Clinical Validation of a Novel Musculoskeletal Modeling Framework to Predict Postoperative Sagittal Alignment

Study Design.

A retrospective radiographic and biomechanical analysis of 108 thoracolumbar fusion patients from two clinical centers.

Objective.

This study aimed to determine the validity of a computational framework for predicting postoperative patient posture based on preoperative imaging and surgical data in a large clinical sample.

Summary of Background Data.

Short-term and long-term studies on thoracolumbar fusion patients have discussed that a preoperative predictive model would benefit surgical planning and improve patient outcomes. Clinical studies have shown that postoperative alignment changes at the pelvis and intact spine levels may negatively affect postural balance and quality of life. However, it remains challenging to predict such changes preoperatively because of confounding surgical and patient factors.

Materials and Methods.

Patient-specific musculoskeletal models incorporated weight, height, body mass index, age, pathology-associated muscle strength, preoperative sagittal alignment, and surgical treatment details. The sagittal alignment parameters predicted by the simulations were compared with those observed radiographically at a minimum of three months after surgery.

Results.

Pearson correlation coefficients ranged from r=0.86 to 0.95, and mean errors ranged from 4.1° to 5.6°. The predictive accuracies for postoperative spinopelvic malalignment (pelvic incidence minus lumbar lordosis>10°) and sagittal imbalance parameters (TPA>14°, T9PA>7.4°, or LPA>7.2°) were between 81% and 94%. Patients treated with long fusion (greater than five segments) had relatively lower prediction errors for lumbar lordosis and spinopelvic mismatch than those in the local and short groups.

Conclusions.

The overall model performance with long constructs was superior to those of the local (one to two segments) and short (three to four segments) fusion cases. The clinical framework is a promising tool in development to enhance clinical judgment and to help design treatment strategies for predictable surgical outcomes.

Level of Evidence.

3

In the last decade, surgical advances for spinal fusion procedures have significantly improved patient outcomes.1,2 However, postoperative (post-op) complications such as adjacent segment disease, nonunion, proximal junctional kyphosis, and implant failure still occur. These mechanical complications may develop because of fusion instrumentation and reciprocal changes (RC), which may require costly revision procedures at greater risk.3 High incidence (37%) of postoperative complications necessitate predictable and improved patient outcomes.4

Postoperative changes in the pelvis and unfused spine levels can affect optimal correction and global balance and ultimately result in clinical failure.5 Foreseeing the extent of these complex interactions remains challenging,6 and clinicians have difficulty formulating treatment plans that will achieve alignment goals long term.7 Advanced methods incorporating patient and surgical factors beyond preoperative (pre-op) radiographic parameters could improve surgical outcomes.8,9

Early predictive studies on adult spinal deformity performed regressions to infer postoperative pelvic tilt (PT) and sagittal vertical axis based on preoperative radiographs and assessed the prediction accuracy.10–13 Langella et al 14 developed a geometry-based graphical model and simulated planned surgery. Lafage et al 3 combined preoperative and postoperative radiographs and adjusted the estimated postoperative PT in virtual alignment models. Recent studies have used machine learning techniques to estimate postoperative PT and thoracic kyphosis (TK) in long fusion cases.15–17 The limitations of regression and geometrical models include negligence of RC, lack of generalizability to different patient cohorts, and explanation of the underlying causes of postural changes regulated by patient and surgical attributes.8,9 Machine learning models share some of these inherent limitations. Furthermore, the need for large and high-quality data and repetitive training for specific applications limits their use clinically.18 Biomechanical models that account for patient and surgical characteristics can overcome these limitations. Recently, a preliminarily validated framework for predicting postoperative alignment using musculoskeletal modeling has been introduced.19 In this study, we evaluated the clinical validity and performance of this modeling framework in a large cohort of patients with different fusion lengths (FLs).

MATERIALS AND METHODS

Patient Population and Data Collection

This retrospective, multicenter, observational, institutional review board–approved study identified 126 adult patients treated for various clinical presentations such as kyphosis, lumbar scoliosis, stenosis, degenerative disc disease, and spondylolisthesis. The inclusion criteria were primary (n=104) and revision (n=22) thoracolumbar spinal fusion. Patients were enrolled based on the number of disc levels fused: local fusion (one to two levels), short fusion (three to four levels), and long fusion (greater than five levels), with 42 patients in each group. Exclusion criteria included coronal T4-T12 Cobb angle >20°, intervertebral disc replacement in the thoracolumbar spine, and more than three adjacent level fusions from a prior surgery. Patients displaying symptoms of pseudoarthrosis, vertebral fracture, neurological deficits, and instrumentation failure before the eligible last postoperative follow-up time point (LFU) were also excluded. The primary and revision cases had a minimum of three-month and six-month follow-up postoperatively, respectively.

The collected data included patient age, sex, body mass index, and indications for surgery. All patients had full-body or long-cassette anteroposterior and lateral standing radiographs preoperatively and at LFU. Radiographic parameters (Figure 1A) were measured using a custom-built graphical user interface in MATLAB 2020a, the MathWorks Inc.

Figure 1.

A, Illustration of the sagittal parameters measured in an example patient-specific musculoskeletal model. Thoracic kyphosis (TK), lumbar lordosis (LL), pelvic incidence (PI), pelvic tilt (PT), T1-pelvic angle (TPA), T9-pelvic angle (T9PA), and L1-pelvic angle (LPA) are depicted in standing position. B, The model-predicted and measured (post-op) radiographic parameters. Pearson correlation coefficients (r) and root-mean-square errors (RMSE) are shown. The solid black line is the best fit line for the data, and the dashed line is the perfect correlation line.

Biomechanical Modeling

The computational framework adopted in this study has been previously described.19 Briefly, full-body musculoskeletal models of 21 fusion patients were constructed and analyzed to predict postoperative sagittal postures from preoperative alignment and surgical treatment parameters. Patient-specific models were developed by scaling the generic full-body musculoskeletal model in the AnyBody Modeling System (v. 7.3) by patient body weight and height. In these models, the vertebra (and bones of the extremities) are represented as rigid segments. The intervertebral articulations are modeled as spherical joints with nonlinear passive stiffness contributions from the disc and spinal ligaments. The mass distributions between individual trunk segments and the location of their centers of mass were modified to reflect the effects of the patient body mass index. Muscle strengths, defined as the maximum voluntary contraction force of the muscles in the musculoskeletal model, were uniformly scaled based on patient body weight. The strength of trunk muscles (erector spinae, multifidus, psoas major, abdominal, and hip muscles) was also adjusted to represent age and spinal-pathology–related muscle atrophy. The muscle scaling was based on data derived from the literature,20–27 and the modeling approach was described previously.28 Patient-specific spinopelvic sagittal alignment was represented in the models based on measurements from the preoperative lateral radiographs. Surgical correction was simulated by modifying the preoperative alignment by achieved disc height and implant angle for fused levels, derived from LFU radiographs. The fused segments were idealized using rigid constraints with no motion allowed, but full force and moment transfer. Knee flexion, PT, and RC at unfused segments for lumbar lordosis (LL) and TK were altered within physiological ranges, exploring a variety of plausible postures, and their values were predicted using an optimization algorithm. Inspired by Dubousset cone-of-economy concept, the algorithm minimized the standing muscle expenditure calculated for each tested posture using inverse-dynamic analysis.29 The prediction routine used derivative-free global and local optimization methods from the NLopt library and launched simulations using the AnyPyTools library in a custom Python script. The optimization problem was solved in two steps. First, a global optimization was executed,30 and global minimums were further refined using local optimization.31 Muscle expenditure, defined as the sum of the simulation-predicted activation of all muscles in the patient model, served as the objective function to be minimized for finding the posture of minimal muscle effort. Thus, the predicted alignments were biomechanically the most ergonomic (i.e. most likely for patients to assume) in response to simulated treatments. Model validity was evaluated in that preliminary study by comparing the model-predicted and measured postoperative sagittal postures for 21 analyzed fusion patient cases, indicating good accuracy (90%) to predict postoperative sagittal imbalance.

This study introduced some changes in the modeling framework. Patients in this study were imaged with their arms raised or holding onto a bar in front of them. Upon visibility, glenohumeral and elbow flexion angles from standing radiographs were implemented.32 When elbow joint was not visible, elbow flexion was either set to 0° or 130° minus glenohumeral angle to best mimic the imaging instructions at the clinical centers. The models also enabled the fusion of segments above the T4 vertebra. The passive stiffnesses of the hip, knee, and ankle joints were simulated based on the computed joint configurations.33 In addition, the optimization routine excluded knee flexion because the knee joints were not visible on most standing 36-inch radiographs. Instead, the initial knee flexion was set to 5° and was further calculated based on the predicted postoperative posture using soft constraints.34 These changes extended the models’ capability to study various patient groups and improved accuracy and simulation run time.

Data and Statistical Analysis

One-way analysis of variance with post hoc Tukey HSD was used to compare the fusion groups’ patient age, body mass index, and preoperative radiographic parameters. The χ² test of independence was performed to compare male-to-female ratio and preoperative spinal pathology among fusion groups. Paired t tests were used to compare preoperative and postoperative parameters. Pearson correlations between the measured and predicted postoperative parameters and the corresponding root-mean-square errors (RMSE) were calculated to evaluate model validity. The primary analysis also included one-sided noninferiority t tests for TK, LL, TPA (T1-pelvic angle), and pelvic incidence minus lumbar lordosis (PI-LL), with limit differences of 9.5°, 9.5°, 6°, and 8.5°, respectively. These thresholds were inspired by previous studies that discussed minimal clinically important differences in spine surgery.35–39 In addition, Bland-Altman plots were used to evaluate the agreement between the measurements and predictions.

Secondary analyses distinguished the accuracy of preoperative planning in predicting postoperative spinopelvic malalignment (PI-LL>10°) and sagittal imbalance (TPA>14°, T9PA (T9-pelvic angle)>7.4°, and LPA (L1-pelvic angle)>7.2°) by using McNemar exact test. These thresholds correspond to mild patient disability.37,40,41 Moreover, we compared the model’s performance for the fusion groups using standard errors calculated as the absolute difference between the measured and predicted postoperative parameters. Statistical tests were performed using MATLAB 2020a, and the significance level was set at P<0.05.

RESULTS

Patient Population

Eleven patients had postoperative changes in LL, TK, and PT beyond 1.5 interquartile ranges in all cases, six patients had pelvic incidence changes over 8°, and one patient could not be simulated because of model errors. These patients were identified as outliers and excluded from the data analysis. Table 1 lists the patient demographics and preoperative parameters of the included 108 cases. After exclusions, the local, short, and long fusion groups included 35, 37, and 36 patients. The patient population had a mean age of 59.3±15.3 years, a mean body mass index of 29.9±6.1 kg/m², and was 64% female. The mean alignment parameters at pre-op were PI=54.4°±17.0°, PI-LL=16.3°±12.0°, PT=19.3°±10.6°, LL=48.9°±19.4°, TK=38.9°±17.3°, TPA=17.8°±11.9°. The fusion groups were mostly similar in terms of demographics and sagittal alignment, except for the TK (P<0.05). Tukey HSD test for multiple comparisons found that the mean TK was significantly higher in the long fusion group compared with the short group (P=0.041, 95% CI, −19.2, −0.3). In addition, χ² tests with Bonferroni correction showed that there were significant relationships between spinal pathology and FLs. The long fusion group included more kyphotic patients (P<3.36e-05), and the short group had more patients with spondylolisthesis (P<8.67e-0.5). The local group had more degenerative disc disease cases (P<2.11e-0.5) and more stenosis cases than the short group (P=0.004). The distributions of FLs and the corresponding LFU time points were as follows: local (n=35, $\overline{\mathit{FL}}$ =1.3, $\overline{\mathit{LFU}}=$ 9 mo), short (n=37, $\overline{\mathit{FL}}$ =3.4, $\overline{\mathit{LFU}}$ =13mo), and long (n=36, $\overline{\mathit{FL}}=$ 9.1, $\overline{\mathit{LFU}}=$ 19 mo). Lower instrumented levels ranged from the sacrum to the T9 vertebra, and upper instrumented levels ranged from L5 to T1.

TABLE 1.

Patient Demographics and Preoperative Radiographic Parameters and Diagnosis.

	All patients (N=108)	Local fusion (N=35)	Short fusion (N=37)	Long fusion (N=36)	P
	Mean±SD	Mean±SD	Mean±SD	Mean±SD
Age	59.3±15.3	58.9±15.2	59.4±14.7	59.5±16.3	0.982
BMI	29.9±6.1	29.6±4.9	30.3±6.9	29.8±6.5	0.884
PI [°]	54.4±17.0	51.9±15.2	59.3±15.1	51.9±19.5	0.098
PI-LL [°]	16.3±12.0	11.8±10.8	16.7±10.9	20.1±13.1	0.141
PT [°]	19.3±10.6	19.8±7.7	20.4±12.6	17.6±10.8	0.487
LL [°]	48.9±19.4	45.6±14.7	49.6±18.3	51.2±24.1	0.470
TK [°]	38.9±17.3	37.6±15.1	34.6±13.1	44.4±21.6	0.047
TPA [°]	17.8±11.9	17.4±10.6	19.7±12.3	16.3±12.6	0.455
M/F	0.57	0.52	0.61	0.57	0.952
Kyphosis	24	2	4	18	5.09e-06
Scoliosis	23	7	7	9	0.797
Stenosis	41	20	9	12	0.013
Degenerative disc disease	54	27	10	17	1.10e-04
Spondylolisthesis	41	10	24	7	1.28e-04

Mean values, standard deviations (SD), and patient counts for spinal pathologies are shown. BMI indicates body mass index; LL, lumbar lordosis; PI, pelvic incidence; PI-LL, pelvic incidence minus lumbar Lordosis; PT, pelvic tilt; TK, thoracic kyphosis; TPA, T1-pelvic angle, and m/f, male-to-female ratio.

See Figure 1 for an illustration of the sagittal parameters.

Validation of the Predictive Modeling Framework

The predicted radiographic parameters correlated well with postoperative measurements (Figure 1B). The strongest correlations were computed for T9PA (r=0.95, RMSE=3.7°), LL (r=0.94, RMSE=6.0°), and PI-LL (r=0.93, RMSE=6.3°). The lowest correlations were observed for the PT (r=0.86, RMSE=5.5°), TK (r=0.89, RMSE=7.1°), and TPA (r=0.91, RMSE=5.2°). A test of noninferiority concluded that the mean absolute differences between the predicted and postoperative parameters were smaller than the corresponding thresholds of clinical significance (P<0.01): TK (predicted 40.8°/postoperative 40.3°), LL (predicted 52.8°/postoperative 50.9°), TPA (predicted 14.7°/postoperative 16.8°), and PI-LL (predicted 11.9°/postoperative 12.7°) (Figure 2). The mean errors between the predicted and actual parameters for all patients were 5.6°, 4.3°, 4.5°, 4.1°, and 4.7° for TK, LL, PT, TPA, and PI-LL, respectively (Table 2). Furthermore, the Bland-Altman method42 was used to assess the prediction accuracy of sagittal alignment. The predictions of eight patient cases for PI-LL and TPA, four cases for PT, six cases for LL, and three cases for TK (18 distinct cases) were outside the limits of the agreement bounds (Figure 3). All outlier cases for PI-LL were also outliers for LL and TPA.

Figure 2.

Comparison of the preoperative, postoperative, and predicted spinopelvic parameters. Paired t tests were used for comparing preoperative and postoperative parameters, and one-sided noninferiority t tests were used for postoperative and predicted parameters. Limit differences tested were as follow, 8.5° (PI-LL), 8° (PT), 9.5° (LL and TK), and 6° (TPA). **Significance at P=0.01. Please note that asterisks symbols indicate the noninferiority of the model predictions to postoperative parameters, that is, mean differences are lower than limit differences that they are NOT significantly different.

TABLE 2.

Comparison of Fusion Groups for Prediction Errors

	TK [°]		LL [°]		PT [°]		TPA [°]		PI-LL [°]
	Mean	SD	Mean	SD	Mean	SD	Mean	SD	Mean	SD
Local	5.9	3.9	6.0	4.6	3.9	2.8	4.3	3.6	5.8	4.8
Short	6.1	4.1	4.9	3.7	4.9	3.2	4.2	3.3	5.0	4.0
Long	5.0	5.0	1.9	3.5	4.6	3.7	3.8	2.7	3.4	3.2
All	5.6	4.3	4.3	4.3	4.5	3.3	4.1	3.2	4.7	4.1

Mean and standard deviations (SD) of the absolute errors are reported.

Figure 3.

Bland-Altman plots for the radiographic parameters: difference (measured minus predicted parameter) versus the average of measured and predicted values are plotted for each patient. Bias (blue lines) and 95% lower and upper limits of agreement (black lines) are plotted. Patient cases outside the limits of agreement bounds are depicted with marker borders highlighted in red.

Comparison of Fusion Groups

Postoperative radiographic parameters did not significantly change from preoperative values in the fusion cohort as large RCs counteracted surgical correction for LL and TK (Figure 2). However, surgery increased LL and TK (P<0.01) and decreased PI-LL, TPA (P<0.01), and PT (P<0.05) in a subset of patients (N=26, mostly short and long fusion cases) with severe spinopelvic mismatch (preoperative PI-LL>20°). The differences between the postoperative and predicted parameters were lower than the acceptable error margins in each fusion group and the severe spinopelvic mismatch group, so they were not significantly different (P<0.01). The prediction accuracy differed across the fusion groups. Overall, the mean absolute errors were lower for the long constructs and similar between the local and short groups. The mean values ranged from 3.9° (PT) to 6.0° (LL) in local, from 4.2° (TPA) to 6.1° (TK) in short, and from 1.9° (LL) to 5.0° (TK) in the long constructs. The fusion groups had similar prediction errors for TK, PT, and TPA but the long group had significantly lower errors for LL (1.9°) and PI-LL (3.4°) than the local (6°, 5.8°) and short groups (4.9°, 5.0°) (P<0.05).

Classification of Postoperative Spinopelvic Malalignment and Sagittal Imbalance

Thirty-six of 108 patients (33%) had a postoperative PI-LL>10°. The prediction accuracy for malalignment was 0.91, with positive (PPV) and negative (NPV) predictive values of 0.91 and 0.91, respectively, sensitivity of 0.81, and specificity of 0.96 (Table 3). The performance of the model in predicting sagittal imbalance differed based on the chosen radiographic parameters. When TPA was used, 67 of 108 (62%) patients experienced a postoperative imbalance. The prediction accuracies for TPA, T9PA, and LPA were 0.81, 0.93, and 0.94, respectively. The performance metrics of PPV, NPV, sensitivity, and specificity substantially improved when using T9PA or LPA over TPA. Also, the exact McNemar test did not find significant differences between the postoperatively measured and predicted TPA. This observation was made when the fusion groups were analyzed separately (local, P=0.72; short, P=0.22; long, P=0.45).

TABLE 3.

Diagnostic Accuracy for Predicting Postoperative Spinopelvic Malignment and Sagittal Imbalance

	PI-LL	TPA	T9PA	LPA
Accuracy	0.91	0.81	0.93	0.94
PPV	0.91	0.90	0.93	0.98
NPV	0.91	0.70	0.92	0.90
Sensitivity	0.81	0.78	0.93	0.92
Specificity	0.96	0.85	0.92	0.98

These clinical values correspond to mild patient disability.37,40,41 PPV stands for positive predictive value and NPV for negative predictive value.

DISCUSSION

This study investigated the validity of a preclinical framework that leverages musculoskeletal modeling to predict postoperative sagittal alignment in thoracolumbar fusion patients. With correlation coefficients (r) from 0.86 to 0.95, predictions were highly correlated with measured radiographic parameters at the last follow-up. Mean errors ranged from 4.1° to 5.6° for TPA, LL, PT, PI-LL, and TK. We also found that the models identified postoperative malalignment (PI-LL>10°) and sagittal imbalance (TPA>14°, T9PA>7.4°, or LPA>7.2°), with predictive accuracy values ranging from 81% to 94%. Although there were large errors in some cases, full-body patient-specific models that simulate preoperative surgical plans can anticipate global alignment after spinal fusion.

The predictive errors in this study were comparable to those of previous studies that investigated long constructs (fusion from T10 or T11 to the sacrum) and reported errors for postoperative PT ranging from 2.5° to 6.5° and TK from 4.8° to 13.2°.3,15,16,43 We found the mean errors for these parameters to be 4.5° and 5.6°, respectively, for the long cases (Table 2). Although we obtained relatively high errors for PT, it is unclear how Lee et al 15 quantified these errors. If absolute values were not used, positive and negative errors might have resulted in lower errors. The error magnitudes indicate that overall model performance is better for long constructs than for short and local constructs. Most patients in the long group had the lower instrumented vertebra terminated at the sacrum with the lumbar spine fused altogether. Musculoskeletal models incorporated the alignment of the fused levels from the last follow-up images and assumed no RCs at these segments. This ultimately resulted in much lower prediction errors for LL and PI-LL in the long fusion cohort. The prediction of patients with greater spinal flexibility (fewer fused segments, as in the local and short groups) is more challenging because of the higher potential for different reciprocal mechanisms. Furthermore, this is the first study comparing prediction errors among different patient cohorts. Patient enrollment data indicates that fusion length is important when treating various spinal pathology (Table 1). Grouping patients based on FL and simulating the effect of spinal pathology on muscle strength and, thus, internal spinal loads, musculoskeletal modeling showed excellent capability for predicting postoperative sagittal alignment for all patient groups diagnosed with TK, lumbar scoliosis, stenosis, degenerative disc disease, and spondylolisthesis. Moreover, although relatively lower correction magnitudes for LL compared with the short and long group (unreported data), the local group had the largest range of RC for LL, the second largest range for TK after the long group, and the lowest alignment changes for PT. These data suggest that postoperative alignment changes in the locally treated spine can be significant and predicting them using musculoskeletal models has clinical value. Finally, an advantage of musculoskeletal models is that they are not limited to preselected patient groups with similar preoperative diagnosis and surgical treatment. Specifically, the extension of fusion to the sacrum or pelvis, completely immobilizing the lumbar spine, is not required. Various configurations of the upper and lower instrumented vertebra can be tested using what-if simulations.

The preclinical framework showed excellent predictive capability to characterize postoperative spinopelvic malalignment based on the PI-LL (Table 3). Notably, the probability of correctly identifying the desired spinopelvic alignment was 96%. Assuming a fixed PI, the required lordosis correction can be calculated preoperatively. The classification of postoperative sagittal imbalance had relatively lower but good prediction accuracy when the TPA parameter was used (81%), yet excellent accuracy when T9PA or LPA was used (93% and 94%, respectively). Full-body models represented patient-specific alignment only from the sacrum to the T4 vertebra because the visibility of the upper thoracic endplates was often obstructed by the shoulder and arm bone anatomy.32 The lack of representation of the exact preoperative alignment at the T3-T1 segments in the model, as well as possible annotation errors in identifying the T1 centroid location for measuring TPA, might explain the relatively lower prediction capability when using TPA. Protopsaltis et al 37 reported significant correlations between the preoperative and two-year follow-up for TPA, T9PA, and LPA. Thus, T9PA or LPA parameters should be used to predict global sagittal balance because they had superior sensitivity and specificity coefficients of prediction. When patients in this study were investigated, the current musculoskeletal model displayed moderate improvement over the original version: the T9PA parameter was predicted with accuracy, PPV, NPV, sensitivity, and specificity values as 0.93, 0.93, 0.92, 0.93, and 0.92, respectively, compared with 0.89, 0.88, 0.92, 0.93, and 0.87, for the original model.19

The cone-of-economy suggests that ideal alignment allows for a balanced standing posture that requires minimal muscle effort.29 Utilizing this concept, we predicted postoperative sagittal alignment corresponding to minimum muscle activation and, thus, biomechanically, the most ergonomic. The optimization of whole-body muscle expenditure, incorporating the impact of patient factors and surgical plans on muscular strength, enabled the prediction of alignment changes. However, the minimum muscle effort was not always associated with the most accurate model predictions. Some patient postures were predicted more accurately when the computed muscle effort was slightly higher than the global minimum. Different configurations of reciprocal mechanisms may result in similar activation levels; however, the prediction accuracy of individual radiographic parameters can differ. Nevertheless, these observations do not disprove that minimal muscle effort cannot be used to predict RC. Some patients might have displayed a more upright posture than in daily life, temporarily compensating during imaging sessions and using increased muscle effort.44

The computational framework presented in this study utilizes three software packages to predict patient alignment after spinal fusion. Radiographic parameters were semi-manually measured in MATLAB and were predicted in Python by running fully automated patient-specific simulations in AnyBody Modeling System. Custom scripts were developed in proprietary software, thus they are not accessible by third-party organizations. Basic knowledge of spinal anatomy and experience with identifying anatomical landmarks from radiographs is required to utilize the presented clinical framework. Our vision is to combine these applications in one platform and leverage the clinical utility of predictive modeling. The preclinical framework can be used then to evaluate various treatment plans suggested by the surgeon and calculate the optimal alignment strategy by simulating the cage height and angle and distributing the surgical correction across targeted disc levels for osteotomy. Although the current framework relies on providing expected correction angle and distraction height, this could be enhanced in the future by predicting these parameters from planned implant and procedure types (e.g. osteotomies), for example, using finite element modeling.45

This study had some limitations. First, some patients had prior fusion at the cervical spine or joint replacement at the shoulder, hip, or knee. These comorbidities might have affected the global alignment; however, we did not assess their impact. Second, although the mean last follow-up time was 14 months after surgery, three and six months follow-up radiographs were included in the primary and revision cases, respectively. Our unpublished data from fusion patients showed that TK, LL, PT, and PI did not change from three months to two years postoperatively. One study found that PI-LL and PT did not change beyond six week postoperative up to two years.46 Another study reported that radiographic TK, LL, and sagittal balance parameters did not change from two years to three to five years.47 It is safe to assume that most alignment changes occur within three months after surgery. Third, biomechanical models do not simulate nonmechanical factors. However, it is crucial to acknowledge nonmechanical factors that limit realignment through intact levels and affect standing balance.1 Fourth, musculoskeletal models assumed symmetrical spinal curvatures in the frontal plane. Although the prediction of sagittal alignment is of utmost importance in patients with deformity either in the coronal or sagittal plane, this study focused on sagittal deformity only and excluded patients with large coronal Cobb angles. The prediction of the postoperative alignment proposed in this work is based on finding the posture of minimal muscle expenditure (i.e. concept of the cone of the economy), the estimation of which relies on the complex analysis of the body biomechanics using a detailed musculoskeletal model. Modeling the biomechanics of three-dimensional deformity would require substantial advancement of the current modeling approach, as well as collection and analysis of three-dimensional data (EOS, MRI), to consider not only three-dimensional orientations of the vertebrae (including axial rotation) but also asymmetry in muscle size and strength. Neglecting coronal deformity (and its impact on muscle activations) and representing only sagittal profile of an analyzed patient, seemed therefore justified only for small coronal Cobb angles. Nevertheless, our vision is to configure the modeling framework to also include coronal deformity in the future and expand its application to broader patient groups. Finally, the models did not include patient-specific hip and knee positions, because the knee joint was not visible in most patient images. Follow-up studies should incorporate these features with the hip joint as an additional compensatory mechanism.

The preclinical framework-simulated treatment details from retrospective patient cohorts by modeling the achieved disc height and angle at the last follow-up. Thus, predicted alignments are only the most ergonomic for the achieved treatments but might be suboptimal if the surgery is suboptimal. Future optimization functions should include the surgical plan as another design variable. For instance, minimizing muscle effort while maintaining LL proportional to PI and TPA below clinical thresholds for sagittal imbalance would allow for predicting the most ergonomic posture for the patient, in terms of optimal surgical alignment correction, while taking into account the body’s response to this treatment (RCs) that might hinder its effects. However, achieving a correction, as suggested by such a simulation framework, may be affected by other factors. Cage subsidence and migration might lessen the surgical goals of disc height restoration and segmental correction, and thus diminish clinical effectiveness.48,49 Finite element simulations incorporating cage design and materials, bone quality, internal loads, and effects of the surgical approach can help predict postoperative changes at the surgical site, disregarded in the current approach.50,51 Such preclinical musculoskeletal models informed by the estimated surgical settling of fusion equipment could predict optimal surgical plans based on the analysis of individual patient biomechanics.

Key Points

• Patient-specific musculoskeletal models simulating preoperative surgical plans and minimizing whole-body muscle expenditure can anticipate global alignment after spinal fusion.

• The preclinical framework showed the excellent predictive capability to characterize postoperative spinopelvic malalignment and sagittal imbalance based on PI-LL and T9PA radiographic parameters, respectively.

• The overall model performance with long constructs was superior to that of the local and short fusion cases.