Spine Publish Ahead of Print

DOI:10.1097/BRS.0000000000004986

D
TE
Machine learning algorithms for predicting Cobb angle beyond 25 degrees in

female adolescent idiopathic scoliosis patients.

EP
Authors:

Shuhei Ohyama MD1, Satoshi Maki PhD1, Toshiaki Kotani PhD2, Yosuke Ogata MD2,
C

Tsuyoshi Sakuma PhD2, Yasushi Iijima PhD2, Tsutomu Akazawa PhD3, Kazuhide

Inage PhD1, Yasuhiro Shiga PhD1, Masahiro Inoue PhD1, Takahito Arai MD1, Noriyasu
AC

Toshi MD1, Soichiro Tokeshi MD1, Kohei Okuyama MD1, Susumu Tashiro MD1,

Noritaka Suzuki MD1, Yawara Eguchi PhD1, Sumihisa Orita PhD1,4, Shohei Minami

PhD2, Seiji Ohtori PhD1

© 2024 Wolters Kluwer Health, Inc. All rights reserved. Unauthorized reproduction of the article is prohibited.
 Affiliations:

1
Department of Orthopedic Surgery, Graduate School of Medicine, Chiba University,

Chiba, Japan

2
Department of Orthopedic Surgery, Seirei Sakura Citizen Hospital, Sakura, Japan

D
3
Department of Orthopedic Surgery, St. Marianna University School of Medicine,

Kawasaki, Japan

4
TE
Center for Frontier Medical Engineering, Chiba University, Chiba, Japan
EP
Corresponding Author’s name and current institution:

Shuhei Ohyama, MD
C

Department of Orthopedic Surgery, Graduate School of Medicine, Chiba University.

1-8-1 Inohana, Chuo-ku, Chiba-city, Chiba, 260-8670, Japan

AC

Tel: +81-43-226-2117, Fax: +81-43-226-2116

[email protected]

Conflict of Interest and Source of Funding

Authors have no conflicts of interest and no financial support related to this study.

© 2024 Wolters Kluwer Health, Inc. All rights reserved. Unauthorized reproduction of the article is prohibited.
 Patient consent

Written informed consent was obtained from the patient for publication of this study

and accompanying images.

D
Ethics approval

TE
The study adhered to the guidelines of the Declaration of Helsinki and the study

protocol was approved by the institutional review board of our hospital (Approval

number: 2023001).
EP
Data Access Statement
C

The data that support the findings of this study are available from the corresponding

author, upon reasonable request.

AC

Abstract:

Study Design: Retrospective cohort study.

Objective: To develop a machine learning (ML) model that predicts the progression of

AIS using minimal radiographs and simple questionnaires during the first visit.

Summary of Background Data: Several factors are associated with angle progression

© 2024 Wolters Kluwer Health, Inc. All rights reserved. Unauthorized reproduction of the article is prohibited.
 in patients with AIS. However, it is challenging to predict angular progression at the

first visit.

Methods: Among female patients with AIS treated at a single institution from July

2011 to February 2023, 1119 cases were studied. Patient data, including demographic

D
and radiographic data based on anterior-posterior and lateral whole-spine radiographs,

TE
were collected at the first and last visits. The last visit was defined differently based on

treatment plans. For patients slated for surgery or bracing, the last visit occurred just

before these interventions. For others, it was their final visit before turning 18 years.
EP
Angular progression was defined as a Cobb angle greater than 25 degrees for each of

the proximal thoracic (PT), main thoracic (MT), and thoracolumbar/lumbar (TLL)
C

curves at the last visit. ML algorithms were employed to develop individual binary

classification models for each type of curve (PT, MT, and TLL) using PyCaret in
AC

Python. Multiple models were explored and analyzed, with the selection of optimal

models based on the area under the curve (AUC) and Recall scores. Feature

importance was evaluated to understand the contribution of each feature to the model

predictions.

Results: For PT, MT, and TLL progression, the top-performing models exhibit AUC

values of 0.94, 0.89, and 0.84, and achieve recall rates of 0.90, 0.85, and 0.81. The

© 2024 Wolters Kluwer Health, Inc. All rights reserved. Unauthorized reproduction of the article is prohibited.
 most significant factors predicting progression varied for each curve: initial Cobb

angle for PT, presence of menarche for MT, and Risser grade for TLL.

Conclusions: This study introduces an ML-based model using simple data at the first

visit to precisely predict angle progression in female patients with AIS.

D
Key Points


TE
The study aimed to develop a machine learning model that predicts adolescent

idiopathic scoliosis (AIS) progression using minimal radiographs and simple

EP
questionnaires during the first visit.

 Predictive models using data from the first visit predicted the progression of the
C

Cobb angle for each of the proximal thoracic, main thoracic, and thoracolumbar

curves with high accuracy.

AC

 The most significant factors predicting progression varied for each curve. The

most important factors for predicting the progression of each curve were the initial

Cobb angle of the PT for the progression of PT, the presence of menarche for the

progression of MT, and the Risser grade for the progression of TLL.

© 2024 Wolters Kluwer Health, Inc. All rights reserved. Unauthorized reproduction of the article is prohibited.
 Introduction

Adolescent idiopathic scoliosis (AIS) is an adolescent-onset disease characterized by

progressive 3-dimensional spinal deformity.1,2 The coronal Cobb angle determines the

D
treatment plan, including surgical or brace treatments.3-5 For example, to reduce

immature stage with a Cobb angle > 25°. 2

TE
progression, brace treatment should be considered for patients with AIS during the

Identifying the risk of AIS progression is important in determining treatment strategies.

EP
Several factors associated with the progression of AIS have been reported, such as the

initial Cobb angle and skeletal growth indicators, including menarche and Risser
C

grade.6-8 However, predicting the progression of AIS at the time of initial diagnosis is

difficult. 9,10 Therefore, patients with AIS require longitudinal visits and radiological
AC

examinations at empirically determined intervals without knowing their own risk of

progression. These treatment strategies are not adapted to each patient's risk of

progression and may impose multiple burdens on patients with AIS, including radiation

exposure, numerous hospital visits, and psychological anxiety.11-14 Several predictive

methods using statistical approaches have been developed using the approaches.15-17

Previous studies using detailed radiographic features based on computed tomography

© 2024 Wolters Kluwer Health, Inc. All rights reserved. Unauthorized reproduction of the article is prohibited.
 (CT) imaging and whole-spine radiographs from multiple visits have provided highly

accurate predictions.15,16 However, these methods are more invasive because of

radiation exposure or difficulty adapting them to the routine practice of patients with

AIS. Therefore, a highly accurate prediction model adapted for routine practice is

D
required to predict the progression of AIS.

TE
Machine learning (ML) is one approach that improves prediction accuracy.18,19 ML, a

subset of artificial intelligence, uncovers patterns and makes predictions by processing

large volumes of training data. In certain cases, ML models have demonstrated

EP
superior performance over traditional statistical approaches, mainly because of their

ability to identify nonlinear relationships and intricate interactions between variables.18

C

Previous studies explored using ML-based models to predict AIS.20-22 However,

studies that used ML to predict AIS progression based on initial radiographs were less
AC

accurate.22 In addition, no study has predicted the progression of AIS at the first visit

using minimal radiographs and simple questionnaires.

This study aimed to create a classification model for predicting AIS progression using

ML based on the data obtained during the first visit.

Materials and Methods

© 2024 Wolters Kluwer Health, Inc. All rights reserved. Unauthorized reproduction of the article is prohibited.
 We reviewed patients with AIS who first visited a single institution specializing in

scoliosis between July 2011 and February 2023. Considering the differences in the

pathophysiology between males with AIS and females with AIS, this study included

female patients with AIS. AIS was defined as follows:1) a Cobb angle of 10° or more

D
in patients aged 10–18 years and 2) patients with only AIS, such as patients without

TE
syndromic or congenital scoliosis.23 The exclusion criteria were as follows:1) patients

with only a single visit, 2) patients who started bracing or surgical intervention at the

first or second visit, 3) patients who had previous bracing or surgical intervention for
EP
scoliosis, and 4) patients with missing values in any patient data. To enroll patients

with no advanced scoliosis from the first visit, patients with Cobb angle of proximal
C

thoracic (PT), main thoracic (MT), or thoracolumbar/lumbar (TLL) equal to or greater

than 25° at the first visit were excluded.

AC

Patient data were collected during the first and last visits. The last visit was defined

differently based on treatment plans. For patients slated for surgery or bracing, the last

visit occurred just before these interventions. For others, it was their final visit before

turning 18 years. Therefore, patient data in this study were not affected by brace or

surgical treatment. The study adheres to the guidelines of the Declaration of Helsinki,

and the study protocol was approved by the institutional review board. Written

© 2024 Wolters Kluwer Health, Inc. All rights reserved. Unauthorized reproduction of the article is prohibited.
 informed consent was obtained from all participants.

Patient data

D
The patient demographic data at the first visit included age, height, weight, presence of

TE
menarche, and duration (months). Patient radiographic data at the first visit included

the Cobb angle of the PT, MT, and TLL; thoracic kyphosis (TK: T5-12); and Risser

grade using anteroposterior and lateral whole-spine radiographs at the first visit. At our
EP
institution, spinal flexibility assessment is performed only on preoperative patients.

Therefore, spinal flexibility is not included in the radiographic data in this study.
C

During the final visit, the Cobb angles of the PT, MT, and TLL were assessed using

anteroposterior whole-spine radiographs. The regional curve was defined as PT with

AC

the location of apex T3-T5, MT with the location of apex T6-T11/12 disc, and TLL

with the location of apex T11/12 disc-L4.24 A modified Lenke classification system

was used to classify the curves into six types.25 In clinical practice, brace intervention

is used when the main curve is greater than 25 degrees before skeletal maturity at our

institution and surgical treatment is indicated when the main curve is 40 degrees or

greater. Therefore, the outcome was defined as whether the Cobb angle had progressed

© 2024 Wolters Kluwer Health, Inc. All rights reserved. Unauthorized reproduction of the article is prohibited.
 to 25° or greater at the last visit, and progression was evaluated for PT, MT, and TLL,

respectively.

Data pre-processing

D
Menarche status was treated as categorical, whereas the others were treated as

TE
numerical. The duration since menarche was set to -1 for patients without menarche.

Curve progression prediction model

EP
We aimed to predict the progression of PT, MT, and TLL by using separate binary

classification models.
C

In this study, we utilized the PyCaret library in Python to compare the

performances of various classification models. The dataset was randomly divided into
AC

a training set (70% of the data) and a validation set (30% of the data) using stratified

random sampling. We adopted several models, including Extreme Gradient Boosting

(XGBoost), Random Forest, Light Gradient Boosting Machine (LightGBM), Category

Boosting (CatBoost), K-nearest neighbor (KNN), Gaussian Naive Bayes, Linear

Discriminant Analysis, and Logistic Regression. These models were selected because

of their high performance and interpretability.26-30 Because few cases progressed to PT,

© 2024 Wolters Kluwer Health, Inc. All rights reserved. Unauthorized reproduction of the article is prohibited.
 MT, or TLL, we employed over-sampling techniques, including the synthetic minority

oversampling technique (SMOTE), considering the imbalanced nature of our data.31

The performance of each model was evaluated using a stratified K-fold

cross-validation (K=10). The patients were randomly divided into 10 equally populated

D
groups. In each iteration, data for the nine subgroups were designated as the training

TE
dataset, whereas the remaining independent subgroup served as validation dataset. The

hyperparameters of each model were tuned to maximize the area under the receiver

operating characteristic curve (AUC). The best model was selected based on the AUC
EP
and Recall scores. AUC is the most common-used metrics for comparing the

performance of the classification models. Recall value is a metric that measures how
C

often a machine learning model correctly identifies true positives from all the actual

positive samples in the dataset., and recall values were emphasized to avoid missing
AC

curve progression cases in this study. Finally, the performance of each model was

assessed using various classification metrics, including Accuracy, AUC, Recall,

Precision, F1 score, receiver operating characteristic (ROC) curve, and Confusion

Matrix. Precision is a metric that measures how often a machine learning model

correctly predicts the positive class. F1 score is a metric that combines precision and

recall scores. A confusion matrix represents the prediction summary in matrix form. It

© 2024 Wolters Kluwer Health, Inc. All rights reserved. Unauthorized reproduction of the article is prohibited.
 shows how many predictions are correct and incorrect per class. Feature importance

was also evaluated to understand the contribution of each feature to the model

predictions.

D
Feature selection

TE
All patient data were used as variables in the prediction model. No feature-selection

method was used in this study.

EP
Statistical analysis

The data is reported as mean ± standard deviation. Subsequently, the dataset was
C

randomly split into training and testing sets at a ratio of 7:3. Using a 10-fold

cross-validation strategy, and each model was trained on the training data and
AC

subsequently assessed for its effectiveness on the testing data. To assess the

performance of the binary classification models, we calculated metrics including

Accuracy, AUC, Recall, Precision, and F1 scores. All computational processes were

conducted using PyCaret 2.2.3 in Python 3.10.9.

Results

© 2024 Wolters Kluwer Health, Inc. All rights reserved. Unauthorized reproduction of the article is prohibited.
 A total of 1922 female patients with AIS were included in this study. We excluded 654

patients: 105 with only a single visit, 517 with bracing or surgical intervention at the

first or second visit, eight with previous bracing or surgical intervention, and 24 with

missing values. After excluding 149 patients with advanced scoliosis at the first visit,

D
1119 patients (12.7 ± 1.5 years) were enrolled in this study (Figure 1). The patient

TE
demographics and radiographic data are presented in Table 1. According to the

Modified Lenke classification, 414 cases had altered curve types. Regarding the study

outcomes, 23 patients exhibited PT progression, 178 had MT progression, and 157

EP
showed TLL progression (Table 1).
C

ML model

1. PT Prediction model
AC

Considering the AUC and Recall, Linear Discriminant Analysis was used to predict PT

progression (Table 2). After tuning the hyperparameters, the performance of the Linear

Discriminant Analysis models was shown in Table 3. The performance of the model for

the validation set is summarized in Table 4. The ROC curve and confusion matrix of

the model are shown in Figure 2.

© 2024 Wolters Kluwer Health, Inc. All rights reserved. Unauthorized reproduction of the article is prohibited.
 2. MT Prediction model

Considering the AUC and Recall, Logistic Regression was used to predict MT

progression (Table 5). After tuning the hyperparameters, the performance of the

Logistic Regression models was shown in Table 3. The performance of the model for

D
the validation set is summarized in Table 4. The ROC curve and confusion matrix of

the model are shown in Figure 3.

3. TLL Prediction model

TE
EP
Considering the AUC and Recall, Logistic Regression was used to predict TLL

progression (Table 6). After tuning the hyperparameters, the performance of the
C

Logistic Regression models was shown in Table 3. The performance of the model for

the validation set is summarized in Table 4. The ROC curve and confusion matrix of
AC

the model are shown in Figure 4.

Feature importance

As shown in Figure 5, the most important factors for predicting the progression of each

curve were the initial Cobb angle of the PT for the progression of PT, the presence of

menarche for the progression of MT, and the Risser grade for the progression of TLL.

© 2024 Wolters Kluwer Health, Inc. All rights reserved. Unauthorized reproduction of the article is prohibited.
 For all models, the initial Cobb angle and skeletal growth indicators were important

factors, but the most important factors differed from model to model (Figure 5).

Discussion

D
This study showed that the progression of AIS at the last visit could be predicted with

TE
high accuracy based on data obtained at the first visit. In addition, this study showed

that the important factors in the angle progression of PT, MT, and TLL were different.

Our model predicted the progression of AIS at the last visit with a high degree of
EP
accuracy, using data from the first visit. Various factors have been reported to predict

the progression of AIS.6-8 The predictors of AIS progression are sex, initial Cobb angle,
C

curve type, thoracic intervertebral disc wedging, and skeletal growth indicators.6-8,15,16

However, the accuracy of spine surgeons in predicting progression based solely on

AC

anteroposterior whole-spine radiographs is documented to be below 50%.22 Multiple

prediction models for AIS progression have been previously reported in the

literature.15-17,20-22 Parent et al. reported that a prediction model using data from the

first visit was accurate, with a median error of 5.5°, with 80% of cases within 10°, and

that the accuracy was further improved using radiographic data from multiple visits.15

Using a statistical approach, a predictive model based on detailed radiographic features

© 2024 Wolters Kluwer Health, Inc. All rights reserved. Unauthorized reproduction of the article is prohibited.
 using computed tomography achieved a determination coefficient (R) of 0.643.16

Meanwhile, with the ML approach, using an efficient capsule network based on

biplanar spinal radiographs yielded an accuracy of 77.1%, sensitivity of 73.5%, and

specificity of 81.0%.20 A random forest regressor model based on multiple radiographs

D
predicted the last major Cobb angle within a 5° error.21 However, these prediction

radiographs and new imaging modalities.20

TE
models require more radiation exposure than anteroposterior and lateral whole-spine

This study also identified the factors predicting the progression of specific
EP
curves, namely PT, MT, and TLL. Few studies have reported predictors of progression

for each curve.32-34 The progression of each curve is associated with the magnitude of
C

the initial curve, the volume of the lumbar muscles, and the wedging of the vertebral

body.32-34 This study observed an association with the initial Cobb angle for all curves.
AC

However, menarche presence in MT patients and Risser grade in TLL patients were

most prominently linked to curve progression. This study is pioneering in reporting

crucial attributes for the progression of each curve. Notably, findings suggest that

skeletal growth's extent plays a pivotal role in MT and TLL progression. Further

research is warranted to elucidate the underlying reasons for the variations in strongly

associated factors among different curves.

© 2024 Wolters Kluwer Health, Inc. All rights reserved. Unauthorized reproduction of the article is prohibited.
 Radiation exposure is a particularly significant concern in pediatric patients, including

those with AIS. The lifetime cancer risk is higher in patients with AIS than in the

general population.11,12,35 Therefore, minimizing radiation exposure is necessary for the

comprehensive management of AIS. In addition to the physical burden, the

D
psychological burden associated with treating AIS is also important. Patients with AIS

TE
have been shown to have greater body image disturbances than healthy individuals,

and 32% of patients with AIS experience psychological and emotional distress.13,14,36

In addition, brace treatments are emotionally distressing.37-39 Spinal surgeons should

EP
pay close attention to this type of distress identifying patients at low risk of progression

in this model will help reduce the psychological and emotional burden by explaining
C

progression risk.

While the findings of this study suggested the potential to refine indications
AC

for brace treatment, they should be interpreted with caution. Our prediction model,

based on simple questionnaires and just two radiographs, can help spine surgeons to

optimize visit intervals and timing for intervention. It suggests the possibility of early

intervention to potentially reduce the progression to surgical thresholds. Yet, it's

important to acknowledge that the current model, while useful for screening and initial

counseling, may not yet warrant a significant shift in clinical practice regarding the

© 2024 Wolters Kluwer Health, Inc. All rights reserved. Unauthorized reproduction of the article is prohibited.
 initiation of brace treatment or the frequency of surveillance imaging. The pursuit of a

model with enhanced predictive accuracy remains a priority to ensure more definitive

guidance in managing patients with AIS, thereby minimizing their physical and

psychological burden.

D
This study presents several limitations. Firstly, not all patients had the same last visit,

TE
as data from some patients encompassed complete skeletal growth, while those

undergoing bracing were evaluated prior to such growth. However, the model's

essential capability in identifying patients at high risk of progressing to bracing or

EP
surgery was preserved, justifying the acceptable endpoint variability. Secondly, the

study lacked detailed features in its items. Prior research has established the relevance
C

of factors like spinal flexibility and morphological features in AIS progression.16,20,21

The absence of these elements likely impacted prediction accuracy. Nonetheless, the
AC

study's strength lies in achieving robust prediction accuracy using minimalistic

questionnaires and only two radiographs, thereby minimizing exposure. Thirdly,

patient bias may arise due to the study's location being an institution specializing in

scoliosis surgery. Patients were referred from various hospitals, potentially causing the

first visit date at our institution to differ from the date of AIS diagnosis. Despite this

bias, our model exhibited accurate predictions of AIS progression. Fourthly, the focus

© 2024 Wolters Kluwer Health, Inc. All rights reserved. Unauthorized reproduction of the article is prohibited.
 of this study is limited to female patients with AIS. Therefore, this model cannot

predict male patients with AIS or scoliosis other than AIS. However, a similar

approach could produce accuracy like the present model. Further research will be

needed.

D
Conclusions

TE
In conclusion, the ML-based model using items commonly evaluated at the first visit

accurately predicted angle progression in female patients with AIS. In addition, the

most important factors for predicting the progression of each curve were the initial
EP
Cobb angle of the PT for the progression of PT, the presence of menarche for the

progression of MT, and the Risser grade for the progression of TLL. Spine surgeons
C

can use the results of this study to consider the intervals between visits and the timing

of the intervention.
AC

© 2024 Wolters Kluwer Health, Inc. All rights reserved. Unauthorized reproduction of the article is prohibited.
 References

1. Kane WJ. Scoliosis prevalence: a call for a statement of terms. Clin Orthop Relat

Res 1977:43–6.

2. Weinstein SL, Dolan LA, Cheng JCY, et al. Adolescent idiopathic scoliosis.

D
Lancet 2008;371:1527–37. https://doi.org/10.1016/S0140-6736(08)60658-3

3.

TE
Morrissy RT, Goldsmith GS, Hall EC, et al. Measurement of the Cobb angle on

radiographs of patients who have scoliosis. Evaluation of intrinsic error. J Bone

Joint Surg Am 1990;72:320–7.

EP
4. Richards BS, Bernstein RM, D’Amato CR, et al. Standardization of criteria for

adolescent idiopathic scoliosis brace studies: SRS Committee on Bracing and

C

Nonoperative Management. Spine 2005;30:2068–75; discussion 2076–7.

https://doi.org/10.1097/01.brs.0000178819.90239.d0
AC

5. Negrini S, Donzelli S, Aulisa AG, et al. 2016 SOSORT guidelines: orthopaedic

and rehabilitation treatment of idiopathic scoliosis during growth. Scoliosis Spinal

Disord 2018;13:3. https://doi.org/10.1186/s13013-017-0145-8

6. Charles YP, Daures J-P, de Rosa V, et al. Progression risk of idiopathic juvenile

scoliosis during pubertal growth. Spine 2006;31:1933–42.

https://doi.org/10.1097/01.brs.0000229230.68870.97

© 2024 Wolters Kluwer Health, Inc. All rights reserved. Unauthorized reproduction of the article is prohibited.
 7. Dimeglio A, Canavese F. Progression or not progression? How to deal with

adolescent idiopathic scoliosis during puberty. J Child Orthop 2013;7:43–9.

https://doi.org/10.1007/s11832-012-0463-6

8. Bunnell WP. The natural history of idiopathic scoliosis before skeletal maturity.

D
Spine 1986;11:773–6. https://doi.org/10.1097/00007632-198610000-00003

9.

TE
Busscher I, Wapstra FH, Veldhuizen AG. Predicting growth and curve progression

in the individual patient with adolescent idiopathic scoliosis: design of a

prospective longitudinal cohort study. BMC Musculoskelet Disord 2010;11:93.

EP
https://doi.org/10.1186/1471-2474-11-93

10. Elattar EA, Saber NZ, Farrag DA. Predictive factors for progression of adolescent
C

idiopathic scoliosis: a 1-year study. Egyptian Rheumatology and Rehabilitation

2015;42:111–9. https://doi.org/10.4103/1110-161X.163943
AC

11. Cool J, Streekstra GJ, van Schuppen J, et al. Estimated cumulative radiation

exposure in patients treated for adolescent idiopathic scoliosis. Eur Spine J

2023;32:1777–86. https://doi.org/10.1007/s00586-023-07651-2

12. Farivar D, Skaggs DL, Gabriel K, et al. Breast Cancer Incidence, Mortality, and

Cost in Adolescent Idiopathic Scoliosis Patients and the Role of Low Dose

Biplanar Radiography. J Am Acad Orthop Surg 2023.

© 2024 Wolters Kluwer Health, Inc. All rights reserved. Unauthorized reproduction of the article is prohibited.
 https://doi.org/10.5435/JAAOS-D-23-00062.

13. Savvides P, Gerdhem P, Grauers A, et al. Self-Experienced Trunk Appearance in

Individuals With and Without Idiopathic Scoliosis. Spine 2020;45:522–7.

https://doi.org/10.1097/BRS.0000000000003308

D
14. Sanders AE, Andras LM, Iantorno SE, et al. Clinically Significant Psychological

TE
and Emotional Distress in 32% of Adolescent Idiopathic Scoliosis Patients. Spine

Deform 2018;6:435–40. https://doi.org/10.1016/j.jspd.2017.12.014

15. Parent EC, Donzelli S, Yaskina M, et al. Prediction of future curve angle using
EP
prior radiographs in previously untreated idiopathic scoliosis: natural history from

age 6 to after the end of growth (SOSORT 2022 award winner). Eur Spine J
C

2023;32:2171–84. https://doi.org/10.1007/s00586-023-07681-w

16. Nault M-L, Beauséjour M, Roy-Beaudry M, et al. A Predictive Model of

AC

Progression for Adolescent Idiopathic Scoliosis Based on 3D Spine Parameters at

First Visit. Spine 2020;45:605–11.

https://doi.org/10.1097/BRS.0000000000003316

17. Skalli W, Vergari C, Ebermeyer E, et al. Early Detection of Progressive

Adolescent Idiopathic Scoliosis: A Severity Index. Spine 2017;42:823–30.

https://doi.org/10.1097/BRS.0000000000001961

© 2024 Wolters Kluwer Health, Inc. All rights reserved. Unauthorized reproduction of the article is prohibited.
 18. Inoue T, Ichikawa D, Ueno T, et al. XGBoost, a Machine Learning Method,

Predicts Neurological Recovery in Patients with Cervical Spinal Cord Injury.

Neurotrauma Rep 2020;1:8–16. https://doi.org/10.1089/neur.2020.0009

19. Maki S, Furuya T, Yoshii T, et al. Machine Learning Approach in Predicting

D
Clinically Significant Improvements After Surgery in Patients with Cervical

TE
Ossification of the Posterior Longitudinal Ligament. Spine 2021;46:1683–9.

https://doi.org/10.1097/BRS.0000000000004125

20. Wang H, Zhang T, Cheung KM-C, et al. Application of deep learning upon spinal
EP
radiographs to predict progression in adolescent idiopathic scoliosis at first clinic

visit. EClinicalMedicine 2021;42:101220.

C

https://doi.org/10.1016/j.eclinm.2021.101220

21. Alfraihat A, Samdani AF, Balasubramanian S. Predicting curve progression for

AC

adolescent idiopathic scoliosis using random forest model. PLoS One

2022;17:e0273002. https://doi.org/10.1371/journal.pone.0273002

22. Yahara Y, Tamura M, Seki S, Kondo Y, Makino H, Watanabe K, et al. A deep

convolutional neural network to predict the curve progression of adolescent

idiopathic scoliosis: a pilot study. BMC Musculoskelet Disord 2022;23:610.

https://doi.org/10.1186/s12891-022-05565-6

© 2024 Wolters Kluwer Health, Inc. All rights reserved. Unauthorized reproduction of the article is prohibited.
 23. Reamy BV, Slakey JB. Adolescent idiopathic scoliosis: review and current

concepts. Am Fam Physician 2001;64:111–6.

24. Lenke LG, Edwards CC 2nd, Bridwell KH. The Lenke classification of adolescent

idiopathic scoliosis: how it organizes curve patterns as a template to perform

D
selective fusions of the spine. Spine 2003;28:S199–207.

TE
https://doi.org/10.1097/01.BRS.0000092216.16155.33

25. Sanders JO, Browne RH, McConnell SJ, Margraf SA, Cooney TE, Finegold DN.

Maturity assessment and curve progression in girls with idiopathic scoliosis. J

EP
Bone Joint Surg Am. 2007 ;89:64-73. https://doi.org/10.2106/JBJS.F.00067.

26. Mitchell R, Adinets A, Rao T, et al. XGBoost: Scalable GPU Accelerated Learning.
C

ArXiv, abs 2018;1806.11248. https://doi.org/10.48550/arXiv.1806.11248

27. Ke G, Meng Q, Finley T, et al. LightGBM: A highly efficient gradient boosting

AC

decision tree. Adv Neural Inf Process Syst 2017;3147–55

28. Breiman L. Random Forests. Mach Learn 2001;45:5–32.

29. Uddin S, Haque I, Lu H, et al. Comparative performance analysis of K-nearest

neighbour (KNN) algorithm and its different variants for disease prediction. Sci

Rep 2022;12:6256.

30. Prokhorenkova L, Gusev G, Vorobev A, et al. CatBoost: unbiased boosting with

© 2024 Wolters Kluwer Health, Inc. All rights reserved. Unauthorized reproduction of the article is prohibited.
 categorical features. arXiv [csLG] 2017.

31. Chawla NV, Bowyer KW, Hall LO, et al. SMOTE: Synthetic Minority

Over-sampling Technique. Jair 2002;16:321–57. https://doi.org/10.1613/jair.953

32. Sun X, Xie Y, Kong Q, et al. Segmental Characteristics of Main Thoracic Curves

D
in Patients with Severe Adolescent Idiopathic Scoliosis. World Neurosurg

TE
2018;119:e174–9. https://doi.org/10.1016/j.wneu.2018.07.086

33. Watanabe K, Ohashi M, Hirano T, et al. The Influence of Lumbar Muscle Volume

on Curve Progression After Skeletal Maturity in Patients With Adolescent

EP
Idiopathic Scoliosis: A Long-Term Follow-up Study. Spine Deform 2018;6:691–

8.e1. https://doi.org/10.1016/j.jspd.2018.04.003
C

34. Labrom FR, Izatt MT, Contractor P, et al. Sequential MRI reveals vertebral body

wedging significantly contributes to coronal plane deformity progression in

AC

adolescent idiopathic scoliosis during growth. Spine Deform 2020;8:901–10.

https://doi.org/10.1007/s43390-020-00138-w

35. Law M, Ma W-K, Lau D, et al. Cumulative effective dose and cancer risk for

pediatric population in repetitive full spine follow-up imaging: How micro dose is

the EOS microdose protocol? Eur J Radiol 2018;101:87–91.

https://doi.org/10.1016/j.ejrad.2018.02.015

© 2024 Wolters Kluwer Health, Inc. All rights reserved. Unauthorized reproduction of the article is prohibited.
 36. Auerbach JD, Lonner BS, Crerand et al. Body image in patients with adolescent

idiopathic scoliosis: validation of the Body Image Disturbance

Questionnaire--Scoliosis Version. J Bone Joint Surg Am 2014;96:e61.

https://doi.org/10.2106/JBJS.L.00867

D
37. Schwieger T, Campo S, Weinstein SL, et al. Body Image and Quality-of-Life in

TE
Untreated Versus Brace-Treated Females With Adolescent Idiopathic Scoliosis.

Spine 2016;41:311–9. https://doi.org/10.1097/BRS.0000000000001210

38. Asada T, Kotani T, Sunami T, et al. What factor induces stress in patients with AIS
EP
under brace treatment? Analysis of a specific factor using exploratory factor

analysis. J Orthop Sci 2021;26:999–1003.

C

https://doi.org/10.1016/j.jos.2020.10.024

39. Asada T, Kotani T, Nakayama K, et al. Japanese adaptation of the Bad Sobernheim
AC

Stress Questionnaire-Brace for patients with adolescent idiopathic scoliosis. J

Orthop Sci 2019;24:1010–4. https://doi.org/10.1016/j.jos.2019.08.003

© 2024 Wolters Kluwer Health, Inc. All rights reserved. Unauthorized reproduction of the article is prohibited.
 40.

Figure 1

The flowchart of study sample selection. AIS; adolescent idiopathic scoliosis

D
TE
EP
C
AC

© 2024 Wolters Kluwer Health, Inc. All rights reserved. Unauthorized reproduction of the article is prohibited.
 Figure 2Receiver operating characteristic curve (A) and Confusion matrix (B) of the

best model for predicting the progression of proximal thoracic curve on the validation

data.

D
TE
EP
C
AC

© 2024 Wolters Kluwer Health, Inc. All rights reserved. Unauthorized reproduction of the article is prohibited.
 Figure 3Receiver operating characteristic curve (A) and Confusion matrix (B) of the

best model for predicting the progression of the main thoracic curve on the validation

data.

D
TE
EP
C
AC

© 2024 Wolters Kluwer Health, Inc. All rights reserved. Unauthorized reproduction of the article is prohibited.
 Figure 4Receiver operating characteristic curve (A) and Confusion matrix (B) of the

best model for predicting the progression of thoracolumbar/lumbar curve on the

validation data.

D
TE
EP
C
AC

© 2024 Wolters Kluwer Health, Inc. All rights reserved. Unauthorized reproduction of the article is prohibited.
 Figure 5Feature importance for the best model for predicting the progression of each

curve. (A) Proximal thoracic curve prediction model; (B) Main thoracic curve

prediction model; (C) Thoracolumbar/lumbar curve prediction model.

D
TE
EP
C
AC

© 2024 Wolters Kluwer Health, Inc. All rights reserved. Unauthorized reproduction of the article is prohibited.
 Number of patients 1119

At first visit

Age (year) 12.7 ± 1.5

D
Body height (cm) 152.6 ± 7.6

TE
Body weight (kg) 43.5 ± 8.1

Menarche/Non-menarche 775/334

PT (°) 3.5 ± 6.9

EP
MT (°) 11.9 ± 8.1

TLL (°) 12.5 ± 8.0

TK (°) 19.2 ± 9.5

C

Risser grade 2.3 ± 1.7

AC

Modified Lenke classification First visit Last visit

1 175 152

2 111 108

3 185 192

4 113 128

5 291 323

© 2024 Wolters Kluwer Health, Inc. All rights reserved. Unauthorized reproduction of the article is prohibited.
 6 243 215

Classification change 414

Number of progression

patients

D
PT 23

MT

TLL
178

157
TE
EP
Table 1. Patient data used in this study.

Data are presented as mean ± standard deviation.

PT, proximal thoracic; MT, main thoracic. TLL, thoracolumbar/lumbar; TK, thoracic
C

kyphosis (T5-12)
AC

© 2024 Wolters Kluwer Health, Inc. All rights reserved. Unauthorized reproduction of the article is prohibited.
 Table 2. Comparison of model performances in predicting PT progression.

Model Accuracy AUC Recall Precision F1 score

Linear 0.86 0.94 0.90 0.12 0.21

Discriminant

D
Analysis

Logistic

Regression
0.93 0.93

TE 0.80 0.20 0.31

EP
Extreme Gradient 0.96 0.93 0.45 0.26 0.28

Boosting

Extra Trees 0.97 0.92 0.30 0.24 0.24

C

Classifier
AC

Random Forest 0.97 0.91 0.20 0.13 0.16

Classifier

PT, proximal thoracic; AUC, area under the receiver operating characteristic curve.

© 2024 Wolters Kluwer Health, Inc. All rights reserved. Unauthorized reproduction of the article is prohibited.
 Table 3. Model performance for each curve in training set after tuning the

hyperparameters.

Model Accuracy AUC Recall Precision F1 score

D
Prediction-PT model Linear 0.86 0.95 0.95 0.13 0.22

Discriminant

Analysis
TE
EP
Prediction-MT Logistic 0.85 0.90 0.83 0.50 0.63

model Regression

Prediction-TLL Logistic 0.75 0.84 0.81 0.35 0.49

C

model Regression
AC

AUC, area under the receiver operating characteristic curve; PT, proximal thoracic. MT,

main thoracic; TLL, thoracolumbar/lumbar.

© 2024 Wolters Kluwer Health, Inc. All rights reserved. Unauthorized reproduction of the article is prohibited.
 Table 4. Model performance for each curve in validation set.

Model Accuracy AUC Recall Precision F1 score

Prediction-PT model Linear 0.84 0.92 0.71 0.09 0.13

D
Discriminant

Prediction-MT
Analysis

Logistic
TE
0.84 0.90 0.87 0.49 0.63
EP
model Regression

Prediction-TLL Logistic 0.73 0.82 0.83 0.32 0.46

model Regression
C

AUC, area under the receiver operating characteristic curve; PT, proximal thoracic. MT,
AC

main thoracic; TLL, thoracolumbar/lumbar.

© 2024 Wolters Kluwer Health, Inc. All rights reserved. Unauthorized reproduction of the article is prohibited.
 Table 5. Comparison of model performance in predicting MT progression.

Model Accuracy AUC Recall Precision F1 score

Logistic 0.82 0.89 0.86 0.49 0.61

D
Regression

Extra Trees

Classifier
0.86

TE
0.90 0.57 0.58 0.56
EP
Random Forest 0.85 0.89 0.58 0.54 0.55

Classifier

Linear Discriminant 0.77 0.89 0.91 0.41 0.56

C

Analysis
AC

Light Gradient Boosting 0.84 0.89 0.61 0.54 0.56

Machine

MT, main thoracic; AUC, area under the receiver operating characteristic curve.

© 2024 Wolters Kluwer Health, Inc. All rights reserved. Unauthorized reproduction of the article is prohibited.
 Table 6. Comparison of model performances in predicting TLL progression.

Model Accuracy AUC Recall Precision F1 score

Logistic 0.75 0.84 0.82 0.36 0.49

D
Regression

Linear Discriminant

Analysis
0.72

TE
0.83 0.82 0.32 0.46
EP
Extra Trees 0.84 0.82 0.44 0.42 0.42

Classifier

Random Forest 0.83 0.81 0.44 0.41 0.41

C

Classifier
AC

Gradient Boosting 0.84 0.81 0.52 0.44 0.47

Classifier

TLL, thoracolumbar/lumbar; AUC, area under the receiver operating characteristic

curve.

© 2024 Wolters Kluwer Health, Inc. All rights reserved. Unauthorized reproduction of the article is prohibited.