Preoperative imaging predictors of subsidence of 3D-printed artificial cervical vertebral bodies: a cohort study of 102 patients with nomogram development and validation

Abstract

Background

Subsidence of an interbody implant, defined as its sinking into adjacent bone, is a common complication that can compromise spinal fusion outcomes and may require revision surgery. Custom 3D-printed titanium vertebral bodies are increasingly used for anterior spinal reconstruction. However, their reported subsidence incidence (15–30%) remains comparable to conventional cages. Early identification of high-risk patients is therefore crucial. This study aimed to identify preoperative imaging predictors of 3D-printed vertebral body subsidence and develop a nomogram for individualized risk prediction.

Methods

We retrospectively analyzed 102 patients (2020–2024) who underwent single-level anterior corpectomy and fusion with a custom 3D-printed titanium vertebral body. Preoperative CT and MRI metrics—including lower adjacent vertebral Hounsfield units (HU), endplate concavity depth, endplate angle, and Modic changes—were assessed as candidate risk factors. Subsidence was defined as implant sinking ≥ 3 mm at the prespecified 3-month postoperative follow-up. Independent predictors were identified by multivariate logistic regression, and a nomogram was constructed. Model performance was evaluated by area under the ROC curve (AUC), along with calibration and decision curve analysis, and internally validated via bootstrap resampling.

Results

Subsidence occurred in 32 of 102 patients (31.4%). Multivariate logistic regression identified lower adjacent vertebral HU, greater endplate concavity depth, and presence of Modic changes as independent predictors of subsidence. The predictive model showed excellent discrimination (AUC = 0.896, 95% CI 0.821–0.970) and remained robust after internal validation (bootstrap-corrected AUC ~ 0.886). It also demonstrated good calibration and a positive net clinical benefit on decision curve analysis.

Conclusions

The major risk factors for 3D-printed vertebral body subsidence (poor bone quality and weak endplate integrity) are similar to those for conventional cages. Our imaging-based nomogram provides accurate risk stratification to guide personalized surgical strategies—such as bone density optimization or use of larger implants in high-risk cases—while avoiding unnecessary interventions in low-risk patients.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12891-026-09676-2.

Introduction

Interbody implant subsidence refers to the sinking of a fusion cage or artificial vertebral implant into the adjacent vertebral endplates, which may compromise reconstruction stability and sagittal alignment and thereby affect surgical outcomes. With the increasing use of customized 3D-printed artificial vertebral bodies (3D-PAVBs) in spinal reconstruction, understanding and preventing subsidence has become a clinically important issue [1–3]. Although 3D-printed implants offer patient-specific geometry and porous architectures that may facilitate osseointegration, the reported incidence of subsidence remains substantial (approximately 15–30%) and can occasionally necessitate revision surgery in severe cases [4]. Despite this non-negligible risk, 3D-PAVBs continue to be used in anterior column reconstruction after corpectomy because the implant geometry (including footprint and endplate interface) can be tailored to the defect and endplate morphology to improve initial contact and anterior support. In addition, porous titanium architectures are designed to promote bone ingrowth and osseointegration and to reduce stress shielding by lowering the effective elastic modulus, which may help maintain postoperative height and alignment in selected settings [1–3]. Therefore, reliable preoperative risk stratification is essential to support surgical planning and preventive strategies.

Preoperative imaging provides accessible markers related to bone quality and endplate integrity. In studies of traditional interbody devices (mainly titanium mesh or PEEK implants), low bone quality, unfavorable endplate morphology, and Modic changes have been repeatedly associated with a higher risk of subsidence [5, 6]. CT-derived Hounsfield unit (HU) values are commonly used as convenient indicators of regional bone density, while endplate morphology (e.g., concavity depth and endplate angle) can reflect local load-bearing capacity and stress distribution [7–9]. In addition, MRI-detected Modic changes have been linked to compromised endplate–vertebral complex integrity and increased subsidence risk [10, 11].

However, whether these imaging-derived risk factors translate directly to 3D-printed, patient-specific implants remains incompletely defined. Emerging studies suggest that 3D-PAVB subsidence may share similar determinants, yet the evidence is still limited and largely focuses on single predictors rather than an integrated tool that provides an individualized probability of early subsidence in this specific implant setting [12,13]. Moreover, in current practice surgeons often rely on single-factor heuristics (such as HU alone), and the incremental value of combining multiple preoperative imaging markers has not been well quantified for 3D-PAVBs. Given the potential consequences of subsidence—including loss of alignment and implant-related failure—there is a need for a practical preoperative multivariable prediction model tailored to 3D-PAVBs that can support individualized risk estimation and guide perioperative decision-making.

Therefore, this study aimed to develop and internally validate an imaging-based prediction model for prespecified 3-month (early) subsidence following 3D-PAVB implantation using preoperative imaging parameters (vertebral HU, endplate concavity depth, endplate angle, and Modic changes) in a retrospective cohort of 102 patients. To clarify the incremental value beyond a single-factor heuristic, we also planned a head-to-head comparison between the full model and a reference model using HU alone. The final model was presented as a nomogram to enable bedside use, and its performance was assessed by discrimination, calibration, decision curve analysis, and bootstrap internal validation.

Methods

Study design and population

This single-center retrospective cohort study retrospectively screened all consecutive patients who underwent single-level anterior corpectomy and anterior column reconstruction at our institution between January 2020 and December 2024 (n = 310), regardless of implant type.

To focus on reconstruction using customized 3D-printed titanium-alloy artificial vertebral bodies (3D-PAVBs), patients who received non-3D-printed implants (e.g., titanium mesh cages or conventional cages) were excluded at the first screening step (n = 66). The remaining patients were assessed according to predefined eligibility criteria. Inclusion criteria were: (1) single-level vertebral body resection and reconstruction with a custom 3D-printed implant; (2) availability of complete preoperative spinal CT and MRI data; and (3) availability of postoperative lateral radiographs for the prespecified 3-month follow-up assessment. Exclusion criteria included multilevel surgeries, any infection or tumor at the operative level, a history of severe trauma at the target level, or incomplete imaging or missing postoperative lateral radiographs at the prespecified 3-month follow-up. Eligibility was determined according to predefined inclusion and exclusion criteria, and no outcome-based selection was performed. Clinical covariates known to influence fusion and implant stability (including systemic bone mineral density testing such as DXA/T-score, smoking history, body mass index, and osteoporosis medication use) were not systematically recorded in this retrospective dataset and showed substantial missingness; therefore, they were not included as candidate predictors in the present model. Finally, 102 patients met the criteria and were included in the analysis (Fig. 1).

Fig. 1.

Flowchart of patient screening and inclusion (310 consecutive single-level anterior corpectomy and reconstruction cases screened; 66 cases reconstructed with non-3D-printed implants excluded; 102 patients with custom 3D-PAVB and complete preoperative imaging and available postoperative lateral radiographs at the prespecified 3-month follow-up included in the final analysis)

Definition of subsidence

Subsidence was defined as implant sinking ≥ 3 mm measured on lateral cervical radiographs at the prespecified 3-month postoperative follow-up; therefore, the model predicts 3-month (early) subsidence. The ≥ 3 mm cutoff has been used to define clinically meaningful/severe subsidence after anterior cervical corpectomy reconstruction and has also been adopted in studies of 3D-printed artificial vertebral bodies [3, 5]. The 3-month time point was selected a priori to capture early postoperative settling, which is the focus of recent 3D-printed artificial vertebral body cohorts investigating early subsidence [12, 13]. The magnitude of subsidence was determined by comparing the intervertebral height immediately postoperatively versus that at last follow-up. Two senior spine surgeons independently performed all measurements; if their results differed by > 5%, a third observer adjudicated the final value (Fig. 2). Because this study was designed around a radiographic endpoint at 3 months, patient-reported outcomes, symptom severity, and revision surgery attributable to subsidence were not systematically collected and were therefore not analyzed.

Fig. 2.

Measurement methods for implant subsidence and endplate parameters. A Schematic illustration of implant subsidence measurement. The intervertebral height between adjacent endplates at the same level iscompared between the immediate postoperative image (A1) and the 3-month postoperative follow-up image (A2). A reduction of ≥ 3 mm is defined as subsidence. B Measurement of lower endplate concavity depth. On the mid-sagittal plane, a baseline is drawn between the anterior and posterior margins of the inferior endplate. A perpendicular line from the deepest concavity point to the baseline represents the concavity depth (mm).C Measurement of endplate angles. On the sagittal plane, the inferior endplate angle of the upper adjacent vertebra (a) and the superior endplate angle of the lower adjacent vertebra (b) are measured as the angle between the endplate tangent and the horizontal reference line. A consistent sign convention is applied to ensure measurement reproducibility. D Measurement of AP diameter and vertebral height.On the preoperative mid-sagittal CT plane, the AP diameter of the resected vertebral body is measured as the maximal anteroposterior distance between the anterior and posterior cortices. The vertebral height is measured on the same plane as the cranio-caudal distance between the midpoints of the superior and inferior endplates (mm). E Measurement of the C2–C7 Cobb angle.On neutral lateral cervical radiographs, lines are drawn parallel to the inferior endplates of C2 and C7; the angle formed by these lines (or by their perpendiculars) is recorded as the C2–C7 Cobb angle.

Implant characteristics and standardization

All reconstructions were performed using a custom 3D-printed titanium-alloy artificial vertebral body (3D-PAVB) designed from preoperative CT data. Across cases, the implant followed a consistent design concept comprising a porous architecture intended to facilitate osseointegration and an endplate-contact interface designed to enhance initial stability, whereas patient-specific customization was primarily applied to overall dimensions (e.g., height and footprint) to match the corpectomy defect and adjacent endplate morphology. Evidence from porous titanium spinal implants supports the biological rationale of bone ingrowth/osseointegration, although this does not preclude radiographic subsidence.

Quantitative device-level parameters (e.g., exact porosity, pore size, lattice type, effective elastic modulus, surface finishing, and the specific additive manufacturing modality) were not available in the clinical records and therefore were not analyzed as covariates in the present model.

Imaging measurements

All imaging parameters were measured on a PACS workstation by two observers blinded to outcomes, using bone-window CT reconstructions (slice thickness ≤ 1.5 mm). The average of the two measurements was used, and if the relative difference between observers exceeded 5%, a third physician repeated the measurement to resolve the discrepancy. Linear distances were recorded in millimeters (mm) and angles in degrees (°), with values recorded to two decimal places unless specified otherwise. All sagittal-plane measurements were performed on multiplanar reconstructed (MPR) CT images. The mid-sagittal slice was predefined as the sagittal MPR plane that bisected the vertebral body and spinal canal in the midline, confirmed by positioning the crosshair at the midpoint of the spinal canal on axial images and selecting the corresponding sagittal slice passing through the vertebral body midline. To ensure interobserver consistency, both observers performed measurements on the same predefined mid-sagittal slice for each patient (with the slice position documented/bookmarked on the PACS before measurement).

Bone quality and endplate morphology(The following parameters were measured to assess bone quality and endplate morphology):

Lower adjacent vertebral HU: On preoperative axial CT, a circular region of interest (ROI) was placed in the center of the cancellous bone of the lower adjacent vertebra, covering approximately 30–50% of the cross-sectional area and avoiding cortical bone, vascular channels, sclerotic endplate zones, and the endplate itself. The average Hounsfield unit (HU) value was calculated from at least 4 consecutive axial slices (Fig. 3).We measured HU values for both the cranial and caudal adjacent vertebral bodies. For model development, we selected the HU of the lower (caudal) adjacent vertebral body as the primary HU metric, because prior ACCF literature has shown that the inferior vertebral body HU is an independent predictor of early implant subsidence. To ensure reproducibility and model parsimony, we used a single prespecified HU definition in the final nomogram.

Fig. 3.

Measurement method for vertebral Hounsfield unit (HU) assessment. A. Preoperative sagittal cervical CT reconstruction showing the lower adjacent vertebral level selected for HU measurement. B–E. Axial CT images from different slices of the same vertebra demonstrating placement of a circular region of interest (ROI, red circle) centrally within the cancellous bone. The ROI covers approximately ≥ 50% of the trabecular area while avoiding cortical bone, subchondral sclerosis, and vascular channels. The average HU value obtained from at least four consecutive slices is recorded as the vertebral Hounsfield unit. CT, computed tomography; HU, Hounsfield unit.

Depth of concavity of the upper adjacent inferior endplate (mm): On the mid-sagittal slice of the upper adjacent vertebra, a line connecting the anterior and posterior edges of its inferior endplate was drawn as a baseline. The perpendicular distance from the deepest point of the endplate concavity to this baseline was measured as the concavity depth. If the deepest concavity point was not clearly visualized on the initial mid-sagittal slice, adjacent sagittal slices were briefly reviewed to identify the slice demonstrating the maximal concavity within the central region, and this slice was then fixed and used by both observers for the final measurement (Fig. 2B).

Endplate angles (°): On the same predefined mid-sagittal slice, two endplate angles were measured for adjacent levels. The inferior endplate angle of the upper adjacent vertebra was defined as the angle between the endplate line and the horizontal plane, and the superior endplate angle of the lower adjacent vertebra was measured similarly. To ensure consistency, both angles were recorded with a clockwise tilt defined as positive (Fig. 2C).

Modic changes: Modic changes on preoperative sagittal MRI were evaluated based on characteristic signal patterns (Type I, II, or III). For statistical analysis, Modic changes were dichotomized as present (Type I or II) versus absent.

Geometric dimensions and alignment(The following parameters related to vertebral dimensions and spinal alignment were measured):

AP diameter of resected vertebra (mm): On the predefined mid-sagittal MPR slice of the resected vertebra on preoperative CT, the shortest distance from the anterior to posterior cortical margin of the resected vertebral body was measured (Fig. 2D).

Height of resected vertebra (mm): On the same mid-sagittal slice, the shortest vertical distance between the superior and inferior endplates of the resected vertebral body was measured (Fig. 2D).

AP/Height ratio: The aspect ratio of the target vertebra was calculated by dividing the AP diameter by its height, reflecting the vertebral body’s shape and load-bearing characteristics.

Upper adjacent vertebral HU: The HU of the upper adjacent vertebra was measured using the same method as described for the lower adjacent vertebra, to assess differences in bone quality between the two adjacent levels.

C2–7 Cobb angle (°): On standing lateral cervical radiographs, the C2–7 Cobb angle was measured between the inferior endplate of C2 and the inferior endplate of C7 (with kyphosis defined as positive) (Fig. 2E).

Interobserver reliability

Interobserver agreement for continuous imaging measurements was assessed using ICCs based on a two-way random-effects model with absolute agreement (ICC(2,1)). ICC estimates with 95% confidence intervals are provided in Supplementary Table S1.

Statistical analysis

Continuous variables were expressed as mean ± standard deviation and were compared between subsidence and non-subsidence groups using the independent-samples t-test or the Mann–Whitney U test, as appropriate based on normality. Categorical variables were compared using the chi-square test or Fisher’s exact test. Candidate predictors were prespecified a priori based on clinical rationale and prior literature and included imaging-derived markers related to bone quality and endplate integrity (vertebral HU values, endplate concavity depth, endplate angles, and Modic changes). Univariate analyses were performed for descriptive comparisons only. Multivariable logistic regression was then fitted using these prespecified predictors, and adjusted odds ratios (ORs) with 95% confidence intervals (CIs) were reported for the final parsimonious model. Because key non-imaging clinical confounders (e.g., DXA/T-score, smoking status, BMI, osteoporosis treatment) were incomplete or unavailable, model development was restricted to prespecified imaging-derived predictors that were consistently available for all included patients.

Given the limited number of subsidence events (n = 32), we restricted model complexity to reduce overfitting and retained a small set of imaging predictors in the final model. The final model included three predictors, corresponding to approximately 10.7 events per predictor parameter (32/3), a commonly cited rule-of-thumb for logistic regression model stability. To contextualize the incremental value of the multifactorial model, we additionally fitted a reference model using lower vertebral HU alone and compared its performance with the full model in terms of discrimination (AUC with 95% CI) and clinical usefulness (decision curve analysis). Penalized regression (e.g., LASSO) was considered; however, because the number of candidate predictors was limited and prespecified and interpretability was prioritized, we used standard logistic regression and minimized overfitting through model parsimony, multicollinearity assessment (VIF), and bootstrap internal validation with optimism correction.

Multicollinearity among predictor variables was evaluated using the variance inflation factor (VIF), with VIF < 5 considered acceptable. Model discrimination was assessed by the area under the receiver operating characteristic curve (AUC). A probability cutoff based on the Youden index was calculated for descriptive reporting of sensitivity and specificity. Clinical decision thresholds, however, should be selected based on clinical context and decision-curve analysis (DCA), considering the relative harms/costs of preventive interventions and patient preferences. Model calibration was evaluated with a calibration curve and internally validated by 1000 bootstrap resamples; the optimism-corrected calibration intercept and optimism-corrected calibration slope, as well as the median absolute error (E50), were reported. Overall predictive accuracy was additionally quantified using the Brier score. Bootstrap resampling was used to estimate optimism and obtain optimism-corrected performance metrics; the optimism-corrected calibration slope was also used as an indicator of the need for coefficient shrinkage. Finally, decision curve analysis (DCA) and a clinical impact curve were used to evaluate the net clinical benefit of the prediction model across a range of threshold probabilities.

Results

Baseline patient characteristics and univariate analysis

A total of 102 patients met the inclusion criteria. All included patients completed the scheduled 3-month postoperative radiographic (X-ray) follow-up for subsidence assessment, of whom 32 (31.4%) experienced subsidence postoperatively. The mean age was 57.0 ± 11.6 years and 37.3% of patients were female. Baseline characteristics for the overall cohort and stratified by subsidence status are summarized in Table 1.

Table 1.

Baseline characteristics of the cohort and univariable comparisons by subsidence

Variable	Total (n = 102)	Subsidence(n = 32)	No subsidence(n = 70)	Statistic (t / χ² / OR)	P value
Age (years)	57.03 ± 11.55	57.44 ± 13.30	56.84 ± 10.75	0.22	0.825
Sex:
Female	38 (37.25%)	13 (40.62%)	25 (35.71%)	1.23	0.664
Male	64 (62.75%)	19 (59.38%)	45 (64.29%)
Surgical segment:
C4	22 (21.57%)	6 (18.75%)	16 (22.86%)	3.05	0.384
C5	42 (41.18%)	15 (46.88%)	27 (38.57%)
C6	37 (36.27%)	10 (31.25%)	27 (38.57%)
C7	1 (0.98%)	1 (3.12%)	0 (0.00%)
AP diameter of replaced vertebra (mm)	17.86 ± 2.34	17.75 ± 2.47	17.92 ± 2.30	-0.33	0.742
Height of replaced vertebra (mm)	11.67 ± 1.53	11.71 ± 1.24	11.66 ± 1.66	0.19	0.847
AP/Height ratio	1.56 ± 0.29	1.53 ± 0.25	1.57 ± 0.30	-0.68	0.499
Upper vertebral HU	374.67 ± 104.85	360.24 ± 110.21	381.26 ± 102.44	-0.91	0.365
Lower vertebral HU	307.71 ± 74.99	250.50 ± 56.31	333.86 ± 67.79	-6.49	< 0.001
Upper endplate angle (°)	14.30 ± 5.08	15.53 ± 5.01	13.73 ± 5.05	1.68	0.099
Lower endplate angle (°)	14.79 ± 5.80	17.27 ± 5.75	13.66 ± 5.50	2.98	0.004
Depth of lower endplate concavity (mm)	1.53 ± 0.84	2.03 ± 0.77	1.31 ± 0.78	4.37	< 0.001
C2–7 Cobb angle (°)	16.28 ± 5.28	15.79 ± 7.34	16.50 ± 4.06	-0.51	0.614
Modic type
└─ Type 1	21 (20.59%)	8 (25.00%)	13 (18.57%)	29.83	< 0.001
└─ Type 2	11 (10.78%)	11 (34.38%)	0 (0.00%)
└─ No change	70 (68.63%)	13 (40.62%)	57 (81.43%)
Modic change (0/1)
No	70 (68.63%)	13 (40.6%)	57 (81.4%)	16.98	< 0.001
Yes	32 (31.37%)	19 (59.4%)	13 (18.6%)

Abbreviations HU Hounsfield Unit, AP anteroposterior, OR odds ratio, SD standard deviation. Notes: All continuous variables use Welch’s t-test; categorical variables use χ² or Fisher’s exact test (2 × 2), with OR reported for Fisher’s exact tests. P values are formatted as “<0.001” when below 0.001

In univariate comparisons, the subsidence group had a significantly lower mean lower vertebral HU (250.5 ± 56.3 vs. 333.9 ± 67.8, p < 0.001) and a greater depth of lower endplate concavity (2.03 ± 0.77 vs. 1.31 ± 0.78 mm, p < 0.001) than the non-subsidence group. The mean lower endplate angle was also larger in the subsidence group (17.3° ± 5.8 vs. 13.7° ± 5.5, p = 0.004). Additionally, Modic type I/II endplate changes were present in a significantly higher proportion of patients with subsidence compared to those without subsidence (59.4% vs. 18.6%, p < 0.001). By contrast, there were no significant between-group differences in age, sex, surgical level, vertebral anteroposterior (AP) diameter, vertebral height, AP/height ratio, upper vertebral HU, upper endplate angle, or overall cervical alignment (C2–7 Cobb angle) (all p > 0.05) (Table 1).

Multivariate logistic regression analysis

All variables with p < 0.05 in univariate analysis were entered into a multivariate logistic regression model (Table 2). Lower endplate angle was significant in univariable analysis but did not remain significant after multivariable adjustment. This attenuation suggests that its univariable association was partly explained by overlap with other predictors; in our cohort, lower endplate angle showed a moderate correlation with lower adjacent vertebral HU (r = − 0.34), and collinearity diagnostics did not indicate problematic multicollinearity (all VIF < 2). Lower vertebral HU, depth of the lower endplate concavity, and Modic change emerged as independent predictors of subsidence in the final model, whereas lower endplate angle was not significant after adjustment (OR = 1.05, 95% CI 0.92–1.19, p = 0.53). Lower vertebral HU was a protective factor, with an OR of 0.979 per 1 HU increase (95% CI 0.968–0.991, p < 0.001), corresponding to approximately an 18–20% reduction in the odds of subsidence for every 10 HU increase. In contrast, each 1 mm increase in concavity depth was associated with a 2.61-fold higher odds of subsidence (95% CI 1.22–5.59, p = 0.015). The presence of Modic type I/II changes conferred an OR of 9.46 (95% CI 2.62–34.16, p = 0.001) compared to no Modic changes. No concerning multicollinearity was detected among the predictors (all VIF < 2).

Table 2.

Univariate and multivariable logistic regression results (Outcome: subsidence, Yes vs. No)

Variable	Univariate OR	Univariate 95% CI	Univariate P	Multivariable OR	Multivariable 95% CI	Multivariable P
Lower vertebral HU (per 1 HU)	0.978	0.968–0.988	< 0.001	0.979	0.968–0.991	< 0.001
Lower endplate angle (° per 1°)	1.126	1.037–1.223	0.005	1.075	0.962–1.202	0.202
Depth of lower endplate concavity (per 1 mm)	3.198	1.718–5.954	< 0.001	2.608	1.202–5.659	0.015
Modic change (Yes vs. No)	6.408	2.534–16.204	< 0.001	9.324	2.560–33.958	< 0.001

Odds ratios (OR) with 95% confidence intervals (CI) are reported; P values are shown as “<0.001” when applicable. Continuous predictors: Lower vertebral HU per 1 HU and Depth of lower endplate concavity per 1 mm. Categorical predictor: Modic change coded as Yes vs. No (reference). Lower endplate angle was not retained in the final multivariable model (its univariate result is shown only)

The final multivariable model includes three predictors; the corresponding coefficients (β, ln of multivariable OR) are:

βHU = − 0.021 (= ln 0.979)

βDepth = 0.959 (= ln 2.608)

βModic = yes = 2.233 (= ln 9.324)

Intercept (β₀) = 3.109; SE = 1.757; Wald z = 1.77; P = 0.077

Predicted probability:

logit(p) = β₀ + βHU·HU + βDepth·Depth + βModic·1(Modic = Yes)

p = 1/(1 + exp(− logit(p)))

Model performance

The three-factor logistic model demonstrated excellent discrimination for predicting subsidence. The area under the ROC curve (AUC) was 0.896 (95% CI 0.821–0.970) in the development cohort (Fig. 4A). Internal validation using 1000 bootstrap resamples yielded an optimism-corrected AUC of approximately 0.886, indicating minimal overfitting (Fig. 4E). For comparison, the reference model using lower vertebral HU alone achieved an AUC of 0.837; thus, the full three-predictor model improved discrimination (AUC 0.896 vs. 0.837; ΔAUC = 0.059). Concavity depth and Modic change alone had AUCs of 0.769 and 0.704, respectively.

Fig. 4.

Discrimination, calibration, and clinical utility of the final logistic-regression model (internal bootstrap validation, B = 1000). A ROC curve in the development cohort (AUC = 0.896, 95% CI 0.821–0.970). B Calibration curves (apparent and bias-corrected); bias-corrected E50 = 0.039 (0.025–0.137), ICI = 0.050 (0.028–0.135). C Decision-curve analysis (net benefit across threshold probabilities). D Clinical-impact curve based on 1000 hypothetical patients (red: number labeled high-risk; blue dashed: high-risk with events). E ROC: apparent vs bootstrap-corrected (B = 1000; shaded area = bootstrap 95% band; AUCapparent = 0.896; AUCbootstrap-corrected = 0.886). F Calibration: apparent vs bootstrap-corrected (B = 1000; bias-corrected E50 = 0.036 [0.018–0.074], ICI = 0.034 [0.027–0.090]).Abbreviations: AUC, area under the curve; E50, median absolute calibration error; ICI, integrated calibration index.

Using the maximum Youden index, a probability cutoff of 0.385 was identified to summarize classification performance (sensitivity/specificity) in this cohort. Importantly, DCA demonstrated net benefit across a broad range of threshold probabilities (~ 0.10–0.80), with net benefit peaking around a ~ 30% threshold (Fig. 4C), supporting the use of decision-analytic considerations rather than a single “optimal” cutoff. At this cutoff, the model achieved a sensitivity of 81.3%, specificity of 85.7%, and overall accuracy of 84.3% in the cohort. In practical terms, 26 of 32 patients who developed subsidence (81.3%) were correctly identified as high risk (true positives), while 60 of 70 patients without subsidence (85.7%) were correctly identified as low risk (true negatives) (Fig. 5).

Fig. 5.

Confusion matrix at probability threshold p = 0.385. “Pred 1”denotes predicted subsidence; “Pred 0” denotes predicted no subsidence. “Actual 1/0” are the observed outcomes. At p = 0.385 the counts were TP = 26, FP = 10, TN = 60, FN = 6. Performance: sensitivity = 0.813, specificity = 0.857, accuracy = 0.843, PPV = 0.722, NPV = 0.909, F1-score = 0.764. The threshold p was pre-specified for the development set (e.g., based on clinical considerations/Youden index). Abbreviations: TP, true positive; FP, false positive; TN, true negative; FN, false negative; PPV, positive predictive value; NPV, negative predictive value

The model also exhibited excellent calibration and clinical utility. The bootstrap-corrected calibration curve closely aligned with the ideal 45° reference line, with an optimism-corrected calibration intercept of ~ 0.05, an optimism-corrected calibration slope of ~ 0.98, and a median absolute error (E50) of ~ 0.04, indicating strong agreement between predicted probabilities and observed outcomes (Fig. 4B and F). The Brier score was 0.104 (bootstrap optimism-corrected Brier score: 0.117), indicating good overall predictive accuracy. Decision-curve analysis showed that the model provided a greater net benefit than either the “treat-none” or “treat-all” strategy across a wide range of threshold probabilities (~ 0.10–0.80), with the net benefit peaking around a 30% risk threshold (Fig. 4C). Consistently, the clinical impact curve demonstrated that at an ~ 30% threshold, the majority of patients classified as high-risk by the model indeed experienced subsidence events (Fig. 4D).

Nomogram for risk prediction

A nomogram was constructed based on the three independent predictors (lower vertebral HU, lower endplate concavity depth, and Modic change) to facilitate individualized risk estimation (Fig. 6). For example, a patient with a lower vertebral HU of 250 (≈ 70 points on the nomogram), a concavity depth of 2.0 mm (≈ 60 points), and Modic changes present (≈ 100 points) would have a total score of about 230 points, corresponding to an estimated ~ 60% probability of subsidence. In contrast, if the lower HU is 350, the concavity is shallow (< 1 mm), and no Modic changes are present, the total score would be under 60 points, translating to a predicted risk below 5%. In clinical practice, the choice of an actionable threshold should be individualized and reflect the burden and potential downsides of the intended preventive strategy. For low-burden measures (e.g., closer radiographic surveillance, patient counseling, and optimization of modifiable factors), clinicians may reasonably adopt a lower threshold, whereas for resource-intensive or higher-burden interventions (e.g., augmented fixation, substantial implant strategy changes, or intensified perioperative optimization), a higher threshold may be preferred. In our cohort, DCA suggested that thresholds around ~ 30% provided favorable net benefit (Fig. 4C–D); thus, we present ~ 30% as an illustrative reference point rather than a universal cutoff. Patients above this risk threshold could be considered for preoperative bone-strengthening treatments, selection of implants with larger endplate contact areas, or augmented fixation to mitigate subsidence risk. Conversely, patients below the threshold may avoid unnecessary additional interventions.

Fig. 6.

Nomogram for predicting postoperative subsidence. The final multivariable logistic-regression model includes three predictors: Lower vertebral HU (unit: HU), Depth of lower endplate concavity (unit: mm), and Modic change (No/Yes). Usage: locate the patient’s value on each predictor axis, project upward to the Points axis to obtain the score; sum across predictors to obtain Total Points and project downward to the Risk of subsidence axis to read the individualized predicted probability. The Linear Predictor axis (logit scale) is provided for cross-checking with the model equation. Discrimination, calibration and clinical utility are presented in Fig. 3; regression coefficients and intercept are reported in Table 2 . Notes: Lower HU, deeper lower-endplate concavity, and the presence of Modic change (Yes) correspond to higher total points and higher predicted risk. Abbreviations: HU, Hounsfield unit; logit, linear predictor

Discussion

Principal findings

In this cohort of 102 patients who underwent anterior reconstruction with custom 3D-printed titanium vertebral bodies, we identified several key preoperative predictors of early implant subsidence. Lower Hounsfield unit (HU) values of the adjacent vertebra (reflecting reduced local bone density) were associated with significantly higher subsidence risk, whereas higher HU (better bone quality) was protective. A greater endplate concavity depth was another independent risk factor, with each 1 mm increase in concavity roughly doubling the odds of subsidence. The presence of Modic type I/II changes at adjacent endplates, indicative of inflammatory or fatty degeneration, emerged as a particularly strong predictor (odds ratio ~ 9.5 for subsidence). Using these imaging factors, we constructed a nomogram-based prediction model that demonstrated excellent discrimination (area under the curve ~ 0.89) and good calibration in internal validation. Compared with a reference model using lower vertebral HU alone (AUC 0.837), the full three-predictor model achieved a higher AUC of 0.896 (ΔAUC = 0.059), indicating incremental discrimination beyond HU alone.

Although customized 3D-printed artificial vertebral bodies are intended to improve anatomic matching and potentially optimize load transfer, early subsidence after ACCF remains largely determined by host-related factors, particularly local bone quality and endplate structural integrity. Comparative clinical studies and meta-analytic evidence indicate that subsidence can still occur at a meaningful rate after 3D-printed vertebral body reconstruction and may not be uniformly lower than titanium mesh cages [14, 15]. Mechanistically, customization primarily optimizes implant geometry/fit but cannot fully compensate for reduced endplate strength in low-bone-quality vertebrae, concave endplate morphology, or degenerative/inflammatory endplate changes. Endplate morphology has been linked to subsidence risk [9], and perioperative factors such as excessive distraction/over-distraction may further increase interface stress and predispose to early settling [16]. These considerations may explain why our key predictors mirror conventional cage literature and highlight that personalization improves fit but does not eliminate early subsidence risk when endplate support is limited. Accordingly, the novelty of our work is not the identification of entirely new risk factors, but the 3D-PAVB–specific integration of routinely available preoperative imaging markers into a prespecified 3-month nomogram, together with internal validation and explicit benchmarking against HU alone.

We selected the prespecified 3-month time point as the endpoint for early subsidence because early mechanical settling after anterior cervical corpectomy reconstruction is commonly captured within the first postoperative months, and recent cohorts focusing on 3D-printed artificial vertebral bodies have similarly assessed “early subsidence” using radiographs obtained around 3 months [12, 13, 17]. In addition, the ≥3 mm cutoff used in the present study is consistent with prior corpectomy reconstruction literature and has also been adopted in studies comparing 3D-printed vertebral bodies with titanium mesh cages [3, 5]. Although the clinical impact of subsidence is heterogeneous across reports [18], severe subsidence has been associated with radiographic and clinical implications in some series, and selected reports have described neurological deterioration and revision in association with clinically significant subsidence [19]. Accordingly, our nomogram should be interpreted as predicting early radiographic (3-month) subsidence rather than directly predicting long-term clinical failure, and longer follow-up with external validation remains warranted.

Comparison with existing literature

Our findings align closely with those reported for traditional interbody cages in prior studies. In particular, we confirmed that the vertebral HU value is a critical determinant of subsidence risk, consistent with earlier investigations on PEEK or titanium cage [7]. Recognizing this, some authors have proposed spine segment–specific HU threshold values to improve risk prediction; for example, a “danger zone” threshold around the low 200s (HU) has been reported for cervical fusion cases [20]. In our cervical subgroup, the optimal HU cutoff (~ 276) was remarkably similar to that reported in a recent anterior cervical corpectomy study (~ 272) [12]. There is also evidence that bone density measured specifically at the endplate region correlates even more strongly with subsidence severity than whole-vertebra average density [21]. Moreover, outside the cervical spine, quantitative grading of preoperative CT attenuation has been explored as a means of stratifying subsidence risk in lumbar fusion, further underscoring the general importance of local bone quality [22].

Endplate structural integrity and preparation technique likewise play a significant role. We observed that deeper concavity of the lower endplate was associated with substantially increased subsidence risk (approximately twofold per millimeter), which is in agreement with previous findings that pronounced endplate depression predisposes to cage subsidence [9]. Biomechanical studies indicate that aggressive curettage or milling of endplate cartilage can weaken the subchondral bone plate and concentrate stress at the bone–implant interface; thus, preserving as much of the endplate’s subchondral bone as possible and maximizing the implant’s footprint contact area are recommended to reduce the risk of implant failure [23].

Degenerative changes in the endplates, as evidenced by Modic changes on MRI, also proved to be a critical factor. Patients with Modic type I/II changes had a markedly higher likelihood of subsidence in our study, consistent with a meta-analysis that found significantly greater cage subsidence rates in the presence of Modic changes [10]. This highlights the impact of underlying endplate pathology on implant stability. It also suggests that combining MRI-based indices of bone quality (such as the vertebral bone quality score) with CT-based measures like HU could further enhance subsidence risk prediction, as proposed by recent studies [24, 25].

Clinical implications

The nomogram developed in this study may serve as a practical tool for preoperative risk stratification in patients receiving 3D-printed vertebral body implants by providing an individualized probability of 3-month (early) subsidence. To translate the predicted probability into action, clinicians can select a risk threshold (threshold probability)—that is, the minimum predicted risk at which additional preventive measures would be considered, balancing the expected benefit of preventing subsidence against the potential harms, cost, and complexity of such measures. Decision curve analysis evaluates the net benefit of applying the model across a range of clinically reasonable thresholds compared with default strategies of intervening in all or none of the patients.

For illustration, we summarized observed 3-month subsidence rates across tertiles of nomogram-predicted risk (Additional file 3: Table S3A) and under two pragmatic thresholds (30% and 40%) (Additional file 4–5: Table S3B–S3C). A lower threshold (e.g., 30%) may be preferred when the planned measures are low-burden, whereas a higher threshold (e.g., 40%) may be selected when the measures increase operative complexity or cost. For patients exceeding the chosen threshold, potential strategies include preoperative bone health optimization (osteoporosis evaluation/treatment when indicated), meticulous endplate preparation emphasizing preservation of the subchondral endplate, maximizing the endplate-contact footprint (e.g., selecting/designing an implant with a larger contact area where feasible), and consideration of construct reinforcement in selected high-risk cases. Conversely, for patients predicted to be at low risk, unnecessary augmentative interventions may be avoided, potentially reducing operative burden without compromising outcomes. Representative de-identified vignette cases (including a low-risk example) are provided to illustrate bedside use of the nomogram (Additional file 2: Table S2). Building on these findings, several practical precautions may be considered for patients classified as high-risk by the nomogram. First, careful endplate preparation is essential: removal of cartilaginous endplate should be balanced with preservation of the subchondral bony endplate to maintain endplate strength and reduce early settling [14]. Second, surgeons should avoid over-distraction and oversizing during reconstruction, as excessive intraoperative distraction/height restoration has been associated with higher cage subsidence in anterior cervical fusion [26]. Third, implant selection and placement should prioritize adequate endplate-contact footprint and accurate centering/alignment to distribute load and mitigate stress concentration; designs that better conform to endplate anatomy have been proposed to improve anti-subsidence performance [16].

Implant material and biomechanical considerations

Despite the theoretical advantages of 3D-printed porous titanium implants—including a lattice structure conducive to bone ingrowth and an elastic modulus closer to that of bone—their tendency to subside is still fundamentally constrained by host bone quality and endplate integrity. Previous reviews have noted that while overall complication rates may be lower with 3D-printed implants compared to traditional cages, the clinical outcomes remain heavily dependent on the patient’s bone condition and endplate health [4]. Our results reinforce this point, as the identified risk factors (poor local bone density and weakened endplates) mirror those long recognized for conventional interbody devices. Notably, recent clinical series on 3D-printed cages have confirmed that low vertebral HU is a significant predictor of early subsidence [12, 13], just as earlier studies of titanium mesh cages found that a low preoperative vertebral CT attenuation is strongly associated with early cage subsidence after corpectomy [5, 27]. Simply substituting a porous implant for a standard cage does not overcome the need for adequate bone quality and endplate support in achieving stable fusion.

Beyond host factors, implant material and structural design may influence load transfer and settling behavior. Current evidence also indicates that 3D-printed titanium cages do not increase subsidence risk relative to polyetheretherketone (PEEK) cages and may in fact confer advantages in certain scenarios, although findings may vary by spinal region and implant system [28–31]. In addition, design details such as porosity/lattice architecture, effective modulus matching, footprint size, and endplate-interface geometry can be important: porous 3D-printed titanium cages have been reported to exhibit lower subsidence than traditional solid (non-porous) titanium cages in some fusion procedures, and hybrid material strategies have also been explored [32, 33]. Collectively, these data underscore the importance of optimizing implant material properties and structural design (including stiffness/modulus matching and footprint/interface contact), together with meticulous endplate preservation, to minimize subsidence risk.

Alternative reconstruction strategies should also be acknowledged. In anterior cervical fusion, structural autograft (e.g., iliac crest bone graft) remains a viable option in selected patients. Spallone et al. reported that a “mini-invasive” iliac crest harvesting technique can provide bicortical autografts with acceptable donor-site morbidity and may represent a practical alternative to cervical cage implantation [34]. In a retrospective analysis of ACDF, the same group found that mini-invasive iliac crest autografting yielded outcomes comparable to PEEK cages and was described as an inexpensive technique with reduced donor-site complications [35]. While indications and biomechanics differ between ACDF and corpectomy reconstruction, these reports highlight that reconstruction choices involve trade-offs among fusion biology, donor-site morbidity, implant-related risks, and costs. Importantly, the present nomogram is intended for risk prediction within patients receiving 3D-PAVBs after single-level corpectomy, rather than comparative effectiveness across reconstruction strategies.

Economic considerations should also be acknowledged. Although 3D printing enables patient-specific geometry and porous architectures, reviews have noted that the technology can be costly and typically requires specialized personnel and equipment, which may limit widespread adoption in many healthcare settings [34]. In cervical anterior corpectomy and fusion, a recent comparative study reported that a more costly 3D-printed off-the-shelf prosthesis did not reduce subsidence compared with a titanium mesh cage, highlighting that the value proposition of 3D-printed devices may vary by implant system and indication [36]. Therefore, subsidence-risk stratification should be interpreted together with local resource availability and economic considerations. Because our study did not capture implant costs or conduct a formal cost-effectiveness analysis, economic conclusions cannot be drawn from our cohort; future studies should integrate radiographic endpoints with standardized health-economic evaluation.

Study limitations

Several limitations should be acknowledged. First, this was a single-center retrospective study with a modest sample size and only 32 subsidence events, which may increase the risk of model optimism, coefficient instability, and restrict generalizability. Although we restricted the final model to three predictors (≈ 10.7 events per predictor parameter, a commonly cited rule-of-thumb threshold) and performed bootstrap internal validation to estimate optimism, the model should be further evaluated and refined in larger, multicenter cohorts. In addition, we reported optimism-corrected calibration metrics (intercept and slope) and the Brier score to provide a more complete assessment of potential optimism. Before clinical implementation, the model should undergo external validation, and coefficient recalibration (e.g., updating the intercept and/or calibration slope) may be required when applied to cohorts with different baseline risk or case-mix. Second, the outcome was assessed at a prespecified 3-month postoperative radiographic time point; therefore, the nomogram predicts early (3-month) subsidence and does not capture mid- to long-term subsidence progression or longer-term clinical outcomes. In addition, we did not systematically collect symptom-based endpoints or patient-reported outcomes (e.g., VAS/NDI/mJOA) and could not evaluate the relationship between early radiographic subsidence and short-term clinical status or revision. Third, although internal validation was performed, the model has not yet undergone external validation; its performance in other institutions, surgeons, and implant systems remains to be confirmed. Fourth, detailed implant-related parameters (e.g., porosity, lattice architecture, effective modulus, footprint/interface geometry, and surface finishing) and other perioperative technical factors were not quantitatively recorded and thus could not be modeled, even though they may influence load transfer and subsidence risk. Additionally, we did not collect implant cost or resource-utilization data and therefore could not evaluate the cost-effectiveness of 3D-PAVBs in relation to subsidence outcomes. In addition, we did not include a comparator cohort (e.g., titanium mesh/PEEK cages or other reconstruction strategies); therefore, the nomogram is intended for risk prediction within 3D-PAVB recipients rather than comparative effectiveness between implant types or surgical techniques. Finally, several established patient-level factors beyond imaging that may affect fusion biology and implant stability—such as systemic bone mineral density (DXA/T-score), smoking status, body mass index, and osteoporosis medication use—were not systematically captured in our retrospective records and therefore could not be adjusted for in the model. This may introduce residual confounding, and future prospective cohorts should collect these variables to enable comprehensive adjustment and external validation/recalibration. Future studies should prioritize multicenter external validation with longer follow-up and broader covariate capture, and should evaluate whether risk-stratified preventive strategies can reduce subsidence and improve outcomes.

Conclusion

This study aimed to identify preoperative imaging predictors of subsidence after 3D-printed vertebral body replacement and to construct a predictive nomogram. The main risk factors identified include low Hounsfield unit values of the inferior adjacent vertebra, a greater depth of concavity in the superior endplate, and the presence of Modic changes. These factors align with those reported for traditional interbody fusion cages, reaffirming that poor bone quality and compromised endplate integrity predispose to implant subsidence. The nomogram integrating these factors demonstrated excellent discrimination and calibration. Clinically, this tool can guide individualized preoperative planning: patients predicted at high risk of subsidence may benefit from bone quality optimization and implants with larger endplate contact area or augmented fixation to mitigate risk, whereas those at low predicted risk can avoid unnecessary interventions.

Abbreviations

3D-PAVB

Three-dimensional printed artificial vertebral body

ACCF

Anterior cervical corpectomy and fusion

AP

Anteroposterior

AUC

Area under the curve

CI

Confidence interval

CT

Computed tomography

DCA

Decision curve analysis

HU

Hounsfield unit

ICC

Intraclass correlation coefficient

MRI

Magnetic resonance imaging

OLIF

Oblique lateral interbody fusion

PEEK

Polyetheretherketone

ROC

Receiver operating characteristic

ROI

Region of interest

SD

Standard deviation

TLIF

Transforaminal lumbar interbody fusion

ULIF

Unilateral lumbar interbody fusion

VIF

Variance inflation factor

Authors’ contributions

Xin Xu and Ruoxian Song* designed the study, collected the clinical and imaging data, performed the statistical analyses, and were the primary contributors to drafting the manuscript. Kun Wang, Fuxin Wang, and Zheng Zhang assisted in data acquisition and imaging review. Ang Li and Junlin Han contributed to clinical data management, figure preparation, and manuscript organization. Ruoxian Song* supervised the entire project, provided critical revision of the manuscript, and approved the final version for submission. All authors read and approved the final manuscript.

Data availability

All data generated or analyzed during this study are included in this published article and its supplementary information files.This study did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors.

Declarations

Ethics approval and consent to participate

This retrospective study was approved by the Ethics Committee of the Affiliated 960th Hospital of PLA, Shandong Second Medical University (Approval No. 2024009). The requirement to obtain written informed consent was waived by the ethics committee due to the retrospective nature of the study. All procedures were performed in accordance with the ethical standards of our institution and with the 1964 Declaration of Helsinki and its later amendments.

Consent for publication

Not applicable. No individual patient details, images or personal data are included in this manuscript. All data have been anonymized, and the need for consent for publication was waived.

Competing interests

The authors declare no competing interests.