Development and Validation of a Multi-Modal Ensemble Model for Predicting Progression in Idiopathic Scoliosis Authors: Hideyuki Arima, Shota Ichikawa, Yu Yamato, Kei Watanabe, Haruki Ueda, Shinji Takahashi, Naobumi Hosogane, Takumi Takeuchi, Hiroki Oba, Masashi Okamoto, Manami Umezu, Yohan Kondo, Shoji Seki Published: 2026 ABSTRACT AbstractStudy DesignRetrospective Cohort Study.ObjectivesAccurate prediction of curve progression in idiopathic scoliosis at the initial visit would facilitate clinical decision-making. In our previous study, progression was predicted using deep learning based on frontal whole-spine radiographs. This study aimed to improve prediction accuracy using a multimodal ensemble model integrating frontal and lateral radiographs with clinical information.MethodsThis multicenter retrospective cohort study included 527 patients with idiopathic scoliosis. Based on the change in Cobb angle over two years, patients were classified into progression (≥10°), non-progression (≤5°), and borderline (6–9°) groups. A total of 471 patients (274 progression and 197 non-progression) were analyzed. Input data included initial whole-spine frontal and lateral radiographs and clinical variables (age, sex, Risser sign, and baseline Cobb angle). Deep learning models (Vision Transformer, Swin Transformer, and ConvNeXtV2) and machine learning models (logistic regression, support vector machine [SVM], and random forest) were applied, generating nine models. Predicted probabilities were integrated using weighted averaging. Performance was evaluated using repeated stratified 10-fold cross-validation with the area under the receiver operating characteristic curve (AUC).ResultsMean age was 12.7 ± 1.8 years, and 433 patients (91.9%) were female. Mean initial Cobb angle was 27.8 ± 10.1°. The AUCs based on averaged predicted probabilities within each modality were 0.791 for frontal radiographs, 0.767 for lateral radiographs, and 0.721 for clinical features. The weighted ensemble achieved highest performance (AUC 0.819, 95% CI: 0.816–0.821), significantly outperforming other ensemble methods (P < 0.001).ConclusionsA multimodal ensemble model integrating radiographs and clinical data improved prediction of idiopathic scoliosis progression compared with single-modality models.Keywords: idiopathic scoliosis, curve progression, deep learning, ensemble model, whole spine radiograph Introduction Idiopathic scoliosis is the most common spinal deformity in adolescents and is characterized by a three-dimensional curvature of the spine. 1 Although many patients remain stable during growth, some experience progressive curve deterioration that may lead to severe deformity and the need for surgical intervention.2,3 Early identification of patients at high risk of progression is therefore crucial for determining appropriate follow-up intervals and treatment strategies, including bracing. 4 Several clinical predictors of curve progression have been reported, including age, sex, skeletal maturity, and initial Cobb angle.4-6 Among these, the Risser sign and baseline Cobb angle are widely used indicators of growth potential and progression risk.4,7 However, these clinical parameters alone often have limited predictive accuracy, and reliable prediction of curve progression at the initial visit remains challenging. Recent advances in artificial intelligence have enabled the application of deep learning techniques to medical imaging analysis.8-10 Several studies have explored the use of deep learning models for scoliosis assessment, including automated Cobb angle measurement and progression prediction using spinal radiographs.11-14 In our previous study, we attempted to predict curve progression using deep learning based solely on frontal whole-spine radiographs. 15 Although this approach demonstrated promising results, relying on a single imaging modality may not fully capture the complex factors associated with scoliosis progression. Integrating multiple data modalities may provide complementary information and improve predictive performance. In clinical practice, physicians evaluate both frontal and lateral radiographs together with clinical information such as age, sex, skeletal maturity, and curve magnitude when assessing progression risk.3-5,16 Therefore, combining imaging features with clinical variables may provide a more comprehensive prediction model. In addition, scoliosis progression is influenced by complex interactions among spinal morphology, growth potential, and biomechanical factors.3,17 Radiographic parameters provide information on curve magnitude and alignment, whereas clinical variables reflect growth status and maturity.4,5 In the context of adolescent idiopathic scoliosis (AIS), combining frontal and lateral radiographs with clinical variables may provide a more comprehensive representation of both structural deformity and biological growth potential, potentially leading to more accurate prediction models. The aim of this study was to develop and evaluate a multimodal ensemble model integrating frontal and lateral radiographs with clinical information to predict curve progression in idiopathic scoliosis. Materials and Methods Study Design and Patients This study was approved by the institutional review boards of all participating institutions, and the requirement for informed consent was waived due to the retrospective nature of the study. The study was conducted in accordance with the ethical principles of the Declaration of Helsinki. This multicenter retrospective cohort study included patients with AIS treated at six institutions. A total of 527 patients with available whole-spine radiographs and clinical information were initially identified. Curve progression was defined based on the change in Cobb angle over a two-year period. Patients were classified into three groups: progression (increase ≥10°), non-progression (increase ≤5°), and borderline (6–9°). The ≥10° threshold was selected to establish clearly distinguishable outcome groups and reduce label ambiguity related to Cobb angle measurement variability, postural variation, and temporal variation in flexible AIS curves. Patients with an intermediate increase of 6–9° were excluded because this range may include both true mild progression and measurement-related variation, which could reduce label reliability for supervised model development. Finally, 471 patients (274 progression and 197 non-progression) were included in the study (Figure 1). Study flow diagram of patient selection Conservative Treatment Conservative treatment strategies were determined according to standard clinical practice at each participating institution. Brace treatment was generally recommended for skeletally immature patients with Cobb angles ≥25°, particularly those with Risser sign 0–II and Cobb angles between 25° and 40°.4,16,18 Bracing was typically indicated for premenarcheal patients or those within one year after menarche. Discontinuation of bracing was considered after skeletal maturity was achieved and no significant curve progression was observed, with a gradual reduction in daily brace wear time.16,19 Brace treatment mainly consisted of underarm thoracolumbosacral orthoses, such as Boston-type braces and active correction braces. Information regarding brace treatment status and daily brace-wearing duration was collected where available. Clinical and Radiographic Data Input data consisted of initial whole-spine frontal and lateral radiographs and clinical information including age, sex, Risser sign, and baseline Cobb angle. These clinical variables were selected a priori based on established predictors of AIS progression reported in previous studies and their consistent availability across all participating institutions.3-5 Radiographs were obtained at the first visit before treatment. Cobb angles were measured using the standard method by experienced spine surgeons. Data Preprocessing All radiographic images were anonymized and uniformly processed prior to analysis. Image preprocessing included contrast-limited adaptive histogram equalization (CLAHE), zero-padding along the shorter side to generate square images, and pixel intensity normalization using the mean and standard deviation of the ImageNet-pretrained models. Clinical variables were preprocessed prior to model development. Continuous variables (age and Cobb angle) were standardized using z-score normalization, while categorical variables were encoded with binary encoding for sex and ordinal encoding for the Risser sign (0–V). Model Development Deep learning and conventional machine learning models were developed for each modality. For image-based modalities, Vision Transformer (ViT), 20 Swin Transformer (SwinT), 21 and ConvNeXtV2 22 architectures were employed. ViT and SwinT use attention mechanisms to identify relationships among different regions of a radiograph, whereas ConvNeXtV2 uses convolutional operations to extract hierarchical image features. All models were initialized with ImageNet-pretrained weights. Input images were resized to 384 × 384 pixels in accordance with the input requirements of the pretrained models. Data augmentation was applied only to the training set and included random horizontal flipping with a probability of 0.5 and random cropping to the target input size, with a scaling range of 50% to 100% of the original image area and a fixed aspect ratio of 1:1. Horizontal flipping was used to increase image diversity and reduce overfitting; no data augmentation was applied to the validation or test folds. A shared classification head was appended to each backbone, consisting of LayerNorm, a linear layer projecting the backbone feature representation to a 512-dimensional embedding, GELU activation, dropout (P = 0.5), and a final linear layer producing binary classification outputs. Models were trained using the AdamW optimizer (learning rate = 1 × 10-5, weight decay = 1 × 10-3) with a batch size of 32 and a weighted binary cross-entropy loss. Class weights were determined based on the inverse frequency of each class in the training set. A cosine annealing learning rate schedule with a five-epoch warm-up was applied. Early stopping was implemented based on validation loss, with training terminated after five consecutive epochs without improvement. The maximum number of training epochs was set to 100. For the clinical modality, age, sex, Risser sign, and baseline Cobb angle were selected a priori based on established predictors of AIS progression and their consistent availability across all participating institutions. No automated or data-driven feature-selection procedure was performed before model development. Machine learning models including logistic regression, support vector machine (SVM), and random forest were constructed using these four clinical variables.23-25 Logistic regression estimates progression probability based on a linear combination of the clinical variables, whereas the SVM identifies a decision boundary that separates progression from non-progression cases. Random forest combines predictions from multiple decision trees to capture potentially nonlinear relationships and interactions among variables. Hyperparameters for all machine learning models were optimized using Optuna. In total, nine models were generated across the three modalities (frontal radiographs, lateral radiographs, and clinical features). Ensemble Model Predicted probabilities of curve progression from each model were integrated to construct a multimodal ensemble model. Four ensemble strategies—weighted averaging, logistic regression, SVM, and random forest—were systematically compared by integrating the predicted probabilities from the nine individual models. The overall architecture of the multimodal ensemble model is illustrated in Figure 2. Specifically, frontal and lateral radiographs were analyzed independently using three deep learning models (ViT, SwinT, and ConvNeXtV2), while clinical variables were analyzed using logistic regression, SVM, and random forest models. The predicted probabilities generated by these nine individual models were subsequently combined using the ensemble methods to generate the final progression prediction. Multimodal ensemble model architecture. Three modalities (frontal radiographs, lateral radiographs, and clinical features) were analyzed using deep learning and machine learning models. The predicted probabilities from nine models were integrated using ensemble methods, primarily weighted averaging, to generate the final prediction Statistical Analysis Baseline characteristics were compared between the progression and non-progression groups. Continuous variables were compared using the Mann–Whitney U test, and categorical variables were compared using the chi-square test. Associations of brace treatment status and brace-wearing duration with progression status were also evaluated using the chi-square test. For image-based deep learning models, model performance was evaluated using repeated stratified 10-fold cross-validation with 10 repetitions. In each repetition, the entire dataset was divided into 10 mutually exclusive folds while preserving the proportion of progression and non-progression cases in each fold. For each outer cross-validation split, one fold was held out as the test set, one of the remaining folds was used as the validation set for early stopping and model selection, and the other eight folds were used for model training. This procedure ensured that test predictions for each patient were generated by a model that had not been trained or optimized using that patient’s data. The complete 10-fold procedure was repeated 10 times using different data partitions to assess the stability of model performance. For machine learning models, nested cross-validation was employed within each outer training fold. Specifically, hyperparameter optimization was performed using inner stratified 10-fold cross-validation, whereas the outer repeated stratified 10-fold cross-validation was used for performance evaluation to prevent information leakage and ensure unbiased performance estimates. This procedure provided internal validation using the present dataset; no independent external dataset was used for model evaluation. The primary outcome measure was the area under the receiver operating characteristic curve (AUC). Mean AUC values and 95% confidence intervals (CIs) were calculated across all cross-validation iterations. The Friedman–Nemenyi test was performed on the AUC values obtained from the repeated cross-validation iterations. A P-value < 0.05 was considered statistically significant. Sensitivity, specificity, accuracy, and F1 score were also calculated to provide a comprehensive evaluation of model performance. To assess the stability of model performance, the distribution of AUC values across repeated cross-validation was visualized using boxplots. All analyses were performed using Python (version 3.10). Results Patient Characteristics A total of 471 patients with AIS were included in the final analysis, comprising 274 progression cases and 197 non-progression cases. The mean age was 12.7 ± 1.8 years, and 433 patients (91.9%) were female. The Risser sign distribution was as follows: Risser sign 0, 166 (35.2%); I, 60 (12.7%); II, 94 (20.0%); III, 44 (9.3%); IV, 105 (22.3%); and V, 2 (0.4%) (Table 1). The mean baseline Cobb angle was 27.8 ± 10.1°, indicating moderate curve severity at presentation. Table 1.Patient Characteristics According to Progression StatusVariableTotal (n = 471)Progression (n = 274)Non-progression (n = 197)P ValueAge (years)12.7 ± 1.812.5 ± 1.713.0 ± 1.90.011Female, n (%)433 (91.9%)249 (90.9%)184 (93.4%)0.412Male, n (%)38 (8.1%)25 (9.1%)13 (6.6%)​Baseline Cobb angle (°)27.8 ± 10.130.3 ± 10.524.2 ± 8.3<0.001Risser sign 0, n (%)166 (35.2%)119 (43.4%)47 (23.9%)<0.001Risser sign I, n (%)60 (12.7%)28 (10.2%)32 (16.2%)​Risser sign II, n (%)94 (20.0%)52 (19.0%)42 (21.3%)​Risser sign III, n (%)44 (9.3%)25 (9.1%)19 (9.6%)​Risser sign IV, n (%)105 (22.3%)49 (17.9%)56 (28.4%)​Risser sign V, n (%)2 (0.4%)1 (0.4%)1 (0.5%)​Values are presented as mean ± standard deviation or number (%). P values were calculated using the Mann–Whitney U test for continuous variables and the chi-square test for categorical variables. Patient Characteristics According to Progression Status Values are presented as mean ± standard deviation or number (%). P values were calculated using the Mann–Whitney U test for continuous variables and the chi-square test for categorical variables. Compared with the non-progression group, patients in the progression group were significantly younger (12.5 ± 1.7 vs. 13.0 ± 1.9 years, P = 0.011) and had larger baseline Cobb angles (30.3 ± 10.5° vs. 24.2 ± 8.3°, P < 0.001). The distribution of Risser signs also differed significantly between groups (P < 0.001), with a higher proportion of Risser sign 0 patients in the progression group (43.4% vs. 23.9%) and a higher proportion of Risser sign IV patients in the non-progression group (28.4% vs. 17.9%). No significant difference in sex distribution was observed between the groups (P = 0.412). The clinical variables used for prediction included age, sex, Risser sign, and baseline Cobb angle. These variables were analyzed together with frontal and lateral radiographic images in the multimodal prediction framework (Table 1). Brace treatment information was available for 415 of the 471 patients (88.1%). Brace treatment was more frequently observed in the progression group than in the non-progression group (145/226 [64.2%] vs. 101/189 [53.4%], P = 0.035; Supplemental Table 1). Detailed brace-wearing duration data were available for 400 patients, and brace-wearing duration was also significantly associated with progression status (P = 0.008; Supplemental Table 2). Performance of Single-Modality Models The predictive performance of models using individual modalities is summarized in Table 2. For frontal radiographs, deep learning models demonstrated moderate performance, with AUC values of 0.766 ± 0.024 for ViT, 0.756 ± 0.052 for SwinT, and 0.764 ± 0.041 for ConvNeXtV2. Simple averaging of the frontal models improved performance to 0.791 ± 0.022. For lateral radiographs, model performance was slightly lower than that for frontal radiographs. The best single model achieved an AUC of 0.746 ± 0.018, and simple averaging improved the AUC to 0.767 ± 0.015. When only clinical variables were used, logistic regression achieved an AUC of 0.719 ± 0.004, while SVM and random forest models showed similar performance (Table 2). Table 2.Performance of Individual Models Across ModalitiesModalityModelAUCSensitivitySpecificityPPVNPVAccuracyF1 scoreFrontal radiographViT0.766 ± 0.0240.734 ± 0.0330.696 ± 0.0520.772 ± 0.0280.653 ± 0.0210.718 ± 0.0170.751 ± 0.016SwinT0.756 ± 0.0520.659 ± 0.1290.753 ± 0.0850.791 ± 0.0370.623 ± 0.0550.698 ± 0.0510.711 ± 0.089ConvNeXtV20.764 ± 0.0410.770 ± 0.0450.666 ± 0.0810.764 ± 0.0400.676 ± 0.0510.726 ± 0.0370.766 ± 0.030Simple averaging0.791 ± 0.0220.753 ± 0.0540.716 ± 0.0550.788 ± 0.0270.679 ± 0.0390.737 ± 0.0230.769 ± 0.026Lateral radiographViT0.746 ± 0.0180.729 ± 0.0620.677 ± 0.0540.759 ± 0.0200.647 ± 0.0410.707 ± 0.0220.742 ± 0.029SwinT0.715 ± 0.0560.714 ± 0.0450.639 ± 0.0680.734 ± 0.0420.616 ± 0.0520.682 ± 0.0440.723 ± 0.038ConvNeXtV20.742 ± 0.0380.734 ± 0.0870.665 ± 0.1040.758 ± 0.0440.649 ± 0.0590.705 ± 0.0330.741 ± 0.040Simple averaging0.767 ± 0.0150.751 ± 0.0430.691 ± 0.0520.773 ± 0.0230.668 ± 0.0260.726 ± 0.0150.761 ± 0.018Clinical featuresLogistic regression0.719 ± 0.0040.763 ± 0.0480.588 ± 0.0440.721 ± 0.0110.644 ± 0.0320.690 ± 0.0120.741 ± 0.019SVM0.707 ± 0.0080.764 ± 0.0380.573 ± 0.0440.714 ± 0.0130.637 ± 0.0210.684 ± 0.0080.737 ± 0.013Random forest0.709 ± 0.0070.683 ± 0.0580.674 ± 0.0590.746 ± 0.0180.607 ± 0.0260.679 ± 0.0110.711 ± 0.023Simple averaging0.721 ± 0.0040.677 ± 0.0900.674 ± 0.0850.746 ± 0.0260.606 ± 0.0390.676 ± 0.0180.706 ± 0.039Values are presented as mean ± standard deviation obtained from repeated cross-validation. Simple averaging refers to the unweighted mean of the predicted probabilities from the three models within each modality. Performance of Individual Models Across Modalities Values are presented as mean ± standard deviation obtained from repeated cross-validation. Simple averaging refers to the unweighted mean of the predicted probabilities from the three models within each modality. Performance of Multimodal Ensemble Models The performance of ensemble models integrating multiple modalities is shown in Table 3. The combination of frontal radiographs, lateral radiographs, and clinical variables, based on nine individual models, demonstrated improved predictive performance compared with single-modality models. Among the four ensemble approaches, weighted averaging of the predicted probabilities from the nine models achieved the best overall performance, with an AUC of 0.819 ± 0.004, sensitivity of 0.766 ± 0.047, specificity of 0.757 ± 0.045, accuracy of 0.762 ± 0.011, and F1 score of 0.789 ± 0.018 (Table 3). Other ensemble methods showed slightly lower AUCs, including random forest (0.810 ± 0.011), SVM (0.782 ± 0.033), and logistic regression (0.779 ± 0.041) (Table 3). The ROC curves demonstrated that the weighted averaging model consistently outperformed the other ensemble approaches across most thresholds (Figure 3). In addition, boxplot analysis of AUC values across repeated cross-validation showed that the weighted averaging model achieved higher performance with lower variability compared to the other methods, indicating more stable and robust predictions (Figure 4). The AUC of the weighted averaging model was significantly higher than those of the SVM and logistic regression models (P < 0.01 and P < 0.001, respectively). Table 3.Performance of Ensemble Models Across Modality CombinationsInput modalitiesEnsemble methodAUCSensitivitySpecificityAccuracyF1 scoreFrontal + Lateral + ClinicalWeighted averaging0.819 ± 0.0040.766 ± 0.0470.757 ± 0.0450.762 ± 0.0110.789 ± 0.018Logistic0.779 ± 0.0410.774 ± 0.0770.692 ± 0.0850.740 ± 0.0290.775 ± 0.036SVM0.782 ± 0.0330.761 ± 0.0730.688 ± 0.1110.730 ± 0.0270.766 ± 0.028Random forest0.810 ± 0.0110.738 ± 0.0540.770 ± 0.0360.751 ± 0.0180.775 ± 0.026Frontal + LateralWeighted averaging0.796 ± 0.0140.764 ± 0.0350.720 ± 0.0380.745 ± 0.0120.777 ± 0.015Logistic0.754 ± 0.0260.791 ± 0.0550.643 ± 0.0780.729 ± 0.0120.772 ± 0.015SVM0.757 ± 0.0510.746 ± 0.1740.657 ± 0.1660.709 ± 0.0540.737 ± 0.103Random forest0.787 ± 0.0190.777 ± 0.0510.701 ± 0.0560.745 ± 0.0190.780 ± 0.022Frontal + ClinicalWeighted averaging0.805 ± 0.0140.771 ± 0.0580.726 ± 0.0740.752 ± 0.0140.783 ± 0.019Logistic0.712 ± 0.0550.633 ± 0.1790.717 ± 0.1280.668 ± 0.0590.676 ± 0.107SVM0.683 ± 0.0820.653 ± 0.2120.641 ± 0.2320.648 ± 0.0590.668 ± 0.102Random forest0.799 ± 0.0190.740 ± 0.0660.741 ± 0.0650.740 ± 0.0170.767 ± 0.026Lateral + ClinicalWeighted averaging0.809 ± 0.0080.693 ± 0.0550.812 ± 0.0540.743 ± 0.0140.757 ± 0.023Logistic0.775 ± 0.0300.753 ± 0.0760.694 ± 0.0780.728 ± 0.0310.762 ± 0.038SVM0.789 ± 0.0190.711 ± 0.0750.759 ± 0.0650.731 ± 0.0260.753 ± 0.037Random forest0.792 ± 0.0140.682 ± 0.0420.805 ± 0.0420.733 ± 0.0120.748 ± 0.019Values are presented as mean ± standard deviation. Performance of Ensemble Models Across Modality Combinations Values are presented as mean ± standard deviation. Receiver operating characteristic (ROC) curves of ensemble models for predicting progression. The weighted averaging model showed the highest performance Boxplot of AUC values across cross-validation. The weighted averaging model showed higher performance with lower variability than other methods Contribution of Each Model and Modality The contributions of the individual models and modalities to the weighted ensemble are illustrated in Figures 5 and 6. Figure 5 demonstrates that the optimized ensemble weights were distributed across multiple individual models rather than concentrated on a single model. Among the individual models, the logistic regression model using clinical variables contributed the largest weight, followed by several deep learning models based on frontal and lateral radiographs. At the modality level, clinical variables contributed the largest weight (mean 36%), followed by frontal radiographs (33%) and lateral radiographs (31%) (Figure 6). Although clinical variables received the highest weight, all three modalities contributed substantially to the final prediction. These findings indicate that the best-performing model relied on complementary information from clinical and radiographic data rather than on a single model or modality. Contribution of individual models in the weighted ensemble. Bars represent mean weights with standard deviations across cross-validation Contribution of each modality in the weighted ensemble. Bars indicate mean weights with standard deviations across repeated cross-validation Discussion In this study, we developed a multimodal ensemble model integrating frontal radiographs, lateral radiographs, and clinical variables to predict curve progression in patients with AIS. The proposed model achieved an AUC of 0.819, outperforming models using a single modality. These findings suggest that combining radiographic and clinical information improves predictive performance for scoliosis progression. Several previous studies have applied machine learning or deep learning approaches to predict curve progression in AIS.13-15,26 Earlier studies primarily relied on radiographic images or clinical variables alone, reporting moderate predictive performance. More recently, multimodal approaches integrating imaging and clinical data have been introduced and have shown improved accuracy. Our findings are consistent with these studies and further extend them by demonstrating that a multimodal ensemble framework integrating multiple models across different modalities can achieve superior predictive performance. To our knowledge, this is the first study to demonstrate the effectiveness of a multimodal ensemble approach combining multiple models and modalities for predicting AIS progression. In our dataset, models using only clinical variables achieved an AUC of 0.719, while models using frontal radiographs alone achieved AUC values of 0.76–0.79. In contrast, integrating imaging and clinical information using an ensemble framework improved predictive performance to an AUC of 0.819, highlighting the complementary value of multimodal data. The superiority of the multimodal approach may be explained by the complementary nature of the included data types. Clinical variables, such as skeletal maturity and baseline curve magnitude, reflect the biological potential for progression, whereas radiographic images provide detailed structural information regarding spinal alignment and deformity patterns.3,5 Lateral radiographs, in particular, may capture sagittal alignment characteristics that are not fully represented in frontal images.27,28 By integrating these heterogeneous sources, the model can leverage both global structural patterns and clinical context, resulting in improved predictive accuracy. This concept is consistent with emerging evidence in medical artificial intelligence, where multimodal models have shown enhanced performance compared with unimodal approaches across various clinical applications.8,29,30 In addition, the improvement in predictive performance observed in the multimodal model may reflect the inherent heterogeneity of scoliosis progression mechanisms.3,31 Curve progression is not determined by a single factor but rather by the interaction between structural deformity, growth dynamics, and biomechanical loading conditions.5,17 Frontal radiographs primarily capture coronal plane deformity, whereas lateral radiographs provide information on sagittal alignment, which has been increasingly recognized as an important factor in spinal deformity. 27 Clinical variables further reflect the biological growth potential of the patient. 4 Therefore, integrating these modalities enables the model to capture both spatial and biological dimensions of disease progression, which may explain the observed improvement in predictive performance. Ensemble weight analysis revealed that clinical variables contributed the largest weight, followed by frontal and lateral radiographs. This finding suggests that established clinical predictors, such as Risser sign and baseline Cobb angle, remain fundamental determinants of progression risk, while radiographic images provide additional structural information that cannot be fully captured by conventional measurements alone.3-5 The integration of these modalities therefore enables a more comprehensive assessment of scoliosis progression risk. Another important finding of this study is the effectiveness of ensemble learning.8,25,32 Among the four ensemble approaches tested, weighted averaging showed the highest performance, outperforming logistic regression and SVM-based ensembles. Ensemble learning allows predictions from multiple models to be combined, reducing variance and improving robustness.8,32,33 This advantage may also be attributed to the ability of ensemble learning to capture diverse decision boundaries from different models, thereby improving generalizability and reducing overfitting. In the present study, predictions from nine individual models were integrated, including deep learning models for radiographs and machine learning models for clinical data. This approach likely contributed to the improved predictive performance observed in the multimodal model. Accurate prediction of curve progression is critical in the management of AIS.4,6,18 Importantly, this model enables risk prediction at the initial visit, which remains a major unmet clinical need in AIS management. Early identification of patients at high risk of progression may help guide treatment decisions, including closer monitoring, brace treatment, or early surgical consideration.4,18,31 The proposed multimodal model may have potential value as a clinical decision-support tool for risk stratification in AIS. However, its predictions should be interpreted in conjunction with established clinical and radiographic findings rather than used as a standalone basis for treatment decisions. External validation and prospective evaluation are required before clinical implementation. If externally validated, the model may support more individualized follow-up strategies by identifying patients who may require closer monitoring while potentially reducing unnecessary imaging in patients at lower predicted risk. Several limitations of this study should be acknowledged. First, this was a retrospective study, and although data were collected from multiple institutions, external validation using independent datasets is necessary to confirm generalizability. Second, the dataset size, while relatively large compared with previous studies, remains limited for deep learning applications. Larger multicenter datasets may further improve model performance. Third, the present model used only baseline radiographs and limited clinical variables. Potentially relevant factors, including curve type, curve flexibility, additional skeletal maturity indicators, and brace-related variables, were not incorporated into the model. Their inclusion may improve prediction accuracy, interpretability, and clinical utility in future studies. Fourth, brace treatment may have influenced progression outcomes in this cohort. Additional analyses demonstrated significant associations of both brace treatment status (P = 0.035) and brace-wearing duration (P = 0.008) with progression status. However, brace treatment was not randomly assigned and was prescribed according to routine clinical indications. Patients in the progression group were younger, had larger baseline Cobb angles, and exhibited lower skeletal maturity, indicating a higher baseline risk of progression. Therefore, the observed associations likely reflect treatment indication and baseline progression risk rather than a direct causal effect of brace treatment. Brace treatment mainly consisted of underarm thoracolumbosacral orthoses, such as Boston-type braces and active correction braces. Because brace treatment status and brace-wearing duration were not incorporated into the predictive model, their potential influence on model performance cannot be excluded. Future studies incorporating standardized information on brace treatment status and objectively measured brace compliance may further improve prediction accuracy and clinical applicability. Fifth, curve progression was defined as an increase of ≥10°, and patients with an intermediate increase of 6–9° were excluded from model development. This threshold was intentionally selected to establish clearly distinguishable outcome groups and reduce label ambiguity related to Cobb angle measurement variability, postural variation, and temporal variation in flexible AIS curves. However, this definition differs from the criteria commonly used in brace studies, in which an increase of ≤5° is generally considered treatment success and an increase of ≥6° is considered progression or treatment failure. 16 Therefore, direct comparison with previous brace studies may be limited. In addition, exclusion of patients with intermediate progression may reduce the generalizability of the model to borderline cases encountered in clinical practice. Future studies should evaluate model performance across the full spectrum of curve progression. Sixth, horizontal flipping was used for image augmentation. Although this procedure was applied only to the training folds to increase image diversity and reduce overfitting, it may have altered potentially relevant laterality information because right- and left-sided curves may differ in their clinical characteristics or progression patterns. Curve direction was not explicitly incorporated into the present model. Future studies should evaluate model performance according to curve laterality and determine whether augmentation strategies that preserve laterality improve clinical validity. Finally, although ensemble weights provided information regarding the relative contributions of individual models and modalities, the contribution of each clinical variable and the specific imaging features driving predictions could not be fully interpreted. Post hoc explainability analyses such as SHAP were not performed in the present study. Although the ensemble-weight analysis provided a model- and modality-level assessment of relative contribution, it did not determine the importance of individual clinical variables. Further studies using explainable AI techniques may help clarify the relative contributions of clinical and radiographic features associated with progression risk.34,35 In conclusion, we developed a multimodal ensemble model integrating radiographic images and clinical variables for predicting curve progression in AIS. The proposed model demonstrated improved predictive performance compared with single-modality models, achieving an AUC of 0.819. These findings suggest that multimodal machine learning may support risk stratification in AIS; however, external validation and prospective clinical evaluation are required before routine clinical use. Supplemental Material Supplemental Material - Development and Validation of a Multi-Modal Ensemble Model for Predicting Progression in Idiopathic Scoliosis Supplemental Material for Development and Validation of a Multi-Modal Ensemble Model for Predicting Progression in Idiopathic Scoliosis by Hideyuki Arima, Shota Ichikawa, Yu Yamato, Kei Watanabe, Haruki Ueda, Shinji Takahashi, Naobumi Hosogane, Takumi Takeuchi, Hiroki Oba, Masashi Okamoto, Manami Umezu, Yohan Kondo, Shoji Seki in Global Spine Journal