Skip to main content
Advertisement
Browse Subject Areas
?

Click through the PLOS taxonomy to find articles in your field.

For more information about PLOS Subject Areas, click here.

  • Loading metrics

Predicting curve progression for adolescent idiopathic scoliosis using random forest model

  • Ausilah Alfraihat,

    Roles Formal analysis, Investigation, Methodology, Software, Visualization, Writing – original draft

    Affiliation School of Biomedical Engineering, Science and Health Systems, Drexel University, Philadelphia, Pennsylvania, United States of America

  • Amer F. Samdani,

    Roles Formal analysis, Resources, Writing – review & editing

    Affiliation Shriners Hospitals for Children, Philadelphia, Pennsylvania, United States of America

  • Sriram Balasubramanian

    Roles Conceptualization, Data curation, Formal analysis, Funding acquisition, Methodology, Project administration, Resources, Supervision, Visualization, Writing – original draft

    sb939@drexel.edu

    Affiliation School of Biomedical Engineering, Science and Health Systems, Drexel University, Philadelphia, Pennsylvania, United States of America

Abstract

Background

Adolescent Idiopathic Scoliosis (AIS) is a three-dimensional (3D) spinal deformity characterized by coronal curvature and rotational deformity. Predicting curve progression is important for the selection and timing of treatment. Although there is a consensus in the literature regarding prognostic factors associated with curve progression, the order of importance, as well as the combination of factors that are most predictive of curve progression is unknown.

Objectives

(1) create an ordered list of prognostic factors that most contribute to curve progression, and (2) develop and validate a Machine Learning (ML) model to predict the final major Cobb angle in AIS patients.

Methods

193 AIS patients were selected for the current study. Preoperative PA, lateral and lateral bending radiographs were retrospectively obtained from the Shriners Hospitals for Children. Demographic and radiographic features, previously reported to be associated with curve progression, were collected. Sequential Backward Floating Selection (SBFS) was used to select a subset of the most predictive features. Based on the performance of several machine learning methods, a Random Forest (RF) regressor model was used to provide the importance rank of prognostic features and to predict the final major Cobb angle.

Results

The seven most predictive prognostic features in the order of importance were initial major Cobb angle, flexibility, initial lumbar lordosis angle, initial thoracic kyphosis angle, age at last visit, number of levels involved, and Risser "+" stage at the first visit. The RF model predicted the final major Cobb angle with a Mean Absolute Error (MAE) of 4.64 degrees.

Conclusion

A RF model was developed and validated to identify the most important prognostic features for curve progression and predict the final major Cobb angle. It is possible to predict the final major Cobb angle value within 5 degrees error from 2D radiographic features. Such methods could be directly applied to guide intervention timing and optimization for AIS treatment.

1. Introduction

Adolescent Idiopathic Scoliosis (AIS) is a three-dimensional (3D) spinal deformity characterized by adolescent-onset, coronal curvature, and rotational deformity. AIS is the most common form of spinal deformity, accounting for 80% of pediatric scoliosis and affects 2–4% of children throughout their growth spurt [1, 2]. The gold standard to quantify the severity of AIS is by Cobb angle measurement from radiographs, which is the angle between the two most tilted vertebrae in the spinal curvature [3]. AIS has a heterogeneous natural history, with some patients presenting with rapidly progressive curves and others progressing slowly [46]. Mild to moderate progressive scoliosis may result in cosmetic deformity, back pain, functional impairments, and psychological issues [7, 8], whereas severe cases are associated with cardiac dysfunction and pulmonary constraints [912]. Treatment of AIS includes braces and surgical interventions. Bracing is recommended for AIS patients with a curvature of 25–45 degrees [2]. Surgery is considered in skeletally immature patients with a structural thoracic curve greater than 45 degrees or patients with continued progression [13].

The prediction of curve progression is important for treatment selection and timing. A major concern of orthopedic surgeons in managing AIS patients with a minor curvature is identifying which curves will progress to moderate or severe deformities that will require surgical intervention [1416]. As the progression of scoliosis deformity is primarily dependent on growth during skeletal immaturity [1719], predicting the risk of curve progression relies on surrogate indicators. Previous studies have associated various clinical and radiographic factors such as gender, age, curve pattern, curve magnitude, apical level location, flexibility, biomarkers, and skeletal maturity with curve progression [17, 2032]. The patient’s younger chronological and skeletal ages at curve onset, greater initial spine curvature, and delayed onset of menarche are known factors associated with increased curve progression [23, 2628, 33]. Thoracic curves are more likely to progress than lumbar curves because of the inherent stiffness associated with the presence of the rib cage [30, 31], and double major curves are more likely to progress than single curves [23, 34]. Wedging of the intervertebral disc and adjacent vertebrae are also factors associated with severe spine curvature [5, 24, 29, 3537]. However, attempts to use each of these features individually as a prognostic factor for curve progression showed inaccurate (i.e., a high number of false positive and false negative) predictions [22]. Although there is a consensus in the literature regarding prognostic factors associated with curve progression, the order of importance, as well as the combination of factors that are most predictive of curve progression is unknown.

Statistical modeling has been used to predict the progression of curves at risk for surgical intervention. Sitoula et al. found that Sanders skeletal maturity and initial major Cobb angle were correlated to curve progression [28]. They proposed a logistic regression model which predicted the probability of curve progression beyond 50 degrees in AIS patients, a threshold typically suggested for surgery. Zhang et al. formulated a multivariate logistic regression model comprising clinical parameters and biomarkers to evaluate the probability of curve progression to a Cobb angle greater than 40 degrees [38]. However, logistic regression models assume a linear relationship between the features and the response, while spine growth and curve progression are non-linear processes [39]. Recent studies have also attempted to predict curve progression using Machine Learning (ML) algorithms [4043]. ML refers to the capability of data-driven algorithms to gain knowledge regarding a system directly from data without predetermining mechanistic relationships that rule the system [44]. Moreover, ML algorithms are self-learning and therefore, able to adaptively improve their predictive performance with each new data sample to detect hidden patterns in high dimensional and heterogeneous data [44]. Deng et al. used a random forest model to predict the final major Cobb angle in AIS patients and suggested a method to handle missing values; however, this model was based on single curves, and therefore, it is not applicable to double and triple curves [45]. Wu et al. merged fuzzy c-means clustering methods and artificial neural network (ANN) to predict the future scoliosis Cobb angle [40]. Scoliosis patients (n = 61) were included with at least four follow-up Cobb angle measurements. However, their model may not be applicable to new patients or those with fewer records. Chalmers et al. proposed a conditional fuzzy clustering model to predict curve progression in braced AIS patients [41, 42]. Although the model achieved good performance in clustering, the importance of features and their relationships may not be properly identified. Ajemba et al. utilized common indicators of AIS patients to predict the progression of scoliosis with support vector machine [43]. However, their study used a small sample size (n = 44), and the results had low accuracy (65%). Moreover, Edgar et al. used statistical generative model to predict the shape variation of the spinal curve. However, 3D reconstructions of the spinal curve from 2D radiographs are sophisticated [6, 4648].

Considering that AIS is a multifactorial pathology, the predictability of progression remains challenging. Hence, the objectives of the current study are to (1) create an ordered list of prognostic factors (features) that most contribute to curve progression using backward elimination feature selection method and (2) develop and validate an interpretable Machine Learning (ML) model to predict final major Cobb angle in AIS patients. We hypothesize that there is a combination of clinical features that significantly contribute to curve progression, and a ML model that uses these features can provide accurate and robust curve progression prediction, provided sufficient patient demographic and radiographic data.

2. Methods

2.1. Inclusion criteria for patients selection

After institutional review board approval (Protocoal# PHL1904R) was obtained by the western institutional review board, pediatric patients from the Shriners hospitals for children database (1995–2019) were retrospectively selected based on the following inclusion criteria: 1) primary diagnosis of AIS with any Lenke type, 2) males and females ages 10–19 years, and 3) at least two follow-up visits with preoperative (posteroanterior (PA) or anteroposterior (AP), lateral and side bending) radiographic images (X-ray, CT or EOS-imaging). Before accessing the database, all data were fully anonymized, and as the data is retrospective, the western institutional review board waived the requirement for informed consent.

2.2. Data collection

Previously reported demographic and radiographic features associated with curve progression were collected. In the current study, eight continuous and six categorical features were used (Tables 1 and 2).

2.2.1. Demographic features.

Based on the inclusion criteria, 193 patients (24 males, 169 females) were selected for the current study. Gender, age, and bracing status at initial and final follow-up visits for each subject are shown in Tables 1 and 2.

2.2.2. Radiographic features.

For all patients, preoperative standing (PA or AP, and lateral) and side bending spinal radiographs, including the pelvis, were collected. Custom script (MATLAB, The MathWorks Inc, Natick, MA) was used to select four landmark points (LMPs) on the vertices of each vertebra in standing and side bending radiographs (Fig 1a and 1b). Using these LMPs, the following features were extracted at initial and final follow-up visits: Cobb angle, vertebral wedge angle, flexibility, apical level, and number of levels involved in the curve. Thoracic Kyphosis (T2-T12) and lumbar lordosis (L1-L5) angles were measured manually from lateral radiographs. Axial rotation of the apical level was measured in the frontal radiograph based on Raimondi’s method [49, 50].

thumbnail
Fig 1. Steps of Cobb angle measurement from frontal radiographs.

a. Plain frontal X-ray b. Four LMPs were selected per vertebra c. Polygon fit through each vertebra vertices (LMPs) d. Best fit ellipse through each polygon to calculate the orientation of each vertebra as the angle between the major axis of the ellipse and the horizontal e. A cubic spline was fit through the centroids of the vertebrae, the most tilted vertebrae above and below the apical level were identified, and the Cobb angle was calculated between these vertebrae.

https://doi.org/10.1371/journal.pone.0273002.g001

Using a custom script, polygons were fit through the LMPs to represent each vertebra’s boundaries and their centroids (Fig 1c). Additionally, an ellipse was fit through each polygon, and the orientation of the vertebra as measured by the angle of the major axis with respect to the horizontal was calculated (Fig 1d). A cubic spline was fit using the centroids, and the maxima and minima points were identified. The apical level was defined as the closest centroid from the maxima or the minima, depending on the sidedness and location of the curve. The most tilted vertebrae above and below the apical level were identified, and the Cobb angle was calculated between these vertebrae. The apical vertebral wedge angle was calculated as the angle between the superior and inferior endplates of the apical vertebra (Fig 1e).

As side bending radiographs were not available for some patients, modified Lenke (mLenke) classification system was used instead of the original Lenke system [28, 51]. The major curve was defined as that with the greatest Cobb angle, and the minor curve with a Cobb angle ≥ 80% of the major curve magnitude was defined as structural. For the mLenke classification, six curve types were used: type 1 (main thoracic), type 2 (double thoracic), type 3 (double major; thoracic curve > lumbar curve), type 4 (triple major), type 5 (thoracolumbar/lumbar), and type 6 (double major; thoracolumbar or lumbar curve > thoracic curve). The flexibility was assessed using the rate of difference in Cobb angle between the standing (PA or AP) and side bending radiographs [52] (Fig 2).

The Risser stage, a widely used indicator of skeletal maturity [53], describes the ossification patterns and fusion of the iliac apophysis [53, 54]. However, the Risser staging system does not predict the growth spurt as the iliac apophysis starts at Risser stage 1. Therefore, its accuracy has been contested [28, 55]. In the current study, an eight-point Risser "+" staging was used as a more precise and detailed marker of skeletal maturity during growth spurt [55]. The Risser "+" system combines the North American (NA) and European (EU) versions of the traditional Risser score [55]. The Risser "+" system consists of grade: 0-(open triradiate cartilage (TRC)), 0+ (Closed TRC), 1 indicates 0–25% coverage, 2 indicates 25%-50% coverage, 3 indicates 50%-75% coverage, 3/4 indicates 75%-100% coverage, 4 indicates start of fusion and 5 indicates complete fusion.

2.3. Sequential Backward Floating Selection (SBFS)

To improve prediction accuracy and ease of data interpretation, as well as to reduce model complexity and eliminate correlated or redundant features, Sequential Backward Floating Selection (SBFS) method was used. SBFS was implemented using mlxtend (version 0.18.0 For python 3.6) [56] and was used as part of a pipeline including Scikit-learn’s (version 0.21.3 for python 3.6) [57] random forest and cross-validation. SBFS starts with the whole feature set as input (Table 3) and performs regression between the features and the target (final major Cobb angle). Based on the regression performance metric (i.e., Mean Absolute Error (MAE)), the feature that minimizes the performance metric upon removal is excluded, and the model is refitted. After each removal step, there is an additional conditional inclusion step to add features after they were excluded, for which the MAE is minimized. The subset of features with the minimum MAE was selected as input for the predictive model.

2.4. Machine learning model selection and optimization

The dataset of the most predictive features (for all patients) was randomly split into training (75%) and testing (25%) datasets. Using scikit-learn (version 0.21.3 for python 3.6) [57], three commonly used machine learning models namely, Random Forest (RF) [58], Support Vector Machine (SVM) [59, 60], and ANN [61, 62] were trained using 5-fold cross-validation and tested to predict final major Cobb angle.

To optimize the models’ performance and avoid overfitting, ranges of hyperparameters were defined and tuned using Scikit-Learn’s GridSearchCV (version 0.21.3 for python 3.6) [57]. Based on a 5-fold cross-validation, the combination of hyperparameters that produced the best performance metric (lowest MAE) was selected for each model.

2.5. Performance evaluation metric

The performance of the ML models was evaluated with Mean Absolute Error (MAE). MAE is the average of the absolute differences (i.e., residuals) between the final major Cobb angle magnitude predicted by the ML model and the actual final major Cobb angle value (Eq 1). An MAE of 5 degrees or less was considered clinically acceptable based on measurement error of medical imaging data [63, 64].

(1)

2.6. Deployment of the curve progression prediction model

To ensure the clinical translation of the reported findings, and for the model to be more widely used, a web user-interface app to predict the final Cobb angle in AIS patients was created. Model deployment was done using Flask and Heroku (Fig 3), where Flask is a micro web framework to create web applications, and Heroku is a cloud platform to host such web applications.

thumbnail
Fig 3. Flowchart of the curve progression model deployment.

https://doi.org/10.1371/journal.pone.0273002.g003

3. Results

Continuous features (n = 9): includes initial major Cobb angle, initial lumbar lordosis angle, initial thoracic kyphosis angle, age at initial follow-up, age at final follow-up, initial apical wedge angle, the time span between age at initial and final follow-ups, flexibility and axial rotation. Data of the continuous features at the initial visit are shown in Table 1.

Categorical features (n = 6) include three nominal features, namely Lenke type, gender, and brace status, and three ordinal features: apex location, number of levels involved in the curve, and Risser "+" stage. Categorical features were encoded. The description of the categorical features is shown in Table 2.

3.1. Most predictive features selected from the SBFS method

The input features to the SBFS and their corresponding indices are listed in Table 3. Table 4 shows a descending order of the features subsets as input to the RF model based on their corresponding 5-fold cross-validation average MAE. The best seven predictive features out of the fifteen features selected by SBFS are: initial major Cobb angle, flexibility, initial lumbar lordosis angle, initial thoracic kyphosis angle, age at last visit, number of levels involved, and Risser "+" stage at the first visit.

3.2. Random forest regressor to predict final major Cobb angle

The range of hyperparameters and the selected values used to optimize the ML models are shown in Table 5. The seven most predictive features selected by the SBFS algorithm were used as inputs for the ML models. Based on its superlative performance (Table 6), the RF model was selected as the ML method to predict the final major Cobb angle. The rank of the importance of the prognostic features and their respective weights are shown in Fig 4a and Table 7, respectively. The testing errors and their frequencies are shown in Fig 4b.

thumbnail
Fig 4.

a. Rank of feature importance of most predictive features b. Frequency of prediction error of the testing dataset.

https://doi.org/10.1371/journal.pone.0273002.g004

thumbnail
Table 5. Hyperparameters grid values and the selected values for optimized models.

https://doi.org/10.1371/journal.pone.0273002.t005

thumbnail
Table 6. Performance of the ML models in terms of Mean Absolute Error (MAE).

https://doi.org/10.1371/journal.pone.0273002.t006

thumbnail
Table 7. Rank and weights of most important features to predict curve progression.

https://doi.org/10.1371/journal.pone.0273002.t007

3.3. Web user-interface app for AIS curve progression prediction

The web user-interface app is publicly available at: https://biomed.drexel.edu/labs/obl/toolkits/predict-curve-progression/

4. Discussion

This is the first study to apply machine learning methods to longitudinal data from a cohort of AIS patients to identify a rank-ordered list of the most important features associated with curve progression. An RF model was developed and validated using these patient-specific prognostic features (i.e., demographic and radiographic measurements) to predict curve progression in AIS patients. The composite model presented in the current study is well suited to characterize the curve progression of a multifactorial AIS pathomechanism.

Previous studies that reported prognostic features associated with curve progression had several limitations. While Wu et al. predicted follow-up Cobb angle based on a single feature (current Cobb angle), radiographs at multiple longitudinal timepoints were needed [40]. The predictive model for curve progression in braced AIS patients developed by Chalmers et al. [41] did not include two skeletal maturity measurements, namely Risser sign and triradiate cartilage status that were previously reported to be highly prognostic [18, 19]. In contrast, the current study used Risser "+" staging, which combines these skeletal maturity grades. Ajemba et al. [43] reduced the number of features using principal component analysis; however, the principal components are difficult to interpret. It is also important to note that none of the previously reported models provided an ordered list of the importance of predictive features.

The current study is the first to predict the magnitude of the final major Cobb angle (with a mean absolute error of 4.64 degrees) to a clinically acceptable error based on 2D radiographic features at the initial curve presentation. While Deng et al. [45] also used a RF model to predict the final major Cobb angle based on the initial major Cobb angle, bracing status, and age, the dataset had a large number of missing values which would increase the risk of model bias. The combined fuzzy c-means clustering, ANN, and cubic spline extrapolation method reported by Wu et al. predicted Cobb angle with good accuracy (mean error of 4.4±1.86 degrees) [40]. However, the model required three Cobb angle measurements at 6- to 12-month intervals and, therefore, could be used for early prognosis. Furthermore, the prediction accuracy was affected by the 3D reconstruction methods used in their study. More recently, 3D spine shape variations based on independent component analysis differed from the actual curvature by 1.83, 5.18, and 4.79 degrees of Cobb angles in the proximal thoracic, main thoracic, and thoracolumbar, lumbar sections, respectively [46]. While this method seems promising, it may be difficult to integrate into the clinical workflow due to limitations associated with sophisticated 3D spine reconstructions from biplanar radiographs. In contrast, the current method uses 2D radiographic features that can be easily measured during routine clinical care.

The current study reports that thoracic kyphosis and lumbar lordosis are important prognostic factors in predicting scoliotic curve progression. Although it might not be obvious that sagittal plane deformity (i.e., kyphosis and lordosis) can be predictive of deformity progression in the coronal plane (i.e., scoliosis), previous studies have reported asymmetric alterations in loading due to coupled motion patterns in the scoliotic spine [65]. Coupled rotation and lateral bending motions, while forcing the posterior elements towards the curve’s concavity [6568], create biplanar asymmetry [65, 68, 69]. Therefore, the structural alterations in the coronal plane are a coupled effect that accompanies the culmination of a sagittal plane deformity and a transverse plane rotation. Such scoliotic deformity progression is a result of a vicious cycle, according to the Heuter-Volkmann principle [70, 71].

DNA-based prognostic tests to predict curve progression in AIS have gained interest since the first genome-wide association studies (GWAS) [72]. Various single nucleotide polymorphisms (SNP) have been reported in several cohorts comparing AIS and control populations [73, 74]. The majority of the associations reported previously are not replicable in different ethnic groups, and their biological roles in AIS remain unclear [75]. A saliva-based genetic prognostic test called ScolioScore analyzes 53 SNPs to identify patients’ risk of AIS curve progression [76]; however, its predictive value is limited [76, 77]. Follow-up replication studies conducted by Ogura [78] and Tang [79] could not confirm the association between the 53 SNPs and curve progression in Japanese and French-Canadian populations, respectively. Due to the aforementioned limitations and since DNA-based testing results were unavailable for the AIS patients in the current study, gene-based features were not included.

The proposed prediction model outperformed existing models in terms of higher accuracy, interpretability, greater sample size, which prevents overfitting, and ease of use, utilizing features that are routine clinical measurements and relatively easy to extract rather than sophisticated 3D shape reconstruction from 2D radiographs. This study has a few limitations: Although RF is a robust regression algorithm, it is unable to discover trends that would enable it to extrapolate values that fall outside the range of the features in the training set. However, this limitation can be reduced by increasing the size of the training set. Moreover, the outcomes of this study may be affected by institutional bias, and future work comparing the results on expanded datasets from other institutions may be beneficial. In addition, because of the retrospective nature of this study, data on menarche status, bracing onset, brace compliance, and brace duration were not available. Also, Sanders skeletal maturity [27] was not used since not all patients had hand radiographs. Menarche status, bracing, and Sanders skeletal maturity can be evaluated as predictive features for future prospective studies.

5. Conclusions and future work

In this study, a machine learning model (Random Forest method) was developed and validated to predict the final major Cobb angle in AIS patients. The results indicate that it is possible to predict the value of the final major Cobb angle with an accuracy of 5 degrees as compared to 2D radiographic measurements. A web user-interface app for AIS curve progression prediction was also created. This web user-interface app could be used by clinicians to help guide the timing and optimization of AIS treatment methods. Also, this app can be used to educate patients and their families about patient-specific spine deformity characteristics, risk of progression, and possible treatment options.

Acknowledgments

We would like to thank Shriners Hospitals for Children-Philadelphia for providing the clinical data used in this study.

References

  1. 1. Reamy BV, Slakey J. Adolescent idiopathic scoliosis: review and current concepts. Am Fam Physician. 2001;64(1):111. pmid:11456428
  2. 2. Weinstein SL, Dolan LA, Cheng JC, Danielsson A, Morcuende JA. Adolescent idiopathic scoliosis. The lancet. 2008;371(9623):1527–37.
  3. 3. Morrissy R, Goldsmith G, Hall E, Kehl D, Cowie G. Measurement of the Cobb angle on radiographs of patients who have. J Bone Joint Surg Am. 1990;72(3):320–7. pmid:2312527
  4. 4. Ede MMN, Jones SW. Adolescent idiopathic scoliosis: evidence for intrinsic factors driving aetiology and progression. Int Orthop. 2016;40(10):2075–80. pmid:26961194
  5. 5. Peters JR, Chandrasekaran C, Robinson LF, Servaes SE, Campbell RM Jr, Balasubramanian S. Age-and gender-related changes in pediatric thoracic vertebral morphology. The spine journal. 2015;15(5):1000–20. pmid:25681580
  6. 6. Peters JR, Servaes SE, Cahill PJ, Balasubramanian S. Morphology and growth of the pediatric lumbar vertebrae. The Spine Journal. 2021;21(4):682–97. pmid:33152510
  7. 7. Ylikoski M. Growth and progression of adolescent idiopathic scoliosis in girls. Journal of pediatric orthopaedics B. 2005;14(5):320–4. pmid:16093941
  8. 8. Bengtsson G, Fällström K, Jansson B, Nachemson A. A psychological and psychiatric investigation of the adjustment of female scoliosis patients. Acta Psychiatr Scand. 1974;50(1):50–9. pmid:4275025
  9. 9. Liu L, Xiu P, Li Q, Song Y, Chen R, Zhou C. Prevalence of cardiac dysfunction and abnormalities in patients with adolescent idiopathic scoliosis requiring surgery. Orthopedics. 2010;33(12).
  10. 10. Winter RB, Lonstein JE. Ultra-long-term follow-up of pediatric spinal deformity problems: 23 patients with a mean follow-up of 51 years. Springer; 2009.
  11. 11. Pehrsson K, Bake B, Larsson S, Nachemson A. Lung function in adult idiopathic scoliosis: a 20 year follow up. Thorax. 1991;46(7):474–8. pmid:1877034
  12. 12. Harris JA, Mayer OH, Shah SA, Campbell RM, Balasubramanian S. A comprehensive review of thoracic deformity parameters in scoliosis. Eur Spine J. 2014;23(12):2594–602. pmid:25238798
  13. 13. Wagner SC, Lehman RA Jr, Lenke LG, editors. Surgical management of adolescent idiopathic scoliosis. Semin Spine Surg; 2015: Elsevier.
  14. 14. Ascani E, Bartolozzi P, Logroscino C, Marchetti P, Ponte A, Savini R, et al. Natural history of untreated idiopathic scoliosis after skeletal maturity. Spine (Phila Pa 1976). 1986;11(8):784–9. pmid:3810293
  15. 15. Karol LA, Johnston C 2nd, Browne RH, Madison M. Progression of the curve in boys who have idiopathic scoliosis. JBJS. 1993;75(12):1804–10. pmid:8258551
  16. 16. Peterson L-E, Nachemson AL. Prediction of progression of the curve in girls who have adolescent idiopathic scoliosis of moderate severity. Logistic regression analysis based on data from The Brace Study of the Scoliosis Research Society. The Journal of bone and joint surgery American volume. 1995;77(6):823–7. pmid:7782354
  17. 17. Charles YP, Daures J-P, de Rosa V, Diméglio A. Progression risk of idiopathic juvenile scoliosis during pubertal growth. Spine (Phila Pa 1976). 2006;31(17):1933–42. pmid:16924210
  18. 18. Dimeglio A, Canavese F. Progression or not progression? How to deal with adolescent idiopathic scoliosis during puberty. J Child Orthop. 2013;7(1):43–9. pmid:24432058
  19. 19. Dimeglio A, Canavese F, Charles YP. Growth and adolescent idiopathic scoliosis: when and how much? Journal of Pediatric Orthopaedics. 2011;31:S28–S36. pmid:21173616
  20. 20. Bunnell WP. The natural history of idiopathic scoliosis before skeletal maturity. Spine (Phila Pa 1976). 1986;11(8):773–6. pmid:3810290
  21. 21. Soucacos PN, Zacharis K, Gelalis J, Soultanis K, Kalos N, Beris A, et al. Assessment of curve progression in idiopathic scoliosis. Eur Spine J. 1998;7(4):270–7. pmid:9765033
  22. 22. Lonstein J, Carlson J. The prediction of curve progression in untreated idiopathic scoliosis. J Bone Jt Surg. 1984;3(2):1061–71. pmid:6480635
  23. 23. Ohrt-Nissen S, Hallager DW, Gehrchen M, Dahl B. Flexibility predicts curve progression in providence nighttime bracing of patients with adolescent idiopathic scoliosis. Spine (Phila Pa 1976). 2016;41(22):1724–30. pmid:27076435
  24. 24. Cheung JPY, Cheung PWH, Yeng WC, Chan LCK. Does Curve Regression Occur During Underarm Bracing in Patients with Adolescent Idiopathic Scoliosis? Clinical Orthopaedics and Related Research®. 2020;478(2):334–45. pmid:31688210
  25. 25. Rogala EJ, Drummond DS, Gurr J. Scoliosis: incidence and natural history. A prospective epidemiological study. The Journal of bone and joint surgery American volume. 1978;60(2):173–6. pmid:641080
  26. 26. Tan K-J, Moe MM, Vaithinathan R, Wong H-K. Curve progression in idiopathic scoliosis: follow-up study to skeletal maturity. Spine (Phila Pa 1976). 2009;34(7):697–700. pmid:19333102
  27. 27. Sanders JO, Khoury JG, Kishan S, Browne RH, Mooney JF III, Arnold KD, et al. Predicting scoliosis progression from skeletal maturity: a simplified classification during adolescence. JBJS. 2008;90(3):540–53. pmid:18310704
  28. 28. Sitoula P, Verma K, Holmes L, Gabos PG, Sanders JO, Yorgova P, et al. Prediction of curve progression in idiopathic scoliosis. Spine (Phila Pa 1976). 2015;40(13):1006–13. pmid:26356067
  29. 29. Scott M, Piggott H. A short-term follow-up of patients with mild scoliosis. The Journal of bone and joint surgery British volume. 1981;63(4):523–5. pmid:7298676
  30. 30. Thompson RM, Hubbard EW, Jo C-H, Virostek D, Karol LA. Brace success is related to curve type in patients with adolescent idiopathic scoliosis. JBJS. 2017;99(11):923–8.
  31. 31. Yao G, Cheung JP, Shigematsu H, Ohrt-Nissen S, Cheung K, Luk KD, et al. Characterization and predictive value of segmental curve flexibility in adolescent idiopathic scoliosis patients. Spine (Phila Pa 1976). 2017;42(21):1622–8. pmid:27997505
  32. 32. Neal KM, Shirley ED, Kiebzak GM. Maturity indicators and adolescent idiopathic scoliosis. Spine (Phila Pa 1976). 2018;43(7):E406–E12.
  33. 33. Sanders JO, Browne RH, McConnell SJ, Margraf SA, Cooney TE, Finegold DN. Maturity assessment and curve progression in girls with idiopathic scoliosis. JBJS. 2007;89(1):64–73. pmid:17200312
  34. 34. Cheung JP, Cheung PW. Supine flexibility predicts curve progression for patients with adolescent idiopathic scoliosis undergoing underarm bracing. The Bone & Joint Journal. 2020;102(2):254–60. pmid:32009436
  35. 35. Alfraihat A, Olson JC, Snyder BD, Cahill PJ, Balasubramanian S. Thoracic vertebral morphology in normal and scoliosis deformity in skeletally immature rabbits: A Longitudinal study. Jor Spine. 2020;3(4):e1118. pmid:33392455
  36. 36. Balasubramanian S, Peters JR, Robinson LF, Singh A, Kent RW. Thoracic spine morphology of a pseudo-biped animal model (kangaroo) and comparisons with human and quadruped animals. Eur Spine J. 2016;25(12):4140–54. pmid:27704284
  37. 37. Peters JR, Campbell RM Jr, Balasubramanian S. Characterization of the age-dependent shape of the pediatric thoracic spine and vertebrae using generalized procrustes analysis. J Biomech. 2017;63:32–40. pmid:28874278
  38. 38. Zhang J, Cheuk K-y, Xu L, Wang Y, Feng Z, Sit T, et al. A validated composite model to predict risk of curve progression in adolescent idiopathic scoliosis. EClinicalMedicine. 2020;18:100236. pmid:31922123
  39. 39. Dimeglio A, Canavese F. The immature spine: growth and idiopathic scoliosis. Annals of Translational Medicine. 2020;8(2). pmid:32055613
  40. 40. Wu H, Ronsky J, Poncet P, Cheriet F, Xue D, Harder J, et al., editors. Prediction of scoliosis progression in time series using a hybrid learning technique. 2005 IEEE Engineering in Medicine and Biology 27th Annual Conference; 2006: IEEE.
  41. 41. Chalmers E, Westover L, Jacob J, Donauer A, Zhao VH, Parent EC, et al. Predicting success or failure of brace treatment for adolescents with idiopathic scoliosis. Med Biol Eng Comput. 2015;53(10):1001–9. pmid:26002592
  42. 42. Chalmers E, Pedrycz W, Lou E, editors. Predicting the outcome of brace treatment for scoliosis using conditional fuzzy clustering. 2013 Joint IFSA World Congress and NAFIPS Annual Meeting (IFSA/NAFIPS); 2013: IEEE.
  43. 43. Ajemba PO, Ramirez L, Durdle NG, Hill DL, Raso VJ. A support vectors classifier approach to predicting the risk of progression of adolescent idiopathic scoliosis. IEEE Trans Inf Technol Biomed. 2005;9(2):276–82. pmid:16138544
  44. 44. Mitchell TM. The discipline of machine learning: Carnegie Mellon University, School of Computer Science, Machine Learning …; 2006.
  45. 45. Deng L, Hu Y, Cheung JPY, Luk KDK. A data-driven decision support system for scoliosis prognosis. IEEE Access. 2017;5:7874–84.
  46. 46. García-Cano E, Cosío FA, Duong L, Bellefleur C, Roy-Beaudry M, Joncas J, et al. Prediction of spinal curve progression in Adolescent Idiopathic Scoliosis using Random Forest regression. Comput Biol Med. 2018;103:34–43. pmid:30336363
  47. 47. Balasubramanian S, D’Andrea CR, Viraraghavan G, Cahill PJ. Development of a Finite Element Model of the Pediatric Thoracic and Lumbar Spine, Ribcage, and Pelvis With Orthotropic Region-Specific Vertebral Growth. J Biomech Eng. 2022;144(10):101007. pmid:35466381
  48. 48. Hadagali P, Peters JR, Balasubramanian S. Morphing the feature-based multi-blocks of normative/healthy vertebral geometries to scoliosis vertebral geometries: development of personalized finite element models. Comput Methods Biomech Biomed Engin. 2018;21(4):297–324. pmid:29528253
  49. 49. Weiss H-R. Measurement of vertebral rotation: Perdriolle versus Raimondi. Eur Spine J. 1995;4(1):34–8. pmid:7749905
  50. 50. Hurtado-Aviles J, Roca-González J, Kurochka KS, Sanz-Mengibar JM, Santonja-Medina F. Developing of a Mathematical Model to Perform Measurements of Axial Vertebral Rotation on Computer-Aided and Automated Diagnosis Systems, Using Raimondi’s Method. Radiol Res Pract. 2021;2021. pmid:33628503
  51. 51. Grothaus O, Molina D, Jacobs C, Talwalkar V, Iwinski H, Muchow R. Is It Growth or Natural History? Increasing Spinal Deformity After Sanders Stage 7 in Females With AIS. Journal of Pediatric Orthopaedics. 2020;40(3):e176–e81. pmid:31181026
  52. 52. Hirsch C, Ilharreborde B, Mazda K. Flexibility analysis in adolescent idiopathic scoliosis on side-bending images using the EOS imaging system. Orthopaedics & Traumatology: Surgery & Research. 2016;102(4):495–500. pmid:27090816
  53. 53. Hacquebord JH, Leopold SS. In brief: the Risser classification: a classic tool for the clinician treating adolescent idiopathic scoliosis. Springer; 2012.
  54. 54. Risser JC. The classic: the iliac apophysis: an invaluable sign in the management of scoliosis. Clinical Orthopaedics and Related Research®. 2010;468(3):646–53. pmid:19763720
  55. 55. Troy MJ, Miller PE, Price N, Talwalkar V, Zaina F, Donzelli S, et al. The “Risser+” grade: a new grading system to classify skeletal maturity in idiopathic scoliosis. Eur Spine J. 2019;28(3):559–66. pmid:30446865
  56. 56. Raschka S. MLxtend: providing machine learning and data science utilities and extensions to Python’s scientific computing stack. Journal of open source software. 2018;3(24):638.
  57. 57. Pedregosa F, Varoquaux G, Gramfort A, Michel V, Thirion B, Grisel O, et al. Scikit-learn: Machine learning in Python. the Journal of machine Learning research. 2011;12:2825–30.
  58. 58. Breiman L. Random forests. Machine learning. 2001;45(1):5–32.
  59. 59. Noble WS. What is a support vector machine? Nat Biotechnol. 2006;24(12):1565–7. pmid:17160063
  60. 60. Pisner DA, Schnyer DM. Support vector machine. Machine learning: Elsevier; 2020. p. 101–21.
  61. 61. Yegnanarayana B. Artificial neural networks: PHI Learning Pvt. Ltd.; 2009.
  62. 62. Hopfield JJ. Artificial neural networks. IEEE Circuits and Devices Magazine. 1988;4(5):3–10.
  63. 63. Wills BP, Auerbach JD, Zhu X, Caird MS, Horn BD, Flynn JM, et al. Comparison of Cobb angle measurement of scoliosis radiographs with preselected end vertebrae: traditional versus digital acquisition. Spine (Phila Pa 1976). 2007;32(1):98–105. pmid:17202899
  64. 64. Greiner AK. Adolescent idiopathic scoliosis: radiologic decision-making. Am Fam Physician. 2002;65(9):1817. pmid:12018804
  65. 65. Panjabi MM, Brand R Jr, White A 3rd. Mechanical properties of the human thoracic spine as shown by three-dimensional load-displacement curves. The Journal of bone and joint surgery American volume. 1976;58(5):642–52. pmid:932062
  66. 66. White AA III. Kinematics of the normal spine as related to scoliosis. J Biomech. 1971;4(5):405–11. pmid:5133357
  67. 67. Lee D. Biomechanics of the Thorax: A Clinical Mode of in Vivo Function. Journal of Manual & Manipulative Therapy. 1993;1(1):13–21.
  68. 68. Sizer PS Jr, Brismée J-M, Cook C. Coupling behavior of the thoracic spine: a systematic review of the literature. J Manipulative Physiol Ther. 2007;30(5):390–9. pmid:17574958
  69. 69. Dickson RA, Lawton J, Archer I, Butt W. The pathogenesis of idiopathic scoliosis. Biplanar spinal asymmetry. The Journal of bone and joint surgery British volume. 1984;66(1):8–15. pmid:6693483
  70. 70. Volkmann R. Die Krankheiten der Bewegungsorgane. Handbuch der allgemeinen und speziellen Chirurgie, Bd II, Abt. 1869;1:350–1.
  71. 71. Stokes IAF, Burwell RG, Dangerfield PH, Ibse. Biomechanical spinal growth modulation and progressive adolescent scoliosis—a test of the ’vicious cycle’ pathogenetic hypothesis: summary of an electronic focus group debate of the IBSE. Scoliosis. 2006;1:16-. pmid:17049077.
  72. 72. Sharma S, Gao X, Londono D, Devroy SE, Mauldin KN, Frankel JT, et al. Genome-wide association studies of adolescent idiopathic scoliosis suggest candidate susceptibility genes. Hum Mol Genet. 2011;20(7):1456–66. pmid:21216876
  73. 73. Wise CA, Gao X, Shoemaker S, Gordon D, Herring JA. Understanding genetic factors in idiopathic scoliosis, a complex disease of childhood. Current genomics. 2008;9(1):51–9. pmid:19424484
  74. 74. Kou I, Otomo N, Takeda K, Momozawa Y, Lu H-F, Kubo M, et al. Genome-wide association study identifies 14 previously unreported susceptibility loci for adolescent idiopathic scoliosis in Japanese. Nature communications. 2019;10(1):1–9.
  75. 75. Cheng JC, Castelein RM, Chu WC, Danielsson AJ, Dobbs MB, Grivas TB, et al. Adolescent idiopathic scoliosis. Nature reviews disease primers. 2015;1(1):1–21.
  76. 76. Roye BD, Wright ML, Williams BA, Matsumoto H, Corona J, Hyman JE, et al. Does ScoliScore provide more information than traditional clinical estimates of curve progression? Spine (Phila Pa 1976). 2012;37(25):2099–103. pmid:22614798
  77. 77. Roye BD, Wright ML, Matsumoto H, Yorgova P, McCalla D, Hyman JE, et al. An independent evaluation of the validity of a DNA-based prognostic test for adolescent idiopathic scoliosis. JBJS. 2015;97(24):1994–8.
  78. 78. Ogura Y, Takahashi Y, Kou I, Nakajima M, Kono K, Kawakami N, et al. A replication study for association of 53 single nucleotide polymorphisms in a scoliosis prognostic test with progression of adolescent idiopathic scoliosis in Japanese. Spine (Phila Pa 1976). 2013;38(16):1375–9.
  79. 79. Tang QL, Julien C, Eveleigh R, Bourque G, Franco A, Labelle H, et al. A replication study for association of 53 single nucleotide polymorphisms in ScoliScore test with adolescent idiopathic scoliosis in French-Canadian population. Spine (Phila Pa 1976). 2015;40(8):537–43.