IntroductionRecurrent patellar dislocation is a common and debilitating patellofemoral disorder, which causes significant pain, dysfunction, decline in quality of life, cartilage damage, and potential patellofemoral arthritis1,2. It is generally acknowledged that patellar dislocation is a multifactorial mechanical condition, arising from the disruption of anatomical restraint homeostasis involving medial patellofemoral ligament injury, trochlear dysplasia, increased tibial tuberosity-trochlear groove distance, patella alta, torsional deformity of lower limbs, and genu valgus3,4.Identifying related risk factors for patellar dislocation could be helpful in screening individuals with the highest risk for occurrence and recurrence and in optimizing clinical interventional decisions. Predictive algorithms based on demographic and anatomical factors have been developed for recurrent patellar dislocation5,6. However, these models are relatively complex, require multiple imaging examinations and pose additional radiation risks and economic burdens. In addition, many known anatomical risk factors for patellar dislocation can only be eliminated by surgical correction, yet there remains no consensus on the optimal surgical strategies, given the complex etiology and the unclear pathogenesis7.Therefore, it is imperative to gain insights from potentially modifiable risk factors. Given that anatomical risk factors for patellar dislocation can change during skeletal growth and trochlear morphology can improve following patellar reduction, patellar dislocation can be considered a developmental dysplasia of the patellofemoral joint, with its recurrence potentially influenced by various metabolic processes8,9,10,11,12. Biomarkers, common reflections of metabolic processes, can be easily collected from bodily fluids at relatively low cost and may play a pivotal role in disease prediction and early diagnosis13. Current evidence has demonstrated that lower preoperative serum calcium level serves as a prognostic factor for poor outcomes after surgical treatment for patellar dislocation13, and vitamin D deficiency is associated with higher rates of primary patellar dislocation and higher risks of requiring additional surgeries14.Obesity is one of the important risk factors for patellar dislocation during adolescent development and is closely associated with lipid metabolism15,16,17. Adipose tissue and bone share common progenitor cells, namely multipotent mesenchymal stem cells in the bone marrow, which can differentiate into various cell phenotypes, including osteoblasts, chondrocytes, and adipocytes18. This lineage overlap provides a critical biological basis for the crosstalk between lipid metabolism and patellofemoral joint development. An increasing number of studies have demonstrated that lipids exert profound regulatory effects on bone and cartilage homeostasis via multiple mechanisms. First, lipid availability dictates the differentiation fate of mesenchymal stem cells. Excess lipids promote their differentiation into adipocytes while inhibiting the formation of osteoblasts and chondrocytes19. Second, dysregulated lipid metabolism induces oxidative stress and chronic low-grade inflammation in the joint microenvironment, prompting the release of pro-inflammatory cytokines, which disrupt cartilage matrix synthesis and accelerate its degradation20. Third, dysregulated lipid metabolism exacerbates joint biomechanical imbalance by promoting adipose accumulation and impairing adipokine signaling, which is an essential pathway for maintaining osteochondral cell survival and matrix turnover21. Based on these mechanistic connections, lipid metabolic factors are more likely to be risk factors for recurrent patellar dislocation. Therefore, in-depth exploration of the relationship between lipid metabolism and recurrent patellar dislocation is expected to provide novel insights for the early diagnosis and non-surgical intervention of this condition.The aim of this study was to develop explainable machine learning algorithms that can accurately discriminate patients with recurrent patellar dislocation from healthy controls based on blood lipid metabolic factors and clinical features, and to screen lipid metabolic factors associated with recurrent patellar dislocation. It was hypothesized that blood lipid metabolic factors could serve as useful discriminatory features for recurrent patellar dislocation. The framework of this study was illustrated in Fig. 1.Fig. 1Full size imageThe framework of this study.MethodsStudy cohortsAfter obtaining the approval from the Ethics Committee of our institution and the written informed consent from all participants, two independent and experienced researchers initiated this retrospective study by reviewing and screening the medical records of all patients diagnosed with recurrent patellar dislocation at our institution between January 2014 and January 2024. The inclusion criteria were as follows: (1) patients had more than 2 episodes of patellar dislocation; (2) confirmation of patellar dislocation by computed tomography or magnetic resonance imaging; and (3) patients had complete blood lipid examination data. Patients were excluded if they met one of the following exclusion criteria: (1) conditions associated with dyslipidemia, such as hypertension and coronary heart disease; (2) endocrine diseases, such as diabetes, thyroid diseases, and adrenocortical diseases; (3) common metabolic comorbidities, such as metabolic syndrome, non-alcoholic fatty liver disease, and polycystic ovary syndrome; (4) patients who had taken drugs that may affect blood lipid levels, such as lipid-lowering drugs, weight-loss drugs, and corticosteroids; (5) systemic connective tissue disorders; and (6) insufficient or poor-quality data. The exclusion of patients with metabolic comorbidities and those using lipid-lowering drugs was intended to accurately explore the direct association between baseline blood lipid levels and recurrent patellar dislocation and to avoid interference from comorbidities and drug interventions on lipid profiles.To compare the blood lipid levels between patients with recurrent patellar dislocation and healthy individuals, a control group was retrospectively selected from the individuals at our institution over the same period. The inclusion criteria were as follows: (1) normal patellofemoral joint structure verified by computed tomography or magnetic resonance imaging; (2) no history of knee injury; and (3) no history of patellofemoral joint disorders including patellar dislocation and patellofemoral pain. The same exclusion criteria were applied to the control group. Ultimately, a total of 260 patients with recurrent patellar dislocation were enrolled in the study group, and 547 healthy individuals without a history of patellar dislocation were included in the control group.Data collection and preprocessingMedical records of the study group and control group were collected. The dataset comprised 14 features categorized into two groups: demographic characteristics and blood lipid metabolic factors. Key demographic features included gender (male = 1.0, female = 0.0), age, height, weight and body mass index (BMI), while blood lipid metabolic factors included total cholesterol (TC), triglycerides (TG), high-density lipoprotein (HDL), low-density lipoprotein (LDL), very low-density lipoprotein (VLDL), apolipoprotein A1 (ApoA1), apolipoprotein B (ApoB), lipoprotein(a) (Lp(a)), and ApoA1/ApoB ratio, all of which were hypothesized to be associated with recurrent patellar dislocation. Key demographic features were included as continuous variables, allowing the models to capture potential non-linear and heterogeneous demographic-dependent patterns rather than assuming a uniform effect across the population. No missing values were observed in the dataset used for model development. For completeness of the preprocessing pipeline, we specified that if missing values were present, they would be imputed using the median of each feature, which is more robust to skewed distributions. Pearson correlation coefficients were computed to calculate the correlation and significance between each pair of feature variables and the target variable, as shown in the heatmap (Fig. 2).Fig. 2Full size imageThe heatmap of the feature correlation matrix.To assess potential redundancy among candidate variables, we performed principal component analysis (PCA) on lipid biomarkers and anthropometric variables. PCA was used to characterize the overall covariance structure and redundancy patterns among candidate variables, rather than to establish predictor independence. Although correlated features were identified, the original variables were retained to preserve clinically interpretable measurements in their native form. This choice may introduce attributional ambiguity in feature-importance and SHapley Additive exPlanations (SHAP) analyses. Therefore, retained correlated predictors should be interpreted as related contributors within a shared feature space rather than as independent biological determinants.Model development: two-layer ensemble frameworkTo take advantage of the traditional machine learning models, this study used a two-layer ensemble learning framework to discriminate between recurrent patellar dislocation and non-dislocation cases, as shown in Fig. 3. In the first layer, ten machine learning models were employed as base classifiers, including Gradient Boosting Decision Trees (GBDT), Random Forest (RF), Extra Trees (ET), Adaptive Boosting (ADA), Support Vector Machine (SVM), XGBoost, CatBoost, Logistic Regression (LR), Multinomial Naive Bayes (MNB), and K-Nearest Neighbors (KNN). These models were selected to represent a diverse range of commonly used classifier families, including linear, tree-based, boosting-based, distance-based, and probabilistic methods, thereby covering different modeling assumptions and inductive biases. GBDT is an ensemble learning method using decision trees, optimizing for deviance loss with slow learning22. RF is a bagging method that builds multiple decision trees with balanced class weights23. ET is similar to RF but uses random splits for improved diversity24. ADA is a boosting algorithm that combines weak learners sequentially to improve accuracy25. SVM is a classification algorithm maximizing the margin with balanced class weights26. XGBoost is an efficient gradient boosting implementation with low tree depth for faster learning27. CatBoost is a gradient boosting algorithm optimized for categorical features with class weight adjustment28. LR is a linear model for classification, using L1 regularization and specific class weightings29. MNB is a probabilistic classifier based on Bayes’ theorem for multiclass data, with Laplace smoothing30. KNN is a non-parametric algorithm that assigns class based on majority vote among nearest neighbors31. In the second layer, an LR model served as the meta-classifier, combining the predictions from the base classifiers to make the final classification. LR was selected as the meta-classifier due to its simplicity, interpretability, and reduced risk of overfitting when integrating correlated base learner outputs, particularly in a small-sample setting. The goal of this multi-model design was to assess whether predictive signals were robust across different model families, which is particularly important in clinical settings.Fig. 3Full size imageThe framework of the proposed two-layer ensemble learning model.Stratified five-fold cross-validation was conducted on the training set for all models, and performance was reported as mean ± standard deviation of area under the receiver operating characteristic (ROC) curve (AUC) and average precision (AP). The hyperparameters and details of all the models used in this step are shown in Table 1. Hyperparameters for all models were optimized using grid search with cross-validation on the training set. The optimal configuration was selected based on mean cross-validation AUC (and AP where applicable). The search space was defined to favor compact model structures in order to mitigate overfitting in the small-sample setting. For the stacking ensemble, out-of-fold predictions generated from the base models were used to construct the input features for the meta-classifier, ensuring that predictions for each sample were obtained from models that had not been trained on that sample. This ensemble strategy was adopted to improve robustness rather than to maximize discriminative performance of any individual model.Table 1 Models and key parameters description.Full size tableThis ensemble approach was designed to capitalize on the strengths of individual models while minimizing their weaknesses. In total, there are 260 positive samples and 547 negative samples in the dataset. Therefore, to handle the class imbalance issue, weighted loss functions were used during training. Model calibration was further assessed using calibration curves and Brier scores based on stratified five-fold out-of-fold predictions on the training set. Decision curve analysis was performed to evaluate the potential clinical utility of the model by quantifying net benefit across a range of threshold probabilities, following standard methodology.Model evaluation metricsTo evaluate the model, 70% of the dataset was used for model training, and 30% was used for testing, where the training set was used to fit the model, and the test set was used to assess its generalization capabilities. The dataset was split into training and test sets using stratified sampling based on the outcome variable to preserve class proportions. Model performance was assessed using sensitivity, specificity, F1-score, ROC curve, and the corresponding AUC, as well as precision-recall (PR) curve and AP. Sensitivity and specificity were calculated to evaluate the performance in detecting positive and negative cases, respectively. The F1-score provided a balanced measure of precision and sensitivity. ROC and AUC quantified overall discriminative performance, whereas PR and AP offered complementary evaluation of positive class prediction under class imbalance. All analyses were conducted using Python (version 3.11) and scikit-learn. Additionally, SHAP values were used to interpret the discriminative performance of the model and identify the features that contributed most to discriminating recurrent patellar dislocation32.ResultsEvaluation of feature redundancyTo evaluate feature redundancy, PCA was performed on both lipid biomarkers and anthropometric variables. As shown in Fig. 4, the first four principal components (PCs) explained 86.78% of the total variance in lipid biomarkers. While PC1 and PC2 represented major pro-atherogenic and protective lipid clusters (Fig. 4A-B), PC3 and PC4 captured unique signals from specific markers, notably Lp(a) and ApoA1/ApoB ratio (Fig. 4C). This multidimensionality justified the inclusion of all markers to ensure a comprehensive metabolic risk profile. Similarly, PCA of anthropometric variables (height, weight, and BMI) confirmed that these variables captured distinct dimensions: height and weight reflected absolute body size, while BMI represented relative adiposity. Although BMI is mathematically derived from height and weight, these clinically familiar measurements were retained in their original form to preserve interpretability. However, due to their structural dependence, their contributions in the model should not be interpreted as independent effects.Fig. 4Full size imagePCA of lipid biomarkers and feature redundancy evaluation. (A) Variance explained by principal components. (B) PC1-PC2 loading distribution. (C) Feature loadings across principal components.Performance of the modelsCross-validation results based on the training set are summarized in Table 2, and ROC and PR curves shown in Fig. 5 were generated using the held-out test set. The ROC curves depicted the ability of the models to discriminate between classes, whereas the PR curves provided complementary information on positive class discrimination. The ensemble model achieved an AUC of 0.82, which was comparable to the best-performing individual models. Among the base models, CatBoost showed the highest AUC (0.82), followed by LR (0.81) and ET (0.80), while MNB and KNN models exhibited relatively lower discriminative performance, with AUC of 0.65 and 0.66, respectively. Consistent trends were observed in the PR analysis, where the ensemble model achieved an AP of 0.71, comparable to the strongest base models.Table 2 Stratified five-fold cross-validation performance on the training dataset of the base models and the proposed ensemble model.Full size tableFig. 5Full size imageThe best ROC and PR curves for model performance evaluation. (A) ROC curves for different models with AUC. (B) PR curves for different models with AP.As shown in Fig. 6, the ensemble model achieved a sensitivity of 0.82, showing its strength in identifying positive cases. Conversely, MNB and KNN exhibited the lowest sensitivity of 0.35 and 0.39, respectively. RF stood out with the highest specificity of 0.84, while LR and KNN also showed strong specificity of 0.81 and 0.80, indicating that these models were useful in identifying negative cases. In contrast, MNB and KNN had significantly lower F1-scores of 0.35 and 0.44, respectively. These findings indicated that, although individual models possessed various strengths, the ensemble model offered a relatively robust balance with the F1-score of 0.68 and the AUC of 0.82, despite not being the top performer on certain individual metrics.Fig. 6Full size imagePerformance metrics of the base and meta classifiers.Calibration analysis based on five-fold out-of-fold predictions showed reasonable agreement between predicted probabilities and observed event rates, particularly in the low-to-moderate risk range, with a Brier score of 0.186, as shown in Fig. 7. Furthermore, in Fig. 8, the net benefit of the ensemble model was shown across a range of threshold probabilities and compared with default strategies of treating all patients or treating none. The ensemble model demonstrated favorable net benefit over clinically relevant thresholds, indicating potential clinical utility.Fig. 7Full size imageCalibration curve of the ensemble model based on five-fold out-of-fold predictions on the training set.Fig. 8Full size imageDecision curve analysis of the ensemble model.Feature importance analysisSHAP analyses were conducted on multiple representative tree-based models to assess the consistency of feature contributions across different model architectures. It should be emphasized that SHAP values reflect model-specific feature contributions and do not imply causal relationships. As shown in Fig. 9, SHAP summary plots across GBDT, XGBoost, and CatBoost models revealed broadly consistent feature contribution patterns. Each row represented a feature, and the x‑axis represented the SHAP values. Each colored point represented a sample, and the color represented the value of the feature (red and blue represented high and low values, respectively). Age and lipid metabolic factors such as LDL, VLDL, and Lp(a) were repeatedly identified as influential features, indicating model-derived importance rather than model-specific artifacts.Fig. 9Full size imageFeature importance analysis using SHAP values. (A) SHAP-based feature importance analysis of the GBDT model. (B) SHAP-based feature importance analysis of the XGBoost model. (C) SHAP-based feature importance analysis of the CatBoost model.The SHAP plots highlighted age and gender as important contributors, indicating that younger age and female gender were each feature associated with recurrent patellar dislocation. However, this did not establish robustness to collinearity, and their attribution ranking may still be affected by correlated contributors. Lipid metabolic factors, particularly LDL and VLDL, were consistently identified as contributors in the model, suggesting their relevance for discrimination. For instance, higher LDL levels were associated with positive SHAP values in discriminating recurrent patellar dislocation, whereas lower VLDL levels were associated with negative SHAP values. These results emphasized the importance of considering blood lipid metabolic factors in discriminating recurrent patellar dislocation.Contributions of age, LDL and ApoBThe SHAP values in Fig. 10A illustrated the contributions of age in discriminating recurrent patellar dislocation, with LDL levels represented by the color gradient. The SHAP values for age varied, showing a clear non-linear relationship between age and recurrent patellar dislocation. Younger subjects (below 15 years old) generally had positive SHAP values, indicating that age in this range contributed to discrimination of recurrent patellar dislocation. As age increased (15–30 years old), the SHAP values gradually decreased. However, in subjects above 30 years old, SHAP values began to increase slightly, which were associated with the discrimination output.Fig. 10Full size imageThe SHAP analysis plot depicting the interaction between features. (A) Interactions between age and LDL; (B) Interactions between LDL and ApoB.Among younger subjects (10–20 years old), those with higher LDL levels tended to have higher SHAP values, whereas lower LDL levels were associated with negative SHAP values, indicating an opposite contribution in younger subjects. Conversely, in subjects over 20 years old, LDL levels appeared more evenly distributed, especially within the age group where SHAP values were negative. This analysis highlighted the interaction effects between age and LDL, suggesting that age-specific variations and blood lipid metabolic factors should be carefully considered in discriminating recurrent patellar dislocation.The interaction between LDL and ApoB was shown in Fig. 10B by the SHAP analysis plot. A strong, non-linear relationship was evident: at an LDL level under 1.5 mmol/L, or between 2.5 mmol/L and 3.5 mmol/L, SHAP values were mainly negative, while LDL of 1.5–2.5 mmol/L or above 3.5 mmol/L led to positive SHAP values. Furthermore, the SHAP analysis plot revealed a gradient of ApoB, with higher ApoB levels clustering in the upper right quadrant where LDL concentrations were also higher.Case-specific feature contribution differencesThe SHAP force plot in Fig. 11A demonstrated the individual contributions of various features to discriminating patients without recurrent patellar dislocation. The model output (f(x) = – 0.75) was below the base value, indicating a lower likelihood compared to the baseline. Features that pushed the output towards higher values were marked in red, while those that pushed it lower were shown in blue. Among the features, TG (0.68 mmol/L), weight (74.0 kg), height (160.0 cm), ApoB (0.68 g/L), and BMI (28.91 kg/m2) had positive SHAP values, favoring the discrimination of patients without recurrent patellar dislocation. TG and ApoB, known markers for cardiovascular risks, were associated with this positive contribution, aligning with their effects in metabolic disorders. Additionally, weight and BMI, which reflected overall body composition, further contributed to higher SHAP values. Conversely, gender (1.0, representing male) and age (27 years old) had negative SHAP values. It likely reflected the reduced impact of age-related risk factors in adults.Fig. 11Full size imageSHAP force plots illustrating individual feature contributions to the discrimination of recurrent patellar dislocation. (A) SHAP force plot for patients without recurrent patellar dislocation. Positive SHAP values (red) contributed to the discrimination of non‑dislocation state. (B) SHAP force plot for patients with recurrent patellar dislocation. Positive SHAP values (red) contributed to the discrimination of recurrent patellar dislocation.Figure 11B showed a SHAP force plot of a patient with recurrent patellar dislocation. The model output (f(x) = – 0.75) was above the base value, indicating a higher likelihood compared to the baseline. Key positive contributors included height (162.0 cm), Lp(a) (111.9 mg/L), gender (0.0, representing female), and age (17 years old). Among these features, Lp(a) showed a strong contribution to discriminating patients with recurrent patellar dislocation.DiscussionThe main finding of this study was that blood lipid metabolic factors, particularly LDL, VLDL and Lp(a), combined with age and gender, enabled successful discrimination of recurrent patellar dislocation via an explainable ensemble machine learning framework. This study provides a non-invasive, interpretable approach for early risk stratification and potential non-surgical intervention for recurrent patellar dislocation.Conventional risk models for recurrent patellar dislocation rely mainly on anatomical risk factors, which are assessed via computed tomography or magnetic resonance imaging, having limitations including high cost, radiation exposure risk, and low accessibility for early screening3,4,5,6,7. Additionally, most anatomical factors are non-modifiable and require surgical correction. Previous studies have shown that some anatomical risk factors, such as trochlear morphology, can be improved by surgically correcting dislocation to restore patellofemoral joint stress9,10,11,12, suggesting that the patellofemoral joint is dynamically responsive to biomechanical and possibly metabolic environments, and that metabolic factors may influence anatomical development. By integrating metabolic and demographic factors, this study achieved discriminative performance for recurrent patellar dislocation without relying on anatomical factors.Among demographic factors, age and gender are widely recognized as key factors for recurrent patellar dislocation. Adolescents (especially those aged 14–18 years old) and females are at higher risk of patellar dislocation1,33. Danielsen et al. emphasized that familial predisposition and age-related skeletal immaturity were important reasons for the elevated risk in adolescents, while hormonal differences, such as estrogen-related soft tissue laxity, may explain the susceptibility in females33. The age-related risk pattern highlights the vulnerability during the critical period of skeletal growth and development, which is closely associated with the development of the patellofemoral joint8,12. The results of our study are consistent with existing literature1,33.Previous studies on metabolic factors have mostly focused on single macroscopic phenotypes or non-specific metabolic indicators, such as obesity, vitamin D deficiency, and abnormal serum calcium level13,14,15,16,17, lacking in-depth exploration of specific metabolic biomarkers directly involved in regulating osteochondral development. The present study thoroughly investigated the discriminative value of lipid components for recurrent patellar dislocation, confirming that blood lipid metabolic factors, especially LDL, VLDL, and Lp(a), contributed significantly to discriminating recurrent patellar dislocation. Lipids play important roles in the normal development of osteochondral tissues. However, the specific mechanism of lipid metabolism in the process of trochlear development is still unclear, which deserves further exploration.Various machine learning models based on lipid metabolism and clinical features were explored in this study to discriminate recurrent patellar dislocation. The choice of LR as the meta-classifier was motivated by a preference for robustness and interpretability over marginal performance gains that may arise from more complex meta-models. The analysis demonstrated the strength of ensemble learning in balancing various performance metrics. Although individual models possessed different strengths, the ensemble model offered a relatively robust balance with the AUC of 0.82 and performed consistently across sensitivity, specificity, and F1-score, despite not being the top performer on certain individual metrics. The integration of multiple base models into the ensemble method likely contributed to its competitive performance by leveraging the complementary strengths of individual classifiers. However, it is worth noting that while the ensemble approach was useful for methodological exploration, simpler models may be more practical for clinical application when discriminative performance is similar.This study showed that age, gender, LDL, VLDL, and Lp(a) were important contributors to discriminating recurrent patellar dislocation. Despite the statistical correlations among lipid markers identified in PCA, the SHAP analysis consistently prioritized age and gender as the key contributors. This suggests that the tree-based ensemble models used in this study are successful in handling feature redundancy and highlighting the most clinically relevant factors without inflating the importance of correlated metabolic biomarkers. This study revealed the complex interaction between age and LDL. Among younger subjects aged 10–20 years old, higher LDL levels tended to have higher SHAP values, while in subjects over 20 years old, LDL showed a trend of even distribution, suggesting that age-specific variations in lipid metabolic factors should be carefully considered in discriminating recurrent patellar dislocation.The observed relationship underscored the complex interplay between LDL and ApoB in the decision-making of the model. ApoB, a primary protein component of LDL particles34, provided a more direct measure of atherogenic lipoprotein burden compared with LDL alone. High LDL levels coupled with elevated ApoB aligned with established evidence linking these factors to cardiovascular risk35,36,37. Notably, the gradient in ApoB concentration suggested its complementary role in amplifying LDL- associated risk, highlighting the importance of considering both LDL and ApoB in risk stratification models. The addition of ApoB may improve the ability of the model to discriminate adverse outcomes by capturing variations in lipoprotein particle number and composition.The potential clinical relevance of this study lies in providing a promising approach for optimizing the clinical management of recurrent patellar dislocation. Unlike anatomical risk factors, lipid metabolic factors identified in this study are modifiable and can be easily accessed through routine blood lipid tests, which may enable surgeons to identify high-risk individuals through non-invasive screening and provide a basis for shifting the clinical intervention paradigm from passive surgical correction to active non-surgical management, such as lipid-lowering therapy, weight management, and lifestyle modification to regulate blood lipid levels, especially important for skeletally immature adolescents. Additionally, for patients at risk of recurrence, integrating lipid metabolism monitoring into follow-up protocols could provide a reference for personalized treatment decisions.This study has several limitations. First, this study focused on discriminating recurrent patellar dislocation cases from healthy controls, and the constructed model is not applicable to identifying high-risk individuals for first-time patellar dislocation events. Second, although internal validation was performed, the sample size remained modest relative to the complexity of the modeling framework, and both the training and test sets were derived from the same institution and time. Therefore, the reported performance should be interpreted as preliminary evidence of internal discriminative ability rather than definitive proof of generalizability. Third, the exclusion of patients with metabolic comorbidities and those using lipid-lowering drugs introduced selection bias in the cohort. The model is currently only suitable for recurrent patellar dislocation patients without metabolic comorbidities and lipid-lowering therapy, limiting its external validity in real-world clinical settings. Fourth, although associations between lipid metabolic factors and recurrent patellar dislocation were identified, causal relationships and underlying biological mechanisms remain to be elucidated. The relatively high SHAP ranking of age and gender should also be interpreted cautiously, as attribution magnitude may be influenced by correlated feature structure and does not in itself demonstrate robustness to redundancy. Fifth, although the model can generate individualized discriminative outputs, no clinically validated decision threshold was established for routine clinical use based solely on the current data. Future studies are needed to define clinical thresholds based on prospective validation and clinical consequences of false-positive and false-negative classifications.ConclusionsBlood lipid metabolic factors, particularly LDL, VLDL and Lp(a), combined with age and gender, enabled successful discrimination of recurrent patellar dislocation via an explainable ensemble machine learning framework. This study provides a non-invasive, interpretable approach for early risk stratification and potential non-surgical intervention for recurrent patellar dislocation.Data availabilityThe data that support the findings of this study are not openly available due to reasons of sensitivity and are available from the corresponding author upon reasonable request.ReferencesSanders, T. L. et al. Incidence of first-time lateral patellar dislocation: A 21-year population-based study. Sports Health. 10 (2), 146–151. https://doi.org/10.1177/1941738117725055 (2018).Article PubMed Google Scholar Sanders, T. L. et al. Patellofemoral arthritis after lateral patellar dislocation: A matched population-based analysis. Am. J. Sports Med. 45 (5), 1012–1017. https://doi.org/10.1177/0363546516680604 (2017).Article PubMed Google Scholar Huntington, L. S., Webster, K. E., Devitt, B. M., Scanlon, J. P. & Feller, J. A. Factors associated with an increased risk of recurrence after a first-time patellar dislocation: A systematic review and meta-analysis. Am. J. Sports Med. 48 (10), 2552–2562. https://doi.org/10.1177/0363546519888467 (2020).Article PubMed Google Scholar Xu, Z., Zhang, H., Chen, J., Mohamed, S. I. & Zhou, A. Femoral anteversion is related to tibial tubercle-trochlear groove distance in patients with patellar dislocation. Arthroscopy 36 (4), 1114–1120. https://doi.org/10.1016/j.arthro.2019.10.033 (2020).Article PubMed Google Scholar Ling, D. I. et al. Development of a multivariable model based on individual risk factors for recurrent lateral patellar dislocation. J. Bone Joint Surg. Am. 103 (7), 586–592. https://doi.org/10.2106/JBJS.20.00020 (2021).Article PubMed Google Scholar Xu, C. et al. Predicting the probability of recurrence based on individualized risk factors after primary lateral patellar dislocation treated nonoperatively. Arthroscopy 40 (5), 1602–1609e1. https://doi.org/10.1016/j.arthro.2023.10.028 (2024).Article PubMed Google Scholar Liu, J. N. et al. Patellar instability management: A survey of the international patellofemoral study group. Am. J. Sports Med. 46 (13), 3299–3306. https://doi.org/10.1177/0363546517732045 (2018).Article PubMed Google Scholar Pruneski, J., O’Mara, L., Perrone, G. S. & Kiapour, A. M. Changes in anatomic risk factors for patellar instability during skeletal growth and maturation. Am. J. Sports Med. 50 (9), 2424–2432. https://doi.org/10.1177/03635465221102917 (2022).Article PubMed Google Scholar Fu, K., Duan, G., Liu, C., Niu, J. & Wang, F. Changes in femoral trochlear morphology following surgical correction of recurrent patellar dislocation associated with trochlear dysplasia in children. Bone Joint J. 100-B, 811–821. https://doi.org/10.1302/0301-620X.100B6.BJJ-2017-1295.R1 (2018).Article CAS PubMed Google Scholar Benoit, B. et al. Long-term outcome of surgically-treated habitual patellar dislocation in children with coexistent patella alta. Minimum follow-up of 11 years. J. Bone Joint Surg. Br. 89, 1172–1177. https://doi.org/10.1302/0301-620X.89B9.19065 (2007).Article CAS PubMed Google Scholar Yang, G. et al. The dysplastic trochlear sulcus due to the insufficient patellar stress in growing rats. BMC Musculoskelet. Disord. 20 (1), 411. https://doi.org/10.1186/s12891-019-2802-y (2019).Article PubMed PubMed Central Google Scholar Hao, K. et al. Morphological improvement of the epiphyseal plate and trochlea after surgical correction in skeletally immature patients with patellar dislocation and trochlear dysplasia. Am. J. Sports Med. 53 (2), 372–384. https://doi.org/10.1177/03635465241301775 (2025).Article PubMed Google Scholar Qiao, Y. et al. Preoperative serum calcium could be a prognostic factor for surgical treatment of recurrent patellar dislocation: a retrospective study. BMC Musculoskelet. Disord. 23 (1), 578. https://doi.org/10.1186/s12891-022-05527-y (2022).Article CAS PubMed PubMed Central Google Scholar Chang, K. et al. A Diagnosis of vitamin D deficiency is associated with increased rates of primary patellar instability and need for recurrent surgical stabilization. Sports Health. 16 (3), 465–472. https://doi.org/10.1177/19417381231172726 (2024).Article PubMed Google Scholar He, H., Liu, W., Teraili, A., Wang, X. & Wang, C. Correlation between flat foot and patellar instability in adolescents and analysis of related risk factors. J. Orthop. Surg. (Hong Kong). 31 (1), 10225536231171057. https://doi.org/10.1177/10225536231171057 (2023).Article PubMed Google Scholar Kurosawa, K. et al. Patellar dislocation in Kabuki syndrome. Am. J. Med. Genet. 108 (2), 160–163. https://doi.org/10.1002/ajmg.10247 (2002).Article PubMed Google Scholar Nussbaumerova, B. & Rosolova, H. Obesity and dyslipidemia. Curr. Atheroscler Rep. 25 (12), 947–955. https://doi.org/10.1007/s11883-023-01167-2 (2023).Article CAS PubMed Google Scholar Tian, L. & Yu, X. Lipid metabolism disorders and bone dysfunction–interrelated and mutually regulated (review). Mol. Med. Rep. 212 (1), 783–794. https://doi.org/10.3892/mmr.2015.3472 (2015).Article CAS Google Scholar van Gastel, N. et al. Lipid availability determines fate of skeletal progenitor cells via SOX9. Nature 579 (7797), 111–117. https://doi.org/10.1038/s41586-020-2050-1 (2020).Article CAS PubMed PubMed Central Google Scholar Medina-Luna, D. et al. Hyperlipidemic microenvironment conditionates damage mechanisms in human chondrocytes by oxidative stress. Lipids Health Dis. 16 (1), 114. https://doi.org/10.1186/s12944-017-0510-x (2017).Article CAS PubMed PubMed Central Google Scholar Mocanu, V., Timofte, D. V., Zară-Dănceanu, C. M., Labusca, L. & Obesity metabolic syndrome, and osteoarthritis require integrative understanding and management. Biomedicines 12 (6), 1262. https://doi.org/10.3390/biomedicines12061262 (2024).Article CAS PubMed PubMed Central Google Scholar Friedman, J. H. Greedy function approximation: A gradient boosting machine. Ann. Stat. 29 (5), 1189–1232. https://doi.org/10.1214/aos/1013203451 (2001).Article Google Scholar Breiman, L. E. O. Random forests. Mach. Learn. 45, 5–32. https://doi.org/10.1023/A:1010933404324 (2001).Article Google Scholar Ghazwani, M. & Begum, M. Y. Computational intelligence modeling of hyoscine drug solubility and solvent density in supercritical processing: gradient boosting, extra trees, and random forest models. Sci. Rep. 13 (1), 10046. https://doi.org/10.1038/s41598-023-37232-8 (2023).Article CAS PubMed PubMed Central Google Scholar Zheng, Z. & Yang, Y. Adaptive boosting for domain adaptation: Toward robust predictions in scene segmentation. IEEE Trans. Image Process. 31, 5371–3582. https://doi.org/10.1109/TIP.2022.3195642 (2022).Article PubMed Google Scholar Noble, W. S. What is a support vector machine? Nat. Biotechnol. 24 (12), 1565–1567. https://doi.org/10.1038/nbt1206-1565 (2005).Article CAS Google Scholar Campi, R., De Santis, A., Colombo, P., Scarpazza, P. & Masseroli, M. Machine learning-based forecast of Helmet-CPAP therapy failure in acute respiratory distress syndrome patients. Comput. Methods Programs Biomed. 260, 108574. https://doi.org/10.1016/j.cmpb.2024.108574 (2024).Article PubMed Google Scholar Hancock, J. T. & Khoshgoftaar, T. M. CatBoost for big data: An interdisciplinary review. J. Big Data. 7 (1), 94. https://doi.org/10.1186/s40537-020-00369-8 (2020).Article PubMed PubMed Central Google Scholar Stoltzfus, J. C. Logistic regression: A brief primer. Acad. Emerg. Med. 18 (10), 1099–1104. https://doi.org/10.1111/j.1553-2712.2011.01185.x (2011).Article PubMed Google Scholar Li, X. et al. An intelligent system for classifying patient complaints using machine learning and natural language processing: Development and validation study. J. Med. Internet Res. 27, e55721. https://doi.org/10.2196/55721 (2025).Article PubMed PubMed Central Google Scholar Gu, C. et al. Data-model interactive Rul prediction of stochastic degradation devices with multiple uncertainty quantification and multi-sensor information fusion. ISA Trans. 157, 293–305. https://doi.org/10.1016/j.isatra.2024.12.024 (2025).Article PubMed Google Scholar Chen, H., Lundberg, S. & Lee, S-I. Explaining models by propagating Shapley values of local components. Explain AI Healthc. Med. 914, 261–270. https://doi.org/10.1007/978-3-030-53352-6_24 (2021).Article Google Scholar Danielsen, O. et al. Familial association and epidemilogical factors as risk factors for developing first time and recurrent patella dislocation: A systematic review and best knowledge synthesis of present literature. Knee Surg. Sports Traumatol. Arthrosc. 31 (9), 3701–3733. https://doi.org/10.1007/s00167-022-07265-z (2023).Article PubMed Google Scholar Xiao, L. et al. The LDL/ApoB ratio predicts cardiovascular and all-cause mortality in the general population. Lipids Health Dis. 22 (1), 104. https://doi.org/10.1186/s12944-023-01869-1 (2023).Article CAS PubMed PubMed Central Google Scholar Silbernagel, G. et al. The LDL apolipoprotein B-to-LDL cholesterol ratio: Association with cardiovascular mortality and a biomarker of small. Dense LDLs Biomed. 10 (6), 1302. https://doi.org/10.3390/biomedicines10061302 (2022).Article CAS PubMed PubMed Central Google Scholar Drexel, H. et al. The LDL-C/ApoB ratio predicts major cardiovascular events in patients with established atherosclerotic cardiovascular disease. Atherosclerosis 329, 44–49. https://doi.org/10.1016/j.atherosclerosis.2021.05.010 (2021).Article CAS PubMed Google Scholar Kaneva, A. M., Potolitsyna, N. N. & Bojko, E. R. Usefulness of the LDL-C/apoB ratio in the overall evaluation of atherogenicity of lipid profile. Arch. Physiol. Biochem. 123 (1), 16–22. https://doi.org/10.1080/13813455.2016.1195411 (2017).Article CAS PubMed Google Scholar Download referencesAcknowledgementsNot applicable.FundingThis work was supported by the Natural Science Research Program of the Education Department of Hebei Province [Grant Number: QN2025378], Hebei Medical University Postdoctoral Fund (Grant Number: 30705010045), Natural Science Foundation of Hebei Province [Grant Number: H2021206162], and Hebei Province Project for Overseas-educated scholars [Grant Number: C20210351].Author informationAuthor notesChunjiao Dong and Kuo Hao contributed equally to this paper.Authors and AffiliationsDepartment of Medical Imaging, Hebei Medical University, Shijiazhuang, 050017, Hebei, ChinaChunjiao DongDepartment of Orthopaedic Surgery, Hebei Medical University Third Hospital, Shijiazhuang, 050051, Hebei, ChinaKuo Hao, Fei Wang, Juncai Wang, Jiahuan Wang & Huijun KangHebei Medical University Clinical Medicine Postdoctoral Research Station (Hebei Medical University Third Hospital), Shijiazhuang, 050051, Hebei, ChinaKuo HaoDepartment of Obstetrics, The Fourth Hospital of Hebei Medical University, Shijiazhuang, 050011, Hebei, ChinaRui ZhengAuthorsChunjiao DongView author publicationsSearch author on:PubMed Google ScholarKuo HaoView author publicationsSearch author on:PubMed Google ScholarFei WangView author publicationsSearch author on:PubMed Google ScholarJuncai WangView author publicationsSearch author on:PubMed Google ScholarJiahuan WangView author publicationsSearch author on:PubMed Google ScholarRui ZhengView author publicationsSearch author on:PubMed Google ScholarHuijun KangView author publicationsSearch author on:PubMed Google ScholarContributionsCD: Writing-original draft, Visualization, Software, Methodology, Investigation, Formal analysis. KH: Writing-original draft, Resources, Methodology, Conceptualization. FW: Writing-review & editing, Supervision, Conceptualization. JW: Data curation, Formal analysis, Validation. JW: Data curation, Formal analysis, Validation. RZ: Writing–review and editing, Conceptualization, Investigation. HK: Writing-review and editing, Conceptualization, Project administration.Corresponding authorsCorrespondence to Rui Zheng or Huijun Kang.Ethics declarationsEthics approval and consent to participateThis study has the formal ethical approval from the Ethics Committee of Hebei Medical University Third Hospital (No. 2023-002-01), and was conducted in full compliance with China’s Ethical Review Measures for Biomedical Research Involving Humans and the Declaration of Helsinki. All participants or their legal guardians provided written informed consent.Consent for publicationWritten informed consent was obtained from all patients to authorize the publication of their data.Competing interestsThe authors declare no competing interests.Clinical trial numberNot applicable.Additional informationPublisher’s noteSpringer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.Rights and permissionsOpen Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/.Reprints and permissionsAbout this article