Machine Learning–Based Prediction of Proximal Junctional Kyphosis in Adult Degenerative Scoliosis After Long-Segment Fusion: A Multicenter Training

Abstract

Study Design

Retrospective case–control study.

Objectives

To identify risk factors for proximal junctional kyphosis (PJK) after long-segment fusion in adult degenerative scoliosis (ADS) and to develop a machine learning–based prediction model with external validation.

Methods

We retrospectively analyzed 142 ADS patients from two institutions undergoing posterior long-segment fusion with ≥24 months follow-up. Patients from center A (n = 105) formed the training cohort, and those from center B (n = 37) served as the external validation cohort. Demographic, radiographic, and surgical parameters were compared between patients with and without PJK. Independent predictors were determined with multivariate logistic regression. Least absolute shrinkage and selection operator (LASSO) regression identified key variables. Six supervised machine learning algorithms were trained using center A data and validated on center B data. Model interpretability was assessed using Local Interpretable Model-agnostic Explanations (LIME).

Results

PJK occurred in 24 patients (16.9%). Logistic regression identified lower T-score, higher T1–pelvic angle, and female sex as independent predictors, with ASA grade III showing a marginal effect. LASSO retained five features: T score, ASA grade, T1PA, sacral slope, and pelvic incidence. Among algorithms, the back-propagation neural network with LASSO feature selection yielded the best discrimination (external validation AUC = 0.882). LIME analysis confirmed T score, T1PA, and PI as the most influential predictors.

Conclusions

Reduced bone density, impaired sagittal balance, and higher ASA grade increase PJK risk after long-segment fusion in ADS. A neural network combined with LASSO feature selection demonstrated superior predictive performance, supporting its potential for individualized preoperative risk assessment and surgical planning.

Keywords

adult degenerative scoliosis proximal junctional kyphosis machine learning predictive model LASSO regression neural network

Background

Adult degenerative scoliosis (ADS) is a common spinal deformity in adulthood, characterized by asymmetric disc degeneration and facet joint arthropathy, which can lead to spinal imbalance and a series of functional impairments.^1,2

The development of ADS is closely associated with progressive disc degeneration, dehydration and collapse, facet joint disease, and ligamentous laxity, eventually resulting in spinal instability, rotation, spondylolisthesis, and kyphotic deformity.¹ Epidemiological studies have reported a high prevalence of ADS in individuals over 60 years of age, with rates ranging from 8.3% to 68%, and a mean onset age of approximately 70 years. Given its typically progressive natural course, ADS not only causes back pain and lower-extremity neurological symptoms but also markedly reduces patients’ quality of life.³ While some patients may undergo conservative treatment, evidence suggests that such strategies often fail to achieve satisfactory outcomes. Consequently, surgery has become the primary treatment option for those with refractory pain, deformity progression, or severe neurological deficits.⁴ The surgical goals are to restore overall balance through decompression, deformity correction, and fusion with instrumentation.

However, adult spinal deformity surgery is associated with a relatively high complication rate, among which proximal junctional kyphosis (PJK) represents one of the most frequent modes of failure. A more severe manifestation is proximal junctional failure (PJF), characterized by vertebral fracture at the upper instrumented vertebra (UIV) or UIV+1, screw pullout, adjacent segment subluxation, pronounced instability, or associated neurological deficits. PJF often necessitates revision surgery, with an acute incidence of approximately 5.6%, and 41% of patients requiring reoperation.⁵ Both PJK and PJF contribute not only to reduced quality of life but also to significant increases in revision rates and overall healthcare costs.⁶

Numerous studies have investigated the risk factors for PJK/PJF, and their etiology is widely considered multifactorial, encompassing patient-, surgery-, and alignment-related parameters. Patient-specific risk factors such as advanced age, osteoporosis or low bone mineral density, sarcopenia, comorbidities, and smoking have all been reported.⁷

Despite the introduction of various preventive strategies in recent years—including minimizing proximal soft tissue disruption, careful UIV selection, application of transition rods or U-hooks, prophylactic vertebral augmentation, and preoperative osteoporosis treatment—no universally effective method has been established. Importantly, PJK/PJF continues to occur frequently even when radiographic parameters are optimized.⁵ Therefore, accurate preoperative risk prediction of PJK/PJF based on multiple clinical and radiographic factors is critical for improving surgical outcomes.⁸

In this context, machine learning (ML) offers a novel avenue of investigation. As a core methodology in artificial intelligence, ML can automatically detect complex patterns in large-scale medical imaging and clinical datasets, outperforming conventional linear models such as logistic regression.^9-13 The present study aimed to develop and validate a machine learning–based prediction model for PJK risk in ADS patients undergoing long-segment fusion, with the goal of enhancing risk assessment and clinical decision-making. Furthermore, we sought to explore the performance of different ML algorithms in small-sample data settings.

Methods

Study Design and Patient Population

From 2013 to 2025, consecutive patients diagnosed with ADS who underwent posterior spinal fusion at two independent tertiary care institutions were retrospectively reviewed. Patients from Center A (n = 105) were assigned to the training cohort, while those from Center B (n = 37) formed the external test cohort to evaluate model generalizability. All patients completed a minimum of 24 months of follow-up.The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments, and was approved by the institutional ethics committee. Owing to the retrospective nature of the analysis, the requirement for individual informed consent was waived.

The inclusion criteria were as follows:

Patients aged over 50 years;

Availability of preoperative spinal CT and MRI;

Preoperative Cobb angle greater than 10°;

Minimum follow-up duration of 24 months.

The exclusion criteria were:

History of childhood or adolescent scoliosis;

Secondary scoliosis due to other conditions, such as autoimmune diseases, tuberculosis, malignancy, post-traumatic deformity, or syndromic scoliosis;

Follow-up of less than 24 months.

Based on the occurrence of proximal junctional problems (PJK) during follow-up, patients were divided into two groups.

Outcomes

The diagnostic criteria for PJK were based on established definitions in prior literature: an increase in the proximal junctional angle (PJA) of ≥10° compared with preoperative measurement, or an absolute PJA of at least 10° at the final 24-month follow-up. Proximal junctional failure (PJF) was defined as a complication involving fracture at the proximal junction, fixation failure, or kyphosis requiring extension of fusion cranially.^14,15 All radiographic parameters were independently assessed by two spine surgeons; discrepancies were resolved by discussion with a third senior surgeon.

Data Collection

A total of 142 patients met the inclusion criteria. Baseline demographic and clinical variables collected included: age, sex, body mass index (BMI), smoking history, alcohol use, hypertension, chronic obstructive pulmonary disease (COPD), malignancy, and diabetes.

Radiographic parameters were measured from pre- and postoperative full-length standing radiographs, including: pelvic incidence (PI), pelvic tilt (PT), lumbar lordosis (LL), sacral slope (SS), sagittal vertical axis (SVA), and T1–pelvic angle (T1PA). Bone density was assessed using both the vertebral bone quality (VBQ) score from preoperative MRI and the femoral neck T-score from dual-energy X-ray absorptiometry (DEXA).

Surgical variables included fusion length and the American Society of Anesthesiologists (ASA) score.

Statistical Analysis

Continuous variables were presented as mean ± standard deviation and compared between groups using independent-sample t-tests, while categorical variables were expressed as counts and percentages and compared using chi-square tests. Variables identified as significant in univariate analyses were further tested in multivariate logistic regression to determine independent risk factors for PJK. The predictive performance of the regression model was also evaluated.

Machine Learning

To identify predictive features for PJK, least absolute shrinkage and selection operator (LASSO) regression was applied to all 20 candidate variables, with the optimal penalty parameter (λ) determined through 10-fold cross-validation. Variables with coefficients reduced to zero were excluded, and the remaining features were used for model development.

Six supervised learning algorithms were trained using both the selected features and the full feature set: Naïve Bayes, support vector machine (SVM), random forest, k-nearest neighbors (KNN), back-propagation neural network, and LightGBM. Each model was trained with 10-fold cross-validation on the training cohort (center A) and subsequently evaluated on the external validation cohort (center B). Model performance was assessed in terms of accuracy, precision, recall, F1 score, and area under the curve (AUC), with the best-performing model chosen based on external validation results.

To improve interpretability, Local Interpretable Model-agnostic Explanations (LIME) was applied to the optimal model. Global feature importance was quantified using mean weight statistics, and local interpretability was explored to assess individual variable contributions to case-level predictions.

Results

Patient Characteristics

A total of 142 patients were included, comprising 95 females and 47 males, with a mean age of 67.5 ± 6.55 years at the time of surgery. Among them, 24 patients (16.9%) developed PJK during follow-up. No significant differences were observed between the PJK and non-PJK groups with respect to age, BMI, fusion length, VBQ, Cobb angle, SVA, PT, or LL (all P > .05).However, several sagittal parameters demonstrated significant differences. Patients with PJK exhibited lower T score (P = .021), higher T1PA values (P = .026), and lower SS (P = .037) compared with controls. In addition, ASA grade distribution differed significantly between groups, with a higher proportion of grade III patients in the PJK group (P = .023). Other demographic and clinical factors—including sex, smoking, alcohol use, hypertension, COPD, malignancy, and diabetes—showed no significant differences (Tables 1 and 2).

Table 1.

Baseline Characteristics of Patients With and Without PJK

Variable	PJK	Mean	SD	P Value
Age (years)	No	67.254	6.516	0.323
Age (years)	Yes	68.708	6.708	0.323
BMI (kg/m²)	No	25.801	3.792	0.621
BMI (kg/m²)	Yes	25.321	6.340	0.621
Fusion length	No	5.203	2.178	0.608
Fusion length	Yes	4.958	1.876	0.608
T score	No	−1.154	1.066	0.021**
T score	Yes	−1.698	0.914	0.021**
VBQ score	No	2.926	3.393	0.693
VBQ score	Yes	2.638	0.706	0.693
Cobb angle (°)	No	19.559	12.326	0.257
Cobb angle (°)	Yes	23.071	19.509	0.257
SVA (mm)	No	36.335	20.992	0.676
SVA (mm)	Yes	38.380	25.578	0.676
PI (°)	No	50.697	13.178	0.091*
PI (°)	Yes	45.817	10.806	0.091*
PT (°)	No	22.319	9.810	0.744
PT (°)	Yes	23.050	10.909	0.744
LL (°)	No	−9.606	34.179	0.889
LL (°)	Yes	−8.554	30.213	0.889
SS (°)	No	28.322	12.123	0.037**
SS (°)	Yes	22.796	9.347	0.037**
T1PA (°)	No	15.327	7.117	0.026**
T1PA (°)	Yes	19.075	8.813	0.026**

Note. ***, **, and * indicate significance levels of 1%, 5%, and 10%, respectively.

Table 2.

Comparison of Demographic and Clinical Variables Between Patients With and Without PJK

Variable	Category	PJK = Yes	PJK = No	Total	P Value
Gender	Male	38	9	47	0.615
Gender	Female	80	15	95	0.615
ASA grade	II	98	15	113	0.023**
ASA grade	III	20	9	29	0.023**
Smoking	No	113	24	137	0.305
Smoking	Yes	5	0	5	0.305
Hypertension	No	54	11	65	0.995
Hypertension	Yes	64	13	77	0.995
Alcohol use	No	114	24	138	0.36
Alcohol use	Yes	4	0	4	0.36
COPD	No	117	24	141	0.651
COPD	Yes	1	0	1	0.651
Malignancy	No	117	23	140	0.208
Malignancy	Yes	1	1	2	0.208
Diabetes	No	91	21	112	0.256
Diabetes	Yes	27	3	30	0.256

Logistic Regression Analysis

Multivariate logistic regression achieved good predictive performance, with accuracy of 0.824, recall of 0.824, precision of 0.781, F1 score of 0.789, and AUC of 0.827 (Figure 1). Significant independent predictors of PJK were lower T score (P = .036), higher T1PA (P = .044), and female sex (P = .050). Specifically, higher T score was associated with decreased PJK risk, whereas higher T1PA increased risk; female sex appeared protective. ASA grade III showed a trend toward significance (OR = 3.258, 95% CI: 0.930-11.412, P = .065), suggesting it may represent a potential risk factor. Other variables, including PI, fusion length, SVA, Cobb angle, BMI, PT, SS, and comorbidities, were not significantly associated with PJK (Table 3).

Figure 1.

ROC curve of the multivariate logistic regression model predicting proximal junctional kyphosis (AUC = 0.827)

Table 3.

Multivariate Logistic Regression Analysis of Risk Factors for Proximal Junctional Kyphosis (PJK) After Long-Segment Fusion in Adult Degenerative Scoliosis

Variable	Coefficient	SE	P Value	95% CI (lower)	95% CI (upper)
PI	0.096	0.206	0.641	0.735	1.65
Fusion length	−0.02	0.158	0.900	0.719	1.337
VBQ score	0.088	0.346	0.798	0.554	2.152
SVA	0.001	0.012	0.939	0.978	1.024
Cobb angle	0.01	0.02	0.628	0.971	1.049
PT	−0.147	0.222	0.508	0.558	1.335
SS	−0.15	0.21	0.476	0.571	1.299
T score	−0.633	0.303	0.036**	0.293	0.96
T1PA	0.105	0.052	0.044**	1.003	1.231
Age	−0.021	0.048	0.665	0.892	1.075
LL	−0.001	0.01	0.914	0.98	1.018
BMI	0.034	0.058	0.552	0.924	1.159
Gender	−1.284	0.654	0.050**	0.077	0.997
Smoking	−15.61	9601.31	0.999	0
COPD	−17.497	7746.475	0.998	0
Malignancy	1.101	1.849	0.552	0.08	112.735
Hypertension	−0.143	0.567	0.800	0.285	2.634
Alcohol	−3.308	14 663.546	1.000	0
Diabetes	−0.728	0.785	0.354	0.104	2.25
ASA grade	1.181	0.64	0.065*	0.93	11.412

Feature Selection by LASSO Regression

A total of 20 candidate variables were considered for feature selection using LASSO regression with 10-fold cross-validation to reduce overfitting. The coefficient path is shown in Figure 2. As the regularization parameter (λ) increased, most coefficients approached zero. At the optimal λ value of 0.03441239—as indicated by the minimum mean squared error (MSE) in Figure 3—the model achieved its best predictive performance. Five predictors were ultimately retained (Figure 4), while others with coefficients shrunk to zero were excluded. Feature weight visualization demonstrated that T score and ASA were the most influential predictors, followed by T1PA, SS and PI were also retained.

Figure 2.

Coefficient path of variables in the LASSO regression model

Figure 3.

Ten-fold cross-validation curve for LASSO regression showing mean squared error across different λ values

Figure 4.

Predictors retained by LASSO regression: T score, ASA grade, T1 pelvic angle (T1PA), sacral slope (SS), and pelvic incidence (PI)

Machine Learning Model Performance

The predictive performance of six supervised algorithms before and after LASSO feature selection is shown in Figure 5 and summarized in Tables 4 and 5. LASSO generally improved model discrimination by reducing redundant variables. The back-propagation neural network achieved the highest test set performance after feature selection, with an AUC of 0.882, followed by SVM and random forest (Figures 6 and 7). Naïve Bayes, however, showed markedly reduced accuracy after LASSO. Based on these results, the back-propagation neural network was selected as the final prediction model for PJK risk.

Figure 5.

Receiver operating characteristic (ROC) curves of six supervised machine learning algorithms before and after LASSO feature selection, illustrating performance changes within each model

Table 4.

Predictive Performance of Six Supervised Machine Learning Algorithms on the External Validation Cohort Using the Full Set of Variables

Model	Naïve bayes	LightGBM	SVM	Back propagation neural network	KNN	Random forest
Accuracy	0.791	0.744	0.628	0.837	0.767	0.744
Recall	0.791	0.744	0.628	0.837	0.767	0.744
Precision	0.734	0.701	0.74	0.816	0.714	0.701
F1 score	0.75	0.72	0.666	0.806	0.735	0.72
AUC	0.824	0.8	0.844	0.882	0.774	0.814

Table 5.

Predictive Performance of Six Supervised Machine Learning Algorithms on the External Validation Cohort after LASSO Feature Selection

Model	Naïve bayes	LightGBM	SVM	Back propagation neural network	KNN	Random forest
Accuracy	0.186	0.744	0.581	0.814	0.814	0.814
Recall	0.186	0.744	0.581	0.814	0.814	0.814
Precision	0.432	0.701	0.787	0.663	0.663	0.663
F1 score	0.16	0.72	0.627	0.73	0.73	0.73
AUC	0.201	0.779	0.864	0.771	0.787	0.826

Figure 6.

Comparative ROC curves of the back-propagation neural network trained with the full variable set vs the LASSO-selected subset, showing improved performance after feature reduction

Figure 7.

Summary comparison of all machine learning models: (A) test set performance with the full variable set; (B) test set performance with LASSO-selected variables. The back-propagation neural network achieved the highest AUC among all algorithms

Model Interpretability

To enhance interpretability, the LIME method was applied to the final model. Global feature ranking identified PI, T score, and T1PA as the three most influential predictors, followed by ASA and SS. Notably, T score consistently emerged as a strong predictor across logistic regression, LASSO regression, and machine learning models, underscoring its stable importance in PJK risk stratification (Figure 8).

Figure 8.

Global feature importance ranking in the final model based on local interpretable model-agnostic explanations (LIME) analysis

Discussion

In this retrospective study of 142 patients with adult degenerative scoliosis (ADS) who underwent long-segment posterior spinal fusion, the overall incidence of proximal junctional kyphosis (PJK) was 16.9%. This rate is consistent with previous studies, supporting the external validity and clinical representativeness of our findings. Through multivariate regression and machine learning modeling, we identified several parameters closely associated with postoperative PJK, including pelvic incidence (PI), femoral neck T-score, T1–pelvic angle (T1PA), ASA grade, and sacral slope (SS). Among these, reduced T-score and increased T1PA—reflecting compromised bone quality and insufficient sagittal balance, respectively—emerged as the most significant risk factors. ASA grade III was also linked to higher PJK risk, underscoring the role of systemic frailty.

Our analysis highlights the importance of sagittal alignment parameters in PJK development. PI, a fixed morphologic parameter, determines the fundamental sagittal balance of an individual. Abnormal PI may predispose to mismatch following correction, leading to compensatory loads at the proximal junction. Elevated T1PA indicated insufficient sagittal compensation, integrating thoracic kyphosis and pelvic retroversion, and was associated with significantly higher PJK risk in this cohort. These findings align with prior reports that maintaining T1PA within 10°-30° may reduce junctional complications.¹⁶ Similarly, decreased SS further reflected limited pelvic compensatory reserve. Taken together, these results echo the view of Xu et al.¹⁷ that patients with impaired pelvic compensation are more susceptible to PJK, and suggest that modest PT overcorrection may be beneficial in such individuals.

Low bone density has repeatedly been linked to proximal junctional complications. Chen et al.¹⁸ demonstrated in a systematic review that reduced T-score and Hounsfield units significantly increase PJK risk. Our findings corroborate this association: patients with lower femoral neck T-scores were more prone to PJK, consistent with the concept that osteoporosis impairs screw–bone anchorage and predisposes the proximal vertebrae to stress fractures. Both logistic regression and machine learning models confirmed the stable predictive value of bone mineral density. Du et al.¹⁹ similarly identified BMD <−3.5 SD as a powerful risk factor for collapse in chronic osteoporotic vertebral fracture patients, achieving an AUC of 0.921. Our results extend these observations to the ADS population, emphasizing low bone quality as a central contributor to junctional failure.

In our cohort, only ASA II and III patients were included. ASA grade III patients were significantly more likely to develop PJK, highlighting the influence of systemic comorbidities and reduced physiological reserve. This is consistent with Hackett et al.,²⁰ who reported ASA grade as an independent predictor of postoperative complications and mortality in spine surgery. Our findings suggest that ASA classification remains a simple yet robust clinical tool for risk prediction, reinforcing the need to identify ASA III patients as high-risk candidates who may benefit from perioperative optimization and closer postoperative monitoring.

The observed risk factors reflect multiple interacting mechanisms. Low BMD weakens the screw–vertebra interface and predisposes to proximal junctional fracture^18,19,21; paraspinal muscle degeneration compromises local biomechanical stability^1,3; impaired pelvic compensation shifts excessive loads to the thoracic spine¹⁷; and poor systemic reserve in ASA III patients limits healing and compensatory capacity.²⁰ Together, these interrelated mechanisms explain the multifactorial etiology of PJK.

Interestingly, the regression model indicated that female sex was associated with a lower risk of PJK (OR <1), which appears counterintuitive given the higher prevalence of osteoporosis among women. This finding should be interpreted cautiously. One possible explanation is that male patients in our cohort tended to present with more severe baseline deformity and a higher proportion of ASA grade III, reflecting greater systemic comorbidity and frailty. Alternatively, the apparent protective effect may reflect sampling variability due to the relatively small cohort size. Larger multicenter studies are warranted to clarify the true direction and magnitude of the sex-related association with PJK risk.

We compared six supervised machine learning algorithms to assess predictive performance. The back-propagation neural network (BP) combined with LASSO feature selection achieved the best performance (test set AUC = 0.882), outperforming SVM, random forest, LightGBM, and other models. BP networks are well suited to capture complex nonlinear relationships, while LASSO mitigated overfitting by excluding redundant features, thus enabling effective learning from a limited sample. Moreover, the relatively shallow architecture of the BP network used in this study, combined with LASSO-based dimensionality reduction, allowed the model to maintain adequate representational power while avoiding overfitting—a common issue in small-sample datasets. Unlike SVMs, which rely on kernel transformations that may become unstable with limited observations, or Random Forests, which can introduce variance through repeated resampling, the BP architecture adjusted its parameters iteratively through gradient descent, achieving balanced bias–variance trade-offs. This likely explains its superior external performance despite the restricted sample size.

By contrast, Naïve Bayes relied on strong independence assumptions, KNN was more vulnerable to noise, and ensemble models such as random forest and LightGBM exhibited tendencies toward overfitting in the small-sample setting. Although SVM performed relatively well, its accuracy and stability were inferior to BP networks.

Our results suggest that in small datasets, the strategy of “feature selection plus moderately complex models” may maximize predictive performance. Compared with prior models, such as Tian’s logistic regression based on radiographic parameters (AUC = 0.806)²² and Wang’s random forest model (AUC = 0.847),²³ our BP + LASSO model achieved superior performance, approaching state-of-the-art levels. This demonstrates the promise of neural networks, when combined with appropriate dimensionality reduction, in enhancing individualized risk prediction for PJK.

Importantly, this study represents a multicenter design, with training and testing performed on independent institutional datasets. The external validation confirmed the robustness of the BP + LASSO model (AUC = 0.882), supporting its potential generalizability to broader clinical populations.

This study has several limitations. First, although a multicenter design with external validation was adopted, the overall sample size remained relatively small, which may limit generalizability; larger prospective studies are required for further validation. Second, only preoperative demographic and radiographic variables were included, without intraoperative or postoperative management factors such as upper instrumented vertebra selection or prophylactic measures. The exclusion of these variables may limit the comprehensiveness of the model, and future work should consider incorporating them to better reflect the multifactorial nature of PJK development. Third, the median follow-up was less than 3 years, restricting conclusions regarding long-term outcomes. Given that mechanical and biological complications of adult degenerative scoliosis surgery may continue to accumulate over time, a longer follow-up period (ideally ≥5 years) would provide a more comprehensive understanding of the temporal evolution and durability of surgical outcomes. Future prospective studies with extended follow-up are therefore warranted to validate the long-term predictive accuracy of the proposed model. Additionally, although this study incorporated an external validation cohort, the relatively small sample size of the validation set (n = 37) may limit the stability and generalizability of the reported AUC values. Small test sets can lead to greater variability in performance metrics, particularly in machine learning models trained on limited data. Accordingly, future studies involving larger multicenter datasets are necessary to further validate and calibrate the model’s robustness across diverse populations.

Future research should focus on prospective cohorts with larger sample sizes to validate and refine predictive models. Incorporating multimodal data—including radiomics, muscle quality, and advanced imaging biomarkers—with deep learning techniques may further enhance predictive power. The development of clinician-friendly, interpretable prediction tools will also be essential to facilitate real-world adoption, enabling rapid preoperative risk stratification and personalized surgical planning.²⁴

Footnotes

Author Note

Present or permanent address Beijing Anzhen Hospital, Capital Medical University, 2 Anzhen Road, Chaoyang District, Beijing, China.

Acknowledgments

We thank the National Natural Science Foundation of China and the R&D Program of Beijing Municipal Education Commission for their financial support. We also acknowledge the contributions of the Chinese Institutes for Medical Research, Beijing, for providing essential resources. The Manuscript submitted does not contain information about medical device(s)/drug(s).

ORCID iD

Sheyang Xu

Ethical Considerations

As this study utilized retrospective data analysis, a waiver of informed consent was granted by the ethics committee in accordance with national regulations and the Declaration of Helsinki. This study was approved by the Ethics Committee of Beijing Anzhen Hospital (Ethical Approval No. 2025228x). This study was approved by the Ethics Committee of the hospitals. All procedures performed involving human participants were conducted in accordance with the ethical standards of the institutional and national research committee and with the 1964 Helsinki Declaration and its later amendments.

Consent to Participate

Owing to the retrospective nature of the analysis, the requirement for individual informed consent was waived.

Author Contributions

All authors contributed to the study conception and design. Yong Hai supervised the study. Xinglin Liu did the material preparation and data collection work, machine learning, analysis and manuscript writing were performed by Xianglong Meng and Sheyang Xu, Zhiheng Zhao did the visual work and all authors commented on previous versions of the manuscript. All authors read and approved the final manuscript.

Funding

The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: We thank the National Natural Science Foundation of China and the R&D Program of Beijing Municipal Education Commission for their financial support. We also acknowledge the contributions of the Chinese Institutes for Medical Research, Beijing, for providing essential resources. The Manuscript submitted does not contain information about medical device(s)/drug(s).

Declaration of Conflicting Interests

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Data Availability Statement

Data availability is not applicable to this study.*

References

Ailon

Smith

Shaffrey

, et al. Degenerative spinal deformity. Neurosurgery. 2015;77(Suppl 4):S75-S91. doi:10.1227/NEU.0000000000000938

Birknes

White

Albert

Shaffrey

Harrop

. Adult degenerative scoliosis: a review. Neurosurgery. 2008;63:94-103. doi:10.1227/01.NEU.0000325485.49323.B2

Yuan

Zeng

Chen

Zhang

Mai

. Degenerative lumbar scoliosis patients with proximal junctional kyphosis have lower muscularity, fatty degeneration at the lumbar area. Eur Spine J. 2021;30:1133-1143. doi:10.1007/s00586-020-06394-8

Phan

Maharaj

, et al. Outcomes of short fusion versus long fusion for adult degenerative scoliosis: a systematic review and meta-analysis. Orthop Surg. 2017;9(4):342-349. doi:10.1111/os.12357

Mika

Mesfin

Rubery

Molinari

Kebaish

Menga

. Proximal junctional kyphosis: a pediatric and adult spinal deformity surgery dilemma. JBJS Rev. 2019;7(4):e4. doi:10.2106/JBJS.RVW.18.00059

Cerpa

Sardar

Lenke

. Revision surgery in proximal junctional kyphosis. Eur Spine J. 2020;29(1):78-85. doi:10.1007/s00586-020-06320-y

Teles

Aldebeyan

Aoude

, et al.

Mechanical complications in adult spinal deformity surgery: can spinal alignment explain everything?

Spine. 2022;47:E1-E9.

Karikari

Metz

. Preventing pseudarthrosis and proximal junctional kyphosis. Neurosurg Clin. 2018;29(3):365-374. doi:10.1016/j.nec.2018.03.005

Yang

Hai

Zhang

. The role of lumbosacral paraspinal muscle degeneration and low vertebral bone mineral density on distal instrumentation-related problems following long-instrumented spinal fusion for degenerative lumbar scoliosis: a retrospective cohort study. Quant Imag Med Surg. 2023;13(7):4475-4492.

10.

Shin

Austin

Ross

, et al. Machine learning vs. conventional statistical models for predicting heart failure readmission and mortality. ESC Heart Fail. 2021;8(1):106-115.

11.

Sofo

Caprino

Schena

Sacchetti

Potenza

Ciociola

. New perspectives in the prediction of postoperative complications for high-risk ulcerative colitis patients: machine learning preliminary approach. Eur Rev Med Pharmacol Sci. 2020;24(24):12781-12787.

12.

Tewari

Kashdan

Walsh

Martin

Parnell

O’Leary

. Estimating the conditional probability of developing HPV related oropharyngeal cancer by combining machine learning and inverse Bayesian modelling. PLoS Comput Biol. 2021;17(8):e1009289.

13.

Wang

Fan

Yang

Lin

Yang

. Development and internal validation of supervised machine learning algorithms for predicting the risk of surgical site infection following minimally invasive transforaminal lumbar interbody fusion. Front Med. 2021;8:771608.

14.

Lau

Clark

Scheer

, et al. Proximal junctional kyphosis and failure after spinal deformity surgery: a systematic review. Spine. 2014;39:2093-2102.

15.

Doodkorte

RJP

Vercoulen

TFG

Roth

de Bie

Willems

. Instrumentation techniques to prevent proximal junctional kyphosis and proximal junctional failure in adult spinal deformity correction: a systematic review of biomechanical studies. Spine J. 2021;21:842-854.

16.

Fisher

Das

Yung

, et al. An analysis of the usage and limitations of the T1 pelvic angle. Bone Joint Lett J. 2025;107-B(3):346-352. doi:10.1302/0301-620X.107B3.BJJ-2024-0800.R2

17.

Fan

Qiu

Liu

Zhu

. Effect of pelvic compensation capacity on proximal junctional kyphosis: a stratified analysis of pelvic tilt in adult spinal deformity surgery. J Orthop Surg Res. 2025;20(1):675. doi:10.1186/s13018-025-06103-5

18.

Chen

McCandless

Bhandarkar

, et al. The association between bone mineral density and proximal junctional kyphosis in adult spinal deformity: a systematic review and meta-analysis. J Neurosurg Spine. 2023;39(1):82-91. doi:10.3171/2023.2.SPINE221101

19.

Jiang

Zhu

Luo

. A predictive scoring system for proximal junctional kyphosis after posterior internal fixation in elderly patients with chronic osteoporotic vertebral fracture: a single-center diagnostic study. Front Endocrinol. 2022;13:923778. doi:10.3389/fendo.2022.923778

20.

Hackett

De Oliveira

Jain

Kim

JYS

. ASA class is a reliable independent predictor of medical complications and mortality following surgery. Int J Surg. 2015;18:184-190. doi:10.1016/j.ijsu.2015.04.079

21.

Liu

Zhu

Wang

Duan

Liu

. Radiological analysis of thoracolumbar junctional degenerative kyphosis in patients with lumbar degenerative kyphosis. Chin Med J. 2017;130(21):2535-2540. doi:10.4103/0366-6999.217090

22.

Tian

, et al. Prediction of proximal junctional kyphosis and failure after corrective surgery for adult spinal deformity: an MRI-Based model combining bone and paraspinal muscle quality metrics. Spine J. 2024;24(12):2389-2399. doi:10.1016/j.spinee.2024.08.010

23.

Wang

Zhao

Chen

. Novel risk factors and personalized risk calculator for predicting proximal junctional kyphosis after adult spinal deformity surgery. Bone Joint Lett J. 2025;107-B(8):829-838. doi:10.1302/0301-620X.107B8.BJJ-2024-1474.R2

24.

Park

Kwon

Choi

, et al. Perioperative teriparatide for preventing proximal junctional kyphosis and failure in patients with osteoporosis after adult thoracolumbar spinal deformity surgery: a prospective randomized controlled trial. Osteoporos Int. 2025;36(5):833-843. doi:10.1007/s00198-025-07449-6

Machine Learning–Based Prediction of Proximal Junctional Kyphosis in Adult Degenerative Scoliosis After Long-Segment Fusion: A Multicenter Training–Validation Study

Abstract

Study Design

Objectives

Methods

Results

Conclusions

Keywords

Background

Methods

Study Design and Patient Population

Outcomes

Data Collection

Statistical Analysis

Machine Learning

Results

Patient Characteristics

Logistic Regression Analysis

Feature Selection by LASSO Regression

Machine Learning Model Performance

Model Interpretability

Discussion

Footnotes

Author Note

Acknowledgments

ORCID iD

Ethical Considerations

Consent to Participate

Author Contributions

Funding

Declaration of Conflicting Interests

Data Availability Statement

References