Pelvic Incidence-Dependent Clustering of Sagittal Spinal Alignment in Asymptomatic Middle-Aged and Elderly Adults: A Machine Learning Approach

Abstract

Study Design.

A cross-sectional cohort study.

Objective.

This study aimed to refine the sagittal morphologic classification of the spine in asymptomatic middle-aged and elderly adult populations using the unsupervised machine learning (ML) techniques and, by leveraging these findings, to propose and validate a surgical correction reference for adult spinal deformity (ASD) patients across different morphologic subtypes.

Summary of Background Data.

Restoration of sagittal alignment is the key to preventing mechanical complications and achieving good clinical outcomes in ASD surgery. However, high variations in the reported incidence of mechanical complications and clinical outcomes under current ASD realignment strategies have severely impeded the decision-making process for the optimal surgical plan.

Materials and Methods.

This study cross-sectionally enrolled asymptomatic middle-aged and elderly Chinese adults. Sagittal spinal morphology clusters and pelvic incidence-based correction criteria for ASD realignment surgery were derived from whole spine radiographs using unsupervised ML algorithms. To externally validate the realignment strategy identified in asymptomatic adults, a consecutive cohort of ASD patients with sagittal deformity who underwent realignment surgery was examined for postoperative mechanical complications, unplanned reoperation, unplanned readmission, and clinical outcomes during follow-up.

Results.

A total of 635 asymptomatic adults were enrolled for morphologic stratification, and 103 ASD patients with sagittal deformity were included for validation. The unsupervised ML algorithm successfully stratified spinal morphology into four clusters. The pelvic incidence-based surgical correction criteria computed by the regression algorithm demonstrated plausible clinical relevance, evidenced by the significantly lower incidence of postoperative mechanical complications, unplanned reoperation, unplanned readmission, and superior patient-reported outcomes in the restored group (conforming to the correction criteria) during follow-up.

Conclusion.

In this study, unsupervised ML algorithm effectively partitioned asymptomatic sagittal spinal morphology into four distinct clusters. Using the pelvic incidence-based proportional correction criteria, ASD patients can anticipate a reduced incidence of mechanical complications and improved clinical outcomes following spinal realignment surgery.

Level of Evidence.

Level Ⅲ.

Restoration of sagittal alignment in adult spinal deformity (ASD) surgery is of paramount importance to achieve a favorable clinical outcome and prevent mechanical complications,¹ and guiding ASD realignment surgery toward the physiological morphology of the normal spine could ensure consistent and improved ASD correction outcomes.² Despite the growing body of literature exploring this objective, existing ASD realignment strategies have shown considerable heterogeneity in the reported incidence of postoperative mechanical failure and clinical outcomes,^1,3–8 which has severely hampered their clinical application. This is because the reference populations used to derive correction targets in these studies had different clinical backgrounds (e.g. different age groups, participants with different symptomatic scales, etc.)^1,9–11 and the propositions of these studies were mostly derived from empirical human judgment.^4,9,12 Consequently, there is an imperative need for a more robust, evidence-based methodology to identify sagittal spinal morphologic variations in standardized asymptomatic cohorts. And a novel surgical correction reference for ASD patients could be established based on the findings from the healthy spine.

In addition, classification schemes such as Lenke-Silva levels of treatment,¹³ Roussouly classification,^9,12 Scoliosis Research Society (SRS)-Schwab classification,¹⁰ age-adjusted sagittal alignment goals,^8,11 and Global Alignment and Proportion (GAP) score¹ have emerged to help achieve a tailored surgical treatment for ASD patients by conceptualizing population subgroups with high intraclass similarities in spinal morphologies, clinical presentations, and postoperative recovery trajectories, and providing each subgroup with a tailored treatment target. Among these, the latter three classifications were primarily established based on the incidence of postoperative mechanical complications^1,11 or clinical outcomes¹⁰ observed in patient populations. The Roussouly classification, however, stood out as the only one to define sagittal spinal correction criteria using asymptomatic subjects,^4,9,12 thereby enabling surgical outcomes that more closely approximate normal spinal parameters and physiology. The five morphologic subtypes of normal spinal sagittal alignment proposed by Roussouly and colleagues were grounded in sacral slope (SS) and spinal curvature morphology.⁹ Nonetheless, the variability in SS during spinal degeneration posed challenges in retrodicting the normal sagittal alignment subtype based on SS measurements obtained during the disease state. Moreover, the applicability of the Roussouly classification to the Chinese population may be compromised by potential racial differences.^14,15

In recent years, pelvic incidence (PI) has emerged as a pivotal measure for assessing sagittal spinal alignment in both patients with ASD and asymptomatic individuals.^4,9,12 Owing to its constancy throughout sagittal spinal degeneration, PI has been recognized as an optimal biomarker for the backward estimation of ideal spinal morphology during the surgical planning phase.^12,16,17 However, studies categorizing spinal morphology using PI have predominantly relied on empirical knowledge.^9,12 Furthermore, the high dimensionality of spinal sagittal alignment data hindered the investigations of the interplays between PI and the whole-spine morphology. An initiative to address these challenges is the adoption of unsupervised machine learning (ML), a data mining approach that can discern patterns within high-dimensional data sets without necessitating subjective human assumptions, thereby ensuring objectivity.¹⁸ To date, there has been no research that employs ML to integrate comprehensive whole-spine parameters for the objective classification of sagittal spinal morphology in asymptomatic adults, nor to offer individualized guidance for the realignment surgery of ASD patients.

In summary, the present study aimed to propose and validate a surgical correction reference for ASD patients based on different sagittal morphologic subtypes of the spine recognized in asymptomatic populations using unsupervised ML. This proposed categorization framework is anticipated to enable the implementation of personalized spinal surgical realignment strategies, ultimately enhancing clinical outcomes, and decreasing the incidence of mechanical complications.

MATERIALS AND METHODS

Study Design and Participants

The overall study design was shown in Figure 1. This study cross-sectionally collected 1250 standing radiographs of the whole-spine from middle-aged and elderly Chinese community-based adult volunteers (age 45 yr or older, the prevalent age groups of ASD¹⁹) between August 2021 and May 2022. Following the exclusion criteria, a total of 635 asymptomatic participants [visual analog scale (VAS) <2, neck disability index (NDI) <20%, and Oswestry disability index (ODI) <20%] were enrolled in the analysis to derive the sagittal spinal morphology clusters and pelvic incidence-based correction criteria for ASD realignment surgery (Fig. 2). In addition, a cohort of 458 consecutive ASD patients (age above 50 yr) who underwent realignment surgery at our center from May 2017 to April 2023 was retrospectively reviewed, and 103 cases with sagittal deformity were included for the external validation of the realignment strategy identified in asymptomatic adults (Fig. 3). Cohort information and inclusion/exclusion criteria were provided in the Supplementary Methods, Supplemental Digital Content 1, http://links.lww.com/BRS/C753.

Figure 1.

Flow diagram showing the overall design of this study. ASD indicates adult spinal deformity; KNN, k-nearest neighbors; NNET, neural networks; PI, pelvic incidence; RF, random forests; t-SNE, t-Distributed Stochastic Neighbor Embedding.

Figure 2.

Participant recruitment flowchart for the cross-sectional study. NDI indicates neck disability index; ODI, Oswestry disability index; VAS, visual analog scale.

Figure 3.

Flowchart for inclusion and exclusion of ASD patient validation cohort. ASD indicates adult spinal deformity.

This study was approved by the regional medical ethics committee and conducted in accordance with the Declaration of Helsinki (Ethical review committee of Xuanwu Hospital, Capital Medical University; KS2022150-1 and KS2022151-1). Written informed consent was obtained from all participants. The Strengthening the Reporting of Observational studies in Epidemiology (STROBE, Supplemental Digital Content 7, http://links.lww.com/BRS/C759) guideline was followed to report this study.

Study Measures

The cross-sectional study measured 21 commonly used sagittal parameters of the cervical, thoracic, lumbar, pelvic and global spine.⁹ Detailed illustrations and descriptions of each parameter were provided in Figure S1, Supplemental Digital Content 2, http://links.lww.com/BRS/C754 and Table S1, Supplemental Digital Content 3, http://links.lww.com/BRS/C755. Basic demographic data including age, sex, height, weight, and body mass index (BMI) were also collected. Clinical and radiologic characteristics were manually checked to identify missing values, erroneous calculations, and possible outliers.

Generation of Morphologic Clusters

To generate the sagittal spinal morphologic clusters, principal component analysis (PCA) was first used to eliminate the inherent correlation and collinearity within the sagittal parameters.²⁰ After PCA, the coordinates on the selected principal components (PCs) were used to perform k-means clustering using Euclidean distance.²¹ The optimal number of clusters was determined by visually examining the reduction in the sum of squared distances with changes in the number of clusters. The detailed descriptions of PCA and k-means clustering were provided in the Supplementary Methods, Supplemental Digital Content 1, http://links.lww.com/BRS/C753.

Development of Morphologic Cluster Classifiers

To facilitate the morphologic cluster recognition in clinical practice, we developed four predictive models, including an easy-to-use PI cutoff-based classification model (PI model) and three ML models of varying complexity [k-nearest neighbors (KNN), random forest (RF), and neural networks (NNET)]. The PI model was established because PI is constant during spinal degeneration (does not change after skeletal maturity) and can provide a backward estimation of the ideal correction target for ASD surgery. Therefore, it could bridge the gap between the ASD and the healthy spine.

To investigate model robustness, the cross-sectional cohort was randomly split into training (70%) and test (30%) sets for model construction and validation. The performance of the ML models was compared using the area under the receiver operating characteristic curve (AUROC) and the area under the precision-recall curve (AUPRC). Accuracy, specificity, sensitivity, positive predictive value (PPV), negative predictive value (NPV), and F1-score were used to compare the performance of PI model and ML models.

We followed the best practices recommended by the Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis (TRIPOD, Supplemental Digital Content 8, http://links.lww.com/BRS/C760) guideline to design and benchmark our ML algorithms. Detailed information on model configurations and evaluations was provided in the Supplementary Methods, Supplemental Digital Content 1, http://links.lww.com/BRS/C753.

Proposing and Validating the Surgical Correction References

First, we used recursive feature elimination (RFE), information gain and Boruta algorithms to identify the most influential variables of the new classification for surgical correction. Next, we applied regression models to calculate the normative range of the surgically modifiable determinant variables to establish the correction criteria for ASD realignment surgery. PI, the pivotal marker for estimating ideal alignment from spinal morphology in degenerative conditions,¹² was adopted as the sole model input because it could bridge the gap between the ASD and the healthy spine due to its constancy during spinal degeneration, which could also provide a convenient approach for clinical application.

In addition, we externally validated the correction references in a retrospective cohort consisting of 103 ASD patients with sagittal deformity who underwent realignment surgery. Patients were divided into the restored and nonrestored groups based on the consistency of their postoperative sagittal spinal parameters with the correction targets derived from the algorithm developed in the present study. The comparison between the two groups encompassed postoperative mechanical complications [e.g. proximal junctional kyphosis (PJK)], instances of unplanned readmission and reoperation, as well as patient-reported outcomes over the course of follow-up. Detailed information on the selection of determinant parameters, prediction of normative values, and validation of the ASD cohort were provided in the Supplementary Methods, Supplemental Digital Content 1, http://links.lww.com/BRS/C753.

Comparison Against Established Classification Framework

The spinal curvatures of each identified subtype were visualized by referencing the mean or median values of the respective determinant parameters. These findings were subsequently benchmarked against the Roussouly classification, which constituted the sole validated classification scheme for spinal morphologic types among asymptomatic individuals. To further elucidate the relationship between the Roussouly classification and the novel classification proposed in this study, a Sankey plot was utilized. In addition, the t-Distributed Stochastic Neighbor Embedding (t-SNE) algorithm was used to evaluate the discriminative capacity of each classification method. This assessment was conducted by calculating the Euclidean distances between the centroids of each subtype. The intricacies of the t-SNE algorithm were detailed in the Supplementary Methods, Supplemental Digital Content 1, http://links.lww.com/BRS/C753.

Statistical Analysis

Normal and non-normal data were presented using mean (SD) and median [interquartile range (IQR)], respectively. Categorical variables were expressed as frequencies with percentages. Statistical variables were compared using χ² test, the Fisher exact test, analysis of variance, Kruskal-Wallis test, t test, and Wilcoxon test as appropriate. The false discovery rate (FDR) was used to correct for multiple comparisons. All analyses were completed using R version 4.3.2 (R Project for Statistical Computing). A two-sided P < .05 was considered to be statistically significant.

RESULTS

Of the 635 asymptomatic middle-aged and elderly participants included in the analysis, the mean (SD) age was 66.2 (6.1) years. Our sample included 483 female participants (76.1%) (Table 1). In general, no features had missing values, erroneous calculations, or physical outliers. Some parameters were highly correlated with each other (the Spearman rho ≥0.8) with each other (Table S2, Supplemental Digital Content 4, http://links.lww.com/BRS/C756).

TABLE 1.

Clinical and Radiologic Characteristics by Spinal Morphologic Subtypes

Characteristics	Total (N=635), n (%)	Subtype 1 (N=103), n (%)	Subtype 2 (N=220), n (%)	Subtype 3 (N=199), n (%)	Subtype 4 (N=113), n (%)	P *
Age, median (IQR) (yr)	65.7 (62.6–69.7)	66.3 (63.4–71.3)	65.1 (62.2–68.9)	66.0 (62.7–70.2)	65.7 (62.6–69.5)	0.15
Sex
Male	152 (23.9)	36 (35.0)	50 (22.7)	49 (24.6)	17 (15.0)	0.007
Female	483 (76.1)	67 (65.0)	170 (77.3)	150 (75.4)	96 (85.0)
Height, median (IQR), cm	160.0 (157.0–166.0)	163.0 (159.0–170.0)	160.0 (156.0–165.0)	160.0 (156.0–167.0)	160.0 (157.0–164.0)	0.003
Weight, median (IQR), kg	63.0 (57.3–70.0)	65.0 (60.0–75.0)	63.0 (56.5–70.0)	63.0 (58.0–70.0)	60.0 (57.0–67.0)	0.02
BMI, median (IQR), kg/m2	24.2 (22.2–26.4)	24.4 (22.5–26.6)	24.2 (22.0–26.4)	24.7 (22.6–26.4)	23.7 (22.2–25.9)	0.62
SVA, mean (SD), mm	-5.8 (28.6)	-1.6 (35.8)	-11.1 (26.6)	-4.5 (28.1)	-1.6 (24.0)	0.003
TPA, median (IQR), °	9.5 (5.9–13.8)	6.9 (3.4–11.6)	7.6 (4.4–11.3)	10.2 (7.6–13.5)	14.8 (10.7–18.8)	<0.001
SSA, median (IQR), °	126.2 (120.2–131.6)	116.6 (111.5–121.5)	124.9 (120.1–130.1)	128.6 (123.8–132.7)	132.9 (127.1–138.5)	<0.001
O–C2, mean (SD), °	18.2 (7.0)	19.3 (7.2)	18.0 (7.2)	18.2 (6.6)	17.5 (7.1)	0.31
C2–7, mean (SD), °	10.4 (10.2)	10.7 (10.4)	8.5 (9.6)	11.9 (10.6)	11.2 (9.9)	0.004
C-apex
C2	8 (1.3)	1 (1.0)	4 (1.8)	3 (1.5)	0	0.14
C3	21 (3.3)	7 (6.8)	1 (0.5)	6 (3.0)	7 (6.2)
C4	254 (40.0)	44 (42.7)	81 (36.8)	83 (41.7)	46 (40.7)
C5	324 (51.0)	46 (44.7)	123 (55.9)	99 (49.7)	56 (49.6)
C6	25 (3.9)	4 (3.9)	9 (4.1)	8 (4.0)	4 (3.5)
C7	1 (0.2)	1 (1.0)	0	0	0
T1	2 (0.3)	0	2 (0.9)	0	0
C2–7 SVA, median (IQR), mm	20.4 (14.0–27.4)	23.2 (16.2–29.5)	19.9 (12.8–27.6)	21.6 (13.9–27.3)	18.7 (14.9–25.0)	0.05
T1 slope, mean (SD), °	23.7 (7.5)	25.1 (8.8)	22.0 (6.7)	24.5 (7.4)	24.2 (7.3)	<0.001
C7 slope, mean (SD), °	21.2 (6.8)	22.7 (8.1)	20.0 (6.2)	21.7 (6.6)	21.2 (6.8)	0.005
CT-point
C5	10 (1.6)	0	5 (2.3)	2 (1.0)	3 (2.7)	<0.001
C6	57 (9.0)	4 (3.9)	34 (15.5)	8 (4.0)	11 (9.7)
C7	110 (17.3)	18 (17.5)	33 (15.0)	37 (18.6)	22 (19.5)
T1	426 (67.1)	64 (62.1)	141 (64.1)	147 (73.9)	74 (65.5)
T2	32 (5.0)	17 (16.5)	7 (3.2)	5 (2.5)	3 (2.7)
TK, median (IQR), °	34.2 (28.0–40.7)	31.6 (25.8–39.9)	30.3 (24.4–37.3)	36.7 (30.8–42.5)	38.2 (31.8–42.5)	<0.001
T-apex
T4	8 (1.3)	1 (1.0)	2 (0.9)	2 (1.0)	3 (2.7)	<0.001
T5	28 (4.4)	1 (1.0)	9 (4.1)	7 (3.5)	11 (9.7)
T6	138 (21.7)	12 (11.7)	53 (24.1)	41 (20.6)	32 (28.3)
T7	242 (38.1)	30 (29.1)	87 (39.5)	80 (40.2)	45 (39.8)
T8	122 (19.2)	20 (19.4)	43 (19.5)	42 (21.1)	17 (15.0)
T9	47 (7.4)	14 (13.6)	13 (5.9)	16 (8.0)	4 (3.5)
T10	28 (4.4)	10 (9.7)	7 (3.2)	10 (5.0)	1 (0.9)
T11	9 (1.4)	5 (4.9)	3 (1.4)	1 (0.5)	0
T12	13 (2.0)	10 (9.7)	3 (1.4)	0	0
TL-point†
T7	1 (0.2)	1 (1.0)	0	0	0	<0.001
T8	2 (0.3)	0	2 (0.9)	0	0
T9	4 (0.6)	0	0	0	4 (3.5)
T10	25 (3.9)	0	7 (3.2)	6 (3.0)	12 (10.6)
T11	115 (18.1)	6 (5.8)	37 (16.8)	35 (17.6)	37 (32.7)
T11/T12	1 (0.2)	0	1 (0.5)	0	0
T12	357 (56.2)	36 (35.0)	124 (56.4)	140 (70.4)	57 (50.4)
T12/L1	1 (0.2)	0	1 (0.5)	0	0
L1	31 (4.9)	9 (8.7)	17 (7.7)	3 (1.5)	2 (1.8)
L2	66 (10.4)	29 (28.2)	21 (9.5)	15 (7.5)	1 (0.9)
L3	31 (4.9)	21 (20.4)	10 (4.5)	0	0
L5	1 (0.2)	1 (1.0)	0	0	0
LL, median (IQR), °	51.0 (42.2–58.4)	37.8 (31.2–44.7)	45.9 (40.0–54.9)	54.9 (49.4–59.9)	61.0 (54.6–65.8)	<0.001
L-apex
T12	1 (0.2)	0	1 (0.5)	0	0	<0.001
L2	1 (0.2)	1 (1.0)	0	0	0
L3	84 (13.2)	0	8 (3.6)	20 (10.1)	56 (49.6)
L4	427 (67.2)	35 (34.0)	170 (77.3)	166 (83.4)	56 (49.6)
L5	122 (19.2)	67 (65.0)	41 (18.6)	13 (6.5)	1 (0.9)
SS, median (IQR), °	34.8 (28.7–39.8)	24.4 (20.8–28.0)	31.8 (28.3–36.8)	37.5 (34.6–40.3)	42.5 (38.5–47.0)	<0.001
PI, median (IQR), °	49.1 (42.9–55.7)	36.8 (32.5–39.2)	45.0 (42.8–47.3)	53.4 (51.2–55.5)	62.3 (60.0–66.2)	<0.001
PT, median (IQR), °	15.1 (11.3–19.0)	11.6 (8.6–15.9)	13.4 (8.9–16.8)	15.9 (12.9–18.8)	20.4 (16.6–24.0)	<0.001
PS-length, mean (SD), mm	124.5 (10.2)	127.8 (9.4)	125.9 (9.8)	124.0 (9.8)	119.3 (10.5)	<0.001
Overhang, mean (SD), mm	34.1 (13.3)	29.2 (12.9)	30.1 (12.8)	36.0 (11.4)	42.7 (13.0)	<0.001
L5-length, median (IQR), mm	38.7 (36.4–41.3)	40.4 (37.3–43.9)	38.6 (36.1–41.2)	38.7 (36.6–40.8)	37.9 (35.8–40.7)	<0.001
Roussouly
Type 1	119 (18.7)	67 (65.0)	39 (17.7)	12 (6.0)	1 (0.9)	<0.001
Type 2	200 (31.5)	36 (35.0)	116 (52.7)	40 (20.1)	8 (7.1)
Type 3 anteverted	71 (11.2)	0	65 (29.5)	6 (3.0)	0
Type 3	188 (29.6)	0	0	133 (66.8)	55 (48.7)
Type 4	57 (9.0)	0	0	8 (4.0)	49 (43.4)
Roussouly (original)
Type 1	119 (18.7)	67 (65.0)	39 (17.7)	12 (6.0)	1 (0.9)	<0.001
Type 2	200 (31.5)	36 (35.0)	116 (52.7)	40 (20.1)	8 (7.1)
Type 3	259 (40.8)	0	65 (29.5)	139 (69.8)	55 (48.7)
Type 4	57 (9.0)	0	0	8 (4.0)	49 (43.4)

^*P-values measure whether clinical or radiologic characteristics differed by spinal morphologic subtypes.

^†T11/T12 and T12/T1 indicated that the inflection point located at the intervertebral level between two vertebrae.

BMI indicates body mass index (calculated as weight in kilograms divided by height in meters squared); C2–7 SVA, C2–7 sagittal vertical axis; C2–7, C2–7 angle; C-Apex, cervical apex; CT-Point, inflection point from cervical lordosis to thoracic kyphosis; L5-Length, length of the L5 vertebra; L-Apex, the apex of lumbar lordosis; LL, lumbar lordosis; O–C2, occipito–C2 angle; overhang, overhang distance; PI, pelvic incidence; PS-Length, pelvis-sacrum length; PT, pelvic tilt; SS, sacral slope; SSA, spinosacral angle; SVA, sagittal vertical axis; T-Apex, the apex of thoracic kyphosis; TK, thoracic kyphosis; TL-Point, inflection point from thoracic kyphosis to lumbar lordosis; TPA, T1 pelvic angle.

Morphologic Classification Through ML

K-means clustering identified four distinct morphologic subtypes (Fig. S2, Supplemental Digital Content 2, http://links.lww.com/BRS/C754). A comprehensive comparison of these subtypes was presented in Table 1. Among the various radiologic and clinical parameters assessed, PI exhibited the most substantial intersubtype variance (Fig. S3, Supplemental Digital Content 2, http://links.lww.com/BRS/C754). A progressive increase in PI was observed from subtype 1 to subtype 4, a trend that was also observed for other pivotal pelvic parameters, including SS, pelvic tilt (PT), and overhang, suggesting a widening and more oblique pelvic configuration. In addition, lumbar lordosis (LL) demonstrated a similar pattern of variation to that of the pelvic parameters. Conversely, the apex of lumbar lordosis (L-Apex) and the thoracic kyphosis-lumbar lordosis inflection point (TL-Point) positions migrated cranially from subtype 1 to subtype 4. The cervical parameters exhibited minimal variability across subtypes. For global parameters, both the T1 pelvic angle (TPA) and the spinosacral angle (SSA) exhibited a gradual increase from subtype 1 to subtype 4, whereas the sagittal vertical axis (SVA) remained relatively stable. No significant differences in age were observed among the subtypes.

Morphologic Subtype Classifiers

The PI model delineated optimal cutoff thresholds of 39.56°, 49.16°, and 58.31° to demarcate the boundaries between subtype 1 and 2, subtype 2 and 3, and subtype 3 and 4, respectively. After excluding the high-correlated variables, ML models were trained and the overall performance of RF model on the test set outscored others, as evidenced by its AUROC (Fig. S4, Supplemental Digital Content 2, http://links.lww.com/BRS/C754) and AUPRC (Fig. S5, Supplemental Digital Content 2, http://links.lww.com/BRS/C754). Finally, we used the four classifiers to generate predicted classes on the test cohort. The performance metrics demonstrated that PI model exhibited good discrimination for different subtypes (Table S3, Supplemental Digital Content 5, http://links.lww.com/BRS/C757), and were comparable to ML models (Fig. S6, Supplemental Digital Content 2, http://links.lww.com/BRS/C754).

Surgical Correction Reference

Twelve parameters were identified as key determinants for the new classification (Fig. S7, Supplemental Digital Content 2, http://links.lww.com/BRS/C754). Among these variables, LL, L-apex, TK, the apex of thoracic kyphosis (T-Apex), and TL-point were the surgically modifiable parameters in ASD realignment surgery.^1,8,10,19 To facilitate clinical application, LL was designated as the principal parameter for correction, as it was deemed more important than others according to the results of the feature selection algorithms (Fig. S7, Supplemental Digital Content 2, http://links.lww.com/BRS/C754). Therefore, regression algorithms were applied to predict the normative range of LL according to PI and specific subtypes for the surgical reference (Table S4, Supplemental Digital Content 6, http://links.lww.com/BRS/C758). To validate the clinical relevance of the surgical reference derived from the new correction algorithm, a retrospective ASD cohort consisting of 103 patients with sagittal deformity [mean (SD) age, 71.5 (7.2), 84 (81.6%) females, mean (SD) follow-up period, 15.7 (12.6) mo] was used (Table 2). The details of the new correction algorithm were shown in Figure 4.

TABLE 2.

Baseline Characteristics of ASD Cohort

Variables	Total (n=103)	Nonrestored (n=56)	Restored (n=47)	P
Age, mean (SD), y	71.53 (7.21)	71.84 (6.39)	71.17 (8.14)	0.641
Sex, N (%)
Female	84 (81.55)	49 (87.50)	35 (74.47)	0.089
Male	19 (18.45)	7 (12.50)	12 (25.53)
Currently smoker, N (%)	8 (7.77)	2 (3.57)	6 (12.77)	0.172
Number of levels fused, mean (SD)	7.12 (2.51)	7.21 (2.70)	7.00 (2.28)	0.668
Duration of symptoms, mean (SD), mo	105.47 (111.44)	130.54 (123.73)	75.60 (86.92)	0.01
Operating time, mean (SD), min	359.30 (91.04)	358.21 (89.48)	360.60 (93.82)	0.896
Estimated blood loss, mean (SD), mL	924.66 (555.15)	940.00 (600.44)	906.38 (501.69)	0.761
Intraoperative blood transfusion, mean (SD), mL	1045.99 (688.42)	1010.75 (659.69)	1087.98 (726.11)	0.573
Length of stay, mean (SD), day	20.34 (6.40)	20.84 (6.58)	19.74 (6.20)	0.39
Nonhome discharge, N (%)	10 (9.71)	6 (10.71)	4 (8.51)	0.966
Preoperative patient-reported outcomes, mean (SD)
VAS back pain	6.11 (1.34)	6.14 (1.29)	6.06 (1.41)	0.766
ODI (%)	51.67 (14.11)	51.87 (14.09)	51.44 (14.27)	0.88
SRS-pain	2.77 (0.60)	2.83 (0.55)	2.71 (0.66)	0.325
SRS-function/activity	2.85 (0.50)	2.86 (0.50)	2.84 (0.51)	0.827
SRS-self-image/appearance	2.88 (0.54)	2.81 (0.54)	2.96 (0.55)	0.171
SRS-mental health	3.09 (0.50)	3.10 (0.55)	3.09 (0.45)	0.944
SRS-subtotal	2.90 (0.30)	2.90 (0.31)	2.90 (0.30)	0.998
Preoperative radiographic assessment, mean (SD)
TK (°)	22.85 (14.35)	18.46 (14.82)	28.08 (11.95)	<0.001
LL (°)	17.74 (17.65)	11.89 (16.79)	24.72 (16.20)	<0.001
SS (°)	19.89 (10.65)	17.58 (10.11)	22.65 (10.74)	0.015
PT (°)	28.68 (10.55)	32.05 (10.98)	24.66 (8.49)	<0.001
PI (°)	48.10 (9.87)	49.21 (9.19)	46.77 (10.57)	0.214
PI-LL (°)	30.35 (17.03)	37.32 (17.61)	22.05 (11.96)	<0.001
TPA (°)	29.14 (11.77)	33.20 (12.42)	24.30 (8.87)	<0.001
SVA (mm)	91.00 (60.33)	103.05 (66.14)	76.65 (49.52)	0.026
Last follow-up radiographic assessment, mean (SD)
TK (°)	30.94 (12.17)	27.07 (11.57)	35.55 (11.34)	<0.001
LL (°)	31.59 (11.46)	24.88 (9.31)	39.59 (8.22)	<0.001
SS (°)	24.87 (8.13)	22.22 (7.88)	28.04 (7.31)	<0.001
PT (°)	24.46 (9.53)	28.47 (9.07)	19.69 (7.76)	<0.001
PI (°)	48.69 (9.84)	49.98 (9.32)	47.15 (10.31)	0.147
PI-LL (°)	17.10 (11.82)	25.10 (8.73)	7.57 (6.91)	<0.001
TPA (°)	22.49 (9.80)	27.61 (9.00)	16.40 (6.82)	<0.001
SVA (mm)	54.76 (50.23)	74.29 (50.88)	31.49 (38.50)	<0.001

ASD indicates adult spinal deformity; LL, lumbar lordosis; ODI, Oswestry disability index; PI, pelvic incidence; PI-LL, pelvic incidence minus lumbar lordosis; PT, pelvic tilt; SRS, Scoliosis Research Society Questionnaire; SS, sacral slope; SVA, sagittal vertical axis; TK, thoracic kyphosis; TPA, T1 pelvic angle; VAS, visual analog scale.

Figure 4.

Proposed algorithm for determining correction target in spinal realignment surgery. If the postoperative LL was within the target LL range, the patient was assigned to the restored group, otherwise to the nonrestored group. PI indicates pelvic incidence; LL, lumbar lordosis.

Consequently, 47 and 56 patients were grouped into the restored and nonrestored groups according to the new correction algorithm (Table 2). Age, sex, preoperative symptoms, PI, and operative variables were similar between groups. PJK (P=0.019), unplanned readmission (P=0.028), and unplanned reoperation (P=0.032) were significantly sporadic in the restored group defined by the new correction algorithm (Table 3). In addition, patient-reported outcomes including VAS back pain (P <0.001), ODI (%) (P=0.038), SRS-function/activity (P=0.04), and SRS-Satisfaction with management (P=0.04) of the restored group were superior to the nonrestored group during follow-up (Table 3).

TABLE 3.

Mechanical Complications and Clinical Outcomes of the ASD Cohort Divided by New Correction Algorithm

Variables	Total (n=103)	Nonrestored (n=56)	Restored (n=47)	RR (95% CI)*	P
Proximal junctional kyphosis, N (%)	13 (12.62)	11 (19.64)	2 (4.26)	0.22 (0.09–0.87)	0.019
Unplanned readmission, N (%)	18 (17.48)	14 (25.00)	4 (8.51)	0.34 (0.17–0.93)	0.028
Unplanned reoperation and causes, N (%)	12 (11.65)	10 (17.86)	2 (4.26)	0.24 (0.09–0.94)	0.032
Proximal junctional failure	2 (1.94)	2 (3.57)	0
Instrumentation failure	3 (2.91)	3 (5.36)	0
Hematoma	2 (1.94)	2 (3.57)	0
Wound infection	1 (0.97)	0	1 (2.13)
Pain relapse	4 (3.88)	3 (5.36)	1 (2.13)
VAS back pain (last follow-up), mean (SD)	2.66 (1.38)	3.09 (1.40)	2.15 (1.18)	NA	<0.001
ODI (%) (last follow-up), mean (SD)	32.04 (15.19)	34.88 (15.03)	28.66 (14.83)	NA	0.038
SRS-pain (last follow-up), mean (SD)	3.70 (0.57)	3.62 (0.59)	3.79 (0.55)	NA	0.146
SRS-function/activity (last follow-up), mean (SD)	3.69 (0.48)	3.61 (0.48)	3.80 (0.47)	NA	0.041
SRS-self image/appearance (last follow-up), mean (SD)	3.84 (0.46)	3.81 (0.48)	3.87 (0.43)	NA	0.515
SRS-mental health (last follow-up), mean (SD)	3.79 (0.52)	3.77 (0.58)	3.81 (0.45)	NA	0.696
SRS-subtotal (last follow-up), mean (SD)	3.75 (0.32)	3.70 (0.30)	3.82 (0.33)	NA	0.068
SRS-satisfaction with management (last follow-up), mean (SD)	3.85 (0.44)	3.77 (0.42)	3.95 (0.46)	NA	0.04
SRS-total (last follow-up), mean (SD)	3.76 (0.32)	3.71 (0.30)	3.83 (0.33)	NA	0.052

^*Restored group versus nonrestored group.

ASD indicates adult spinal deformity; ODI, Oswestry disability index; RR, relative risk; SRS, Scoliosis Research Society Questionnaire; VAS, visual analog scale.

Comparative Analysis With Established Classification Systems

The spinal curvatures of each subtype were depicted utilizing the determinant parameters (Fig. 5), revealing a pattern that aligned with the original Roussouly classification.⁹ The Sankey plot provided further evidence of this correspondence. However, it was observed that the new subtyping strategy, which incorporated a global assessment of spinal alignment, may result in certain discrepancies with the Roussouly classification. For instance, certain individuals classified as Roussouly type 1 with a small PI were reassigned to subtype 3 or subtype 4 according to the present classification scheme. Moreover, the intersubtype distances calculated using the t-SNE algorithm demonstrated that the new subtyping strategy had a higher discriminatory power compared with the Roussouly classification (Fig. S8, Supplemental Digital Content 2, http://links.lww.com/BRS/C754).

Figure 5.

Spinal curvatures of four morphologic subtypes and their association with the Roussouly classification. (A) Representation of spinal curvatures for each of the four morphologic subtypes, utilizing the mean or median values of respective determinant parameters. Subtype 1 is characterized by an atypical curve with an elongated thoracolumbar lordosis and a reduced lumbar lordosis. Subtype 2 exhibits a uniform, flat-backed appearance. Subtype 3 demonstrates a harmonious curve with comparable thoracic kyphosis and lumbar anterior convexity. Subtype 4 features a coordinated curve with an extended anterior lumbar lordosis and a shorter, more pronounced posterior thoracic lordosis. The observed spinal curvatures for each subtype align with the original Roussouly classification framework. (B) Sankey plot substantiating the correspondence between the four morphologic subtypes and the Roussouly classification.

DISCUSSION

In the present cross-sectional study of 635 asymptomatic Chinese adults, the k-means clustering algorithm identified four distinct subtypes of spinal alignment. Notably, the spinal parameters exhibited extensive variability across these subtypes, with PI emerging as the variable with the greatest intersubtype difference. Considering the degenerative changes of spinopelvic parameters in ASD patients, backward estimation of an ideal spine shape using SS or other parameters was challenging.^22–24 Given that PI remained constant throughout spinal degeneration, it served as a critical link between physiological and deformed spinal morphologies, thus warranting its role as an ideal biomarker for retrospective estimation and guidance of sagittal realignment surgery in ASD patients.^12,16,17 Consequently, we stratified the spinal subtypes based on PI thresholds, and proposed a PI-based proportional surgical correction reference for ASD patients with sagittal deformity.

In summary, our classification system is similar to the Roussouly classification in that both are based on asymptomatic populations^9,12 (Table 4). This allows ASD realignment surgeries to be guided to restore the physiological alignment of the spine,² which cannot be achieved using methods such as the GAP score¹ or the SRS-Schwab classification.¹⁰ However, our methodology is based on a larger data set of 635 asymptomatic volunteers than was used in Roussouly study (Table 4). Moreover, this data set covers the age range at which ASD occurs most frequently¹⁹ (Table 4), therefore offering better guidance for ASD surgery. Furthermore, our classification is grounded in ML algorithms, lending objectivity to our results, as opposed to the empirical summaries on which the Roussouly^9,12 and SRS-Schwab¹⁰ classifications are based (Table 4). Lastly, our classification system incorporates cervical, thoracic, lumbar, pelvic, and global spinal parameters, as well as the apex and inflection point of the spine (Table 4). It also has clear PI boundaries between subtypes (Table 4). Other classifications, such as the Roussouly classification, do not consider such a comprehensive range of spinal parameters.^9,12 Furthermore, they do not provide PI cutoff values for subtype identification,^9,12 which hinders their clinical application.

TABLE 4.

Comparisons Between Surgical Realignment Strategies for ASD

Traits	Lenke-Silva	SRS-Schwab*	Age-adjusted sagittal alignment goals	GAP score	Roussouly algorithm	New algorithm
Study population	NA	947 American ASD patients [mean (SD) age, 48 (18) yr]	773 American ASD patients [mean (SD) age, 53.7 (16.4)  yr]	222 European ASD patients [mean (SD) age, 52.2 (19.3) yr]	160 European asymptomatic volunteers (mean age 27 years, with a range from 18 to 48 yr)	635 Chinese asymptomatic volunteers [median (IQR) age, 65.7 (62.6–69.7) yr]
Spinal parameters included	Lumbar and global	Lumbar, pelvic, and global	Thoracic, lumbar, pelvic, and global	Lumbar, pelvic, and global	Thoracic, lumbar, and pelvic	Cervical, thoracic, lumbar, pelvic, and global
Whether included apex or inflection point	No	No	No	No	Yes	Yes
PI boundary for subtyping	NA	NA	NA	NA	Blurred†	Clear
Individualized correction target based on PI	No	No	Yes	Yes	No	Yes
Primary outcomes of the study	Clinical outcome measures	Clinical outcome measures	Clinical outcome measures	Clinical outcome measures and mechanical complications	Mechanical complications	Clinical outcome measures and mechanical complications
Study conclusion	Empirical	Empirical	Evidence-based	Evidence-based	Empirical	Evidence-based

^*Only sagittal modifiers were included.

^†Blurred PI boundary between Roussouly type 1 and 2 and Roussouly type 3 and 4.

ASD indicates adult spinal deformity; GAP score, Global Alignment and Proportion score; PI, pelvic incidence; SRS, Scoliosis Research Society.

Identifying effective strategies to mitigate the high incidence of mechanical failures and improve clinical outcomes is of paramount importance in the context of the growing practices of ASD surgery worldwide.^25,26 In pursuit of this objective, numerous studies have endorsed the concept of adhering to both individualized and normative spinal physiology in the determination of surgical realignment targets.^6,7 As the only sagittal spinal correction algorithm that used populations with normal physiological status, the Roussouly classification was initially applied to surgical goal setting by Bari et al. ⁴ and Sebaaly et al ⁵. Beyond the Roussouly classification, various other systems have been developed, including the Lenke-Silva treatment levels,¹³ SRS-Schwab system,^10,27 age-adjusted sagittal alignment goals^8,11 and the PI-based proportional GAP scoring system,¹ each with a specific focus on either preventing mechanical complications or improving postoperative health-related quality of life (HRQoL) measures. However, subsequent validations of these algorithms have yielded conflicting results,^{3–7,28–37} and none have demonstrated a consistent correlation with both improved clinical outcomes and reduced rates of mechanical complications, with the exception of the GAP score.^1,13,38–40 In contrast, the PI proportional correction algorithm proposed based on our new classification has addressed these shortcomings and provides better guidance for the surgical correction of ASD patients (Table 4). On the basis of this algorithm, the spine surgeons can compute a personalized correction target for each ASD patient to aid their surgical decision-making process in clinical practice. This is achieved by measuring the PI magnitude of a specific patient to confirm his/her sagittal spinal morphologic subtype, and then using the regression algorithm to derive the correction target (Fig. 4). This approach is notably more user-friendly when compared with other surgical realignment strategies (Table 4). Moreover, the marked differences observed between the restored and nonrestored surgical correction groups in the validation study suggested that our ASD correction algorithm yielded plausible results regarding both clinical outcomes and the incidence of mechanical complications. Furthermore, LL, the correction target selected by ML algorithms, is a pivotal component of the sagittal plane essential for maintaining an upright posture.⁴¹ Surgical correction of LL has been associated with the restoration of global sagittal alignment,⁴² thereby validating the rationale underpinning our surgical correction algorithm.

Moreover, our findings have elucidated that ML represented an efficacious methodology for identifying spine morphology variations through the analysis of high-dimensional data sets. Compared with conventional subgrouping techniques, ML does not necessitate empirical human input, thereby augmenting its rationale and precision.⁴³ To our knowledge, this constitutes the inaugural study to employ ML techniques for sagittal spinal categorization, culminating in the proposal of a novel surgical reference tailored for Chinese ASD patients.

There were a few limitations to this study. First, our analysis only included the canonical parameters in spinal studies. Second, for convenience and simplicity, PI was the only input parameter in the regression algorithm. The inclusion of demographic variables, such as age and gender, would improve the individualization of correction targets. Third, the validation of our surgical references was conducted within a single-center, retrospective ASD cohort, which restricted the generalizability of the findings. Consequently, future research necessitates validation within a large-scale, prospective, multicenter cohort to ensure broader applicability.

CONCLUSION

In the present study of 635 asymptomatic Chinese adults aged 45 years or above, we developed and validated a surgical correction reference for ASD patients based on four different sagittal morphologic subtypes of the spine recognized by unsupervised ML techniques. Among the 21 measured sagittal spinal parameters, LL was selected as the correction target because it is a crucial spinal morphologic determinant that can be corrected by surgery, and PI was used for backward estimation of the ideal correction target due to its constancy throughout sagittal spinal degeneration. The knowledge gained from this study had the potential to better guide the spinal realignment surgery and to facilitate the development of highly individualized, tailor-made surgical interventions for Chinese ASD patients. Such advancements would contribute to the improvement of clinical outcomes and the reduction of mechanical complications associated with ASD surgery.

Key Points

• The unsupervised machine learning algorithm identified four sagittal spinal morphologic clusters in asymptomatic middle-aged and elderly adults.

• A user-friendly rule-based classifier consisting of a set of pelvic incidence cutoffs was proposed for sagittal morphologic subtype prediction.

• A pelvic incidence-based proportional surgical correction criteria was proposed for ASD patients with different sagittal morphologic subtypes.

• Reduced risk of mechanical complications and improved clinical outcomes after ASD realignment surgery were observed under the guidance of these new surgical correction criteria.

Supplementary Material

brs-50-1715-s001.docx ^{(29KB, docx)}

brs-50-1715-s002.docx ^{(4.7MB, docx)}

brs-50-1715-s003.docx ^{(20.4KB, docx)}

brs-50-1715-s004.docx ^{(24.8KB, docx)}

brs-50-1715-s005.docx ^{(18KB, docx)}

brs-50-1715-s006.docx ^{(16.9KB, docx)}

brs-50-1715-s007.docx ^{(32.1KB, docx)}

brs-50-1715-s008.docx ^{(89.1KB, docx)}