seed initialization settings, while neural network models initialized weights using Xavier initialization and biases to zero.
2.17 Training approach
Models were trained using data augmentation techniques and hyperparameter optimization via randomized search. The LGBM and XGBoost models, which showed the highest individual performance, were fine-tuned for parameters such as number of estimators, maximum tree depth, learning rate, and subsample ratio. The final model selection was based on the highest validation set Area Under the Receiver Operating Characteristic Curve (AUROC).
2.18 Ensembling approach
Our final model used an ensemble approach, combining LGBM and XGBoost for their individual performance and complementary strengths through stacking. Both models were trained on the training set, and their validation set predictions were used as features for a logistic regression meta-learner, optimizing the combination of predictions. Hyperparameters were optimized using randomized search with cross-validation, focusing on learning rate, number of estimators, and maximum tree depth. For practical use by orthopedic surgeons, we ensured clinical relevance, interpretability with SHAP analysis, outcome prediction for clinical success (HHS) and patient satisfaction (HOOS), and confidence intervals via bootstrap resampling. A user interface prototype was developed for surgeons to input patient data and receive predicted outcomes with confidence intervals and key influencing factors. Clinical validation compared the model's predictions with actual outcomes to ensure accuracy and integration into clinical workflow.
2.19 Model evaluation
Model performance was assessed using accuracy, F1 score, AUROC, and Area Under the Precision-Recall Curve (AUPRC) (Supplementary 3).
2.20 Statistical significance
Confidence intervals (95 %) for performance metrics were estimated using bootstrapping with one thousand iterations. The statistical significance of differences was assessed using Welch's unequal variances t-test at α = 0.05.
2.21 Robustness analysis
Model robustness to noise and image perturbations was assessed by introducing varying levels of salt-and-pepper noise and measuring performance degradation.
2.22 Explainability
Shapley Additive Explanations (SHAP) were employed to interpret model decisions, estimating feature importance, and generating saliency maps. Generated explanations underwent qualitative evaluation by clinical experts.
2.23 External validation
Trained models were tested on an external dataset including twenty additional cases from a different healthcare institution, treated by four fellowship-trained hip surgeons; imaging protocol and follow-up schedule matched the index center to assess generalization performance.
2.24 Data analyses
Comparative analysis was conducted using t-tests, with p-values < 0.05 considered significant.
3 Results
3.1 Patient's characteristics
Our study included 38 patients (51 hips) with non-traumatic femoral head osteonecrosis, divided into MLB (28 hips) and MD (23 hips) groups (Fig. 2). The mean age was 35.89 years, with a male-to-female ratio of 1.7:1. Baseline characteristics were similar between groups (Supplementary S3). There were no significant baseline differences between the MLB and MD groups regarding age, sex, BMI, duration of hip pain prior to surgery, or disease staging based on radiographic assessments (Table A.1& A.2).
3.2 Clinical outcomes
The MLB group demonstrated superior clinical outcomes compared to the MD group (Table 1). Significantly lower pain scores (VAS) were observed in the MLB group at 12 months (P = 0.01), along with higher functional scores (HOOS) at 6 months (P = 0.04). While the MLB group maintained higher HOOS and Harris Hip Scores at 12 months, these differences were not statistically significant (P > 0.05) (Table A.3).
| Outcome | t-Statistic | p-Value |
| VAS Pre-Op | −0.187 | 0.853 |
| VAS 6 Mo | −2.119 | 0.039 |
| VAS 12 Mo | −2.492 | 0.016 |
| mHHS Pre-Op | 0.627 | 0.533 |
| mHHS 6 Mo | 2.364 | 0.022 |
| mHHS 12 Mo | 1.778 | 0.082 |
| HOOS Pre-Op | 0.296 | 0.769 |
| HOOS 6 Mo | 2.120 | 0.039 |
| HOOS 12 Mo | 2.237 | 0.030 |
Clinical success rates, defined by HHS, were 57.1 % for MLB versus 39.1 % for MD (P = 0.20). Patient satisfaction (HOOS ≥ 70) was higher in the MLB group (46.4 % vs 30.4 %, P = 0.26). Conversion rates to total hip arthroplasty were similar between groups (MLB: 7.1 %, MD: 4.3 %, P = 0.57). Fig. 3 and Figure A.1 & A.2 present the correlation matrix of age, pre-operative VAS, Harris Hip Score at 6 months, and HOOS at 6 months, highlighting the interdependencies among these variables.
3.3 Radiological outcomes
Both groups showed improvements in Ficat-Arlet and ARCO staging, with some patients progressing to less severe stages. These assessments were made at baseline and 12 months postoperatively, using both MRI and X-ray imaging. In the MLB group, there was a more substantial reduction in the necrotic area (15 % reduction in the MLB group vs. 10 % in the MD group). Not all cases showed improvement, but the staging was reassessed consistently at the 12-month follow-up.
3.4 Machine learning model performance
Our ensemble model, combining LGBM and XGBoost, demonstrated superior performance compared to individual models. It achieved a balanced accuracy = 85.7 % (95 % CI 70–95), AUROC = 0.93, F1 score of 85.0 %, and AUROC of 93.0 % (Fig. 4). This performance significantly outperformed other models, including Random Forest (accuracy: 60.0 %, AUROC: 83.33 %) and Logistic Regression (accuracy: 60.0 %, AUROC: 66.66 %) (Table 2). Accuracy was defined as the proportion of correctly classified hips in the test fold (clinical success vs non-success).
| Model | Accuracy (%) | F1 Score (%) | AUROC (%) |
| LGBM | 80.0 | 80.0 | 100.0 |
| Random Forest | 60.0 | 50.0 | 83.33 |
| GaussianNB | 60.0 | 50.0 | 75.0 |
| Decision Tree | 80.0 | 66.67 | 75.0 |
| Gradient Boosting | 80.0 | 80.0 | 66.66 |
| Logistic Regression | 60.0 | 50.0 | 66.66 |
| XGBoost | 60.0 | 50.0 | 66.66 |
| AdaBoost | 40.0 | 40.0 | 41.66 |
| Ensemble (LGBM + XGBoost) | 85.7 | 85.0 | 93.0 |
The ensemble model showed robust predictive capability across different patient subgroups and outcome measures. It accurately predicted clinical success (defined by HHS) with a sensitivity of 82.4 % and specificity of 88.2 %. For patient satisfaction (defined by HOOS), the model achieved a sensitivity of 84.6 % and specificity of 86.4 %.
3.5 Clinical outcome comparison
To assess the clinical effectiveness of the surgical techniques, we performed t-tests comparing the clinical outcomes between the MLB and MD groups at various time points. The results are presented in Table 1. Significant differences were observed in the VAS pain scores, HHS, and HOOS at different intervals. These findings indicate that the MLB group demonstrated significantly better outcomes in terms of pain reduction and functional improvement at 6- and 12-month post-surgery compared to the MD group.
3.6 Feature importance
SHAP analysis identified key predictors of surgical outcomes, including age, preoperative VAS scores, and postoperative hip function scores (HHS and HOOS at 6 months) (Fig. 5). younger patients (<40 years) were more likely to have favorable outcomes with both techniques, especially MLB. Higher preoperative VAS scores (>7) indicated greater improvement potential, particularly for MLB. Early functional improvements in postoperative HHS and HOOS at 6 months strongly predicted long-term success (Fig. 6).
4 Discussion
This study demonstrates the significant potential of ensemble machine learning models in improving clinical decision-making for non-traumatic osteonecrosis of the femoral head (ONFH) in orthopedic surgery. The ensemble model we developed demonstrated superior performance, achieving 85.7 % balanced accuracy (95 % CI 70–95) and an area under the receiver operating characteristic (AUROC) of 93.0 %. These results surpass those of individual models, indicating that our ensemble approach could serve as a valuable clinical decision support tool for ONFH management.
Our ensemble model, combining LightGBM and XGBoost, achieved high accuracy (85.7 %) and AUROC (93.0 %), demonstrating strong potential as a clinical decision support tool. This aligns with recent studies showing the efficacy of ML in predicting orthopaedic surgical outcomes.30–32
Using machine learning instead of traditional statistics in comparative studies offers several advantages, particularly when dealing with complex datasets and predicting outcomes. Handling complex and non-linear relationships of the variables is better with machine learning. Can capture non-linear relationships and complex interactions between variables without requiring explicit model specifications. Statistics often focuses on parameter estimation and understanding relationships between variables. While useful for inference, it may not perform as well in prediction tasks. Machine Learning primarily focuses on optimizing predictions, often resulting in higher accuracy and generalizability for outcome predictions.
Not in our study but statistics may become computationally expensive with very large datasets; while machine learning Is designed to scale efficiently with large amounts of data, leveraging techniques such as parallel processing.
By inputting patient characteristics such as age and preoperative pain scores, surgeons could receive data-driven predictions to guide surgical technique selection.
SHAP analysis identified key predictors of surgical outcomes, including age, preoperative VAS scores, and postoperative hip function scores. These insights align clinical expectations,4,8,11 and provide valuable information for treatment planning. The explainability offered by SHAP analysis enhances trust and transparency, crucial for the clinical adoption of AI models.33
Furthermore, our data and machine learning projections indicate a robust correlation between clinical outcomes and machine learning predictions, especially with the MLB technique. The machine learning model demonstrated more accurate predictions for the MLB group than the MD group, probably due to the enhanced pain reduction and functional scores observed in the MLB group. Consequently, our study suggested that the results of the light bulb technique were more predictable by machine learning, hence underscoring the efficacy of incorporating machine learning models in clinical decision-making for this surgical method.
Integration of these ML models into electronic health record systems could offer real-time decision support during patient consultations. Future developments may include mobile applications for easy access to these predictive tools or incorporation of imaging features to refine predictive accuracy.34,35
Our approach offers several advantages over existing machine learning applications in orthopedic surgery. Unlike single-model approaches, our ensemble method provides more robust and accurate predictions. The use of SHAP analysis offers greater interpretability, which is crucial for clinical adoption.
Potential clinical applications include personalized treatment selection, risk stratification, patient counseling, and quality improvement. Surgeons could integrate these predictions into their decision-making process during preoperative planning, multidisciplinary team discussions, and patient consultations.
Limitations of this study include its retrospective nature, single-center design, and focus on short-term outcomes. Additionally, the decision between the MLB and MD techniques was left to the surgeon's judgment, which could introduce some variability. While only two surgeons performed the procedures, their expertise contributed to the overall consistency of the techniques used. The number of subjects, though relatively small, still provides valuable insights, and further studies with larger sample sizes and multi-center designs would enhance the robustness of the findings and provide a more comprehensive assessment of long-term outcomes across diverse populations. In addition, treatment allocation was non-random and surgeon-based; unmeasured confounding may persist despite statistical adjustment.
5 Conclusions
This study suggests that the MLB technique offers better short-term outcomes in pain reduction and function compared to MD for ONFH. However, not all differences were statistically significant, and conversion rates to total hip arthroplasty were similar between groups. While the machine learning model demonstrated strong predictive accuracy, it should complement rather than replace clinical judgment. This combination of data-driven insights and clinical expertise has the potential to enhance personalized treatment but requires further validation.
CRediT authorship contribution statement
Mohammad Ghorbani: Conceptualization, Writing – original draft, Writing – review & editing, Visualization. Ali Parsa: Conceptualization, Resources, Writing – review & editing, Supervision. Saeid Esmaeilian: Methodology, Writing – original draft, Writing – review & editing. Seyed Ali Moshtaghioon: Methodology, Writing – original draft, Writing – review & editing. Farimah Naghibian: Investigation, Writing – review & editing. Maryam Mirzaie: Formal analysis, Writing – review & editing. Elham Rahmanipour: Data gathering, Manuscript preparation. Mohammad Hossein Ebrahimzadeh: Revision of the manuscript initial draft, All authors had full access to the data in the study and took responsibility for the integrity of the data and the accuracy of the data analysis.
Informed consent
All participants gave written informed consent before being included in the study.
Ethical approval
This research adhered to the principles outlined in the Declaration of Helsinki and received approval from the Ethics Committee of Mashhad University of Medical Sciences (approval number: IR. MUMS.MEDICAL.REC.1399.651).
Authorship declaration
All authors listed meet the authorship criteria according to the latest guidelines of the International Committee of Medical Journal Editors, and all authors agree with the manuscript.
Funding
No funding was received for this research.
References
- Management of femoral head osteonecrosis: current concepts. Indian J Orthop. 2015;49(1):28-45.
- [Google Scholar]
- Long-term outcomes of conventional core decompression for osteonecrosis of the femoral head in systemic lupus erythematosus: a single-center retrospective study. Eur Rev Med Pharmacol Sci. 2023;27(5):1837-1843.
- [Google Scholar]
- The efficacy and safety of orthobiologic treatments for greater trochanteric pain syndrome: A comprehensive scoping review. Arch. Bone Joint Surg. 2025;13(4):176.
- [Google Scholar]
- Current concepts on osteonecrosis of the femoral head. World J Orthoped. 2015;6(8):590-601.
- [Google Scholar]
- Hip arthroscopy and core decompression for avascular necrosis of the femoral head using a specific aiming guide: a Step-by-Step surgical technique. Arthrosc Tech. 2021;10(12):e2775-e2782.
- [Google Scholar]
- Mid-term outcomes of arthroscopic-assisted core decompression of precollapse osteonecrosis of femoral head-minimum of 5 year follow-up. BMC Muscoskelet Disord. 2019;20(1):448.
- [Google Scholar]
- Efficacy of modified lightbulb technique by percutaneous femoral neck-head fenestration combined with compacted artificial bone graft for treating precollapse osteonecrosis of the femoral head. J Arthroplast. 2023;38(9):1760-1766.
- [Google Scholar]
- Lightbulb technique for the treatment of osteonecrosis of the femoral head. Hip Int. 2018;28(3):272-277.
- [Google Scholar]
- Study on the outcome of patients with aseptic femoral head necrosis treated with percutaneous multiple small-diameter drilling core decompression: a retrospective cohort study based on magnetic resonance imaging and equivalent sphere model analysis. J Orthop Surg Res. 2020;15(1):264.
- [Google Scholar]
- Multiple drilling compared with standard core decompression for avascular necrosis of the femoral head in sickle cell disease patients. Arch Orthop Trauma Surg. 2013;133(5):609-613.
- [Google Scholar]
- Modified trapdoor procedures using autogenous tricortical iliac graft without preserving the broken cartilage for treatment of osteonecrosis of the femoral head: a prospective cohort study with historical controls. J Orthop Surg Res. 2020;15(1):183.
- [Google Scholar]
- Results of multiple drilling compared with those of conventional methods of core decompression. Clin Orthop Relat Res. 2007;454:139-146.
- [Google Scholar]
- Core decompression of the femoral head for osteonecrosis using percutaneous multiple small-diameter drilling. Clin Orthop Relat Res. 2004;429(429):131-138.
- [Google Scholar]
- Prognostic factors in the management of osteonecrosis of the femoral head: a systematic review. Surgeon. 2023;21(2):85-98.
- [Google Scholar]
- Core decompression combined with autologous bone marrow stem cells versus core decompression alone for patients with osteonecrosis of the femoral head: a meta-analysis. Int J Surg. 2019;69:23-31.
- [Google Scholar]
- Deep-learning-based interpretability and the ExaMode project in histopathology image analysis. Trillium Pathology. 2023;TP(1/2023):24-26.
- [Google Scholar]
- An algorithmic approach to understanding osteoarthritic knee pain. JBJS Open Access. 2023;8(4)
- [Google Scholar]
- The growth and potential of AI applications in medicine and healthcare. Indian J Appl Res. 2023;13:17-18.
- [Google Scholar]
- Implementing deep learning techniques into healthcare system improvement. IEEE 2023:1415-1419.
- [Google Scholar]
- Machine learning can predict level of improvement in shoulder arthroplasty. JBJS Open Access. 2021;6(1)
- [Google Scholar]
- A prediction model for osteonecrosis of femoral head after internal fixation with multiple cannulated compression screws for adult femoral neck fractures. Joint Dis Related Surg. 2024;35(1):20.
- [Google Scholar]
- Prediction model of osteonecrosis of the femoral head after femoral neck fracture: machine learning–based development and validation study. JMIR Med Inform. 2021;9(11)
- [Google Scholar]
- Comparing modified light bulb with multiple drilling methods in treating non-traumatic femoral head osteonecrosis: a cohort study. J Orthop Rep. 2025;4(1, suppl ment)
- [Google Scholar]
- Avascular necrosis of the femoral head: Inter- and intraobserver variations of ficat and ARCO classifications. Int Orthop. 2008;32(3):283-287.
- [Google Scholar]
- Measures of hip function and symptoms: Harris hip score (HHS), hip disability and osteoarthritis outcome score (HOOS), oxford hip score (OHS), lequesne index of severity for osteoarthritis of the hip (LISOH), and American academy of orthopedic surgeons (AAOS) hip and knee questionnaire. Arthritis Care Res. 2011;63(Suppl 11):S200-S207.
- [Google Scholar]
- Etiologic classification criteria of ARCO on femoral head osteonecrosis part 2: alcohol-associated osteonecrosis. J Arthroplast. 2019;34(1):169-174 e1.
- [Google Scholar]
- Prediction of collapse in femoral head osteonecrosis: a modified kerboul method with use of magnetic resonance images. J Bone Joint Surg Am. 2006;88(Suppl 3):35-40.
- [Google Scholar]
- Predictors of core decompression success in patients with femoral head avascular necrosis. Arch Bone Jt Surg. 2023;11(8):517-523.
- [Google Scholar]
- Using machine learning to predict nonachievement of clinically significant outcomes after rotator cuff repair. Orthop J Sports Med. 2023;11(10)
- [Google Scholar]
- Machine learning for benchmarking adolescent idiopathic scoliosis surgery outcomes. Spine. 2023;48(16):1138-1147.
- [Google Scholar]
- Machine learning can predict level of improvement in shoulder arthroplasty. JBJS Open Access. 2021;6(1):e20.
- [Google Scholar]
- Interpretation of compound activity predictions from complex machine learning models using local approximations and shapley values. J Med Chem. 2020;63(16):8761-8777.
- [Google Scholar]
- Screening for femoral head osteonecrosis following COVID-19: is it worth it? Archives of Bone and Joint Surgery. 2023;11(12):731.
- [Google Scholar]