Development and validation of a nomogram for predicting low bone mineral density in male patients with ankylosing spondylitis

Abstract

Objective: This retrospective cohort study aimed to develop and validate clinical nomogram models for predicting site-specific low bone mineral density (BMD) risk in male patients with ankylosing spondylitis (AS).

Methods: This study enrolled male AS patients treated at the Rheumatology Department of Jiangsu Provincial Hospital of Traditional Chinese Medicine between January 2017 and September 2024. A total of 322 eligible patients were randomly allocated to training and validation cohorts at a 7:3 ratio. Potential predictors of low BMD at the lumbar spine (LS) and left hip (LH) were initially screened through univariate logistic regression (p < 0.05), followed by stepwise bidirectional multivariate logistic regression (entry criteria p < 0.05) to identify independent predictors for each anatomical site. Based on the regression coefficients, we developed visualized nomogram prediction models for LS and LH low BMD, accompanied by an interactive online prediction tool. The models were comprehensively evaluated for discrimination, calibration, and clinical utility. After identifying the primary predictive factors, exploratory subgroup analyses were conducted to assess effect heterogeneity of key variables (BMI and serum uric acid).

Results: This study included 322 male AS patients randomly allocated to training (n = 225) and validation (n = 97) cohorts with balanced baseline characteristics (all p > 0.05). Multivariate logistic regression identified age at onset (LS OR = 0.96, 95%CI:0.93-0.99; LH OR = 0.97, 95%CI: 0.95-0.99), BMI (LS OR = 0.90, 95%CI: 0.81-0.99; LH OR = 0.81, 95%CI: 0.72-0.91), serum uric acid (LS/LH OR = 0.99, 95%CI: 0.99-0.99), and hip involvement (LS OR = 3.22, 95%CI: 1.71-6.05; LH OR = 8.03, 95%CI: 4.01-16.09) as common independent predictors for low BMD at both sites, while serum calcium (OR = 12.19, 95%CI: 1.44-103.25) was specific to LS. The developed nomograms, including web-based versions, demonstrated good discrimination (LS AUC: 0.77 training/0.73 validation; LH AUC: 0.82/0.85) and calibration. Decision curve analysis revealed significant net clinical benefit across probability thresholds (LS: 0.17-0.86 training/0.20-0.82 validation; LH: 0.15-0.92/0.27-0.91). The protective effect of BMI exhibited site-specific patterns: LS (low-TC: OR = 0.86; high-TC: OR = 0.77), LH (low-TC: OR = 0.77; mid-TC: OR = 0.74), with the most pronounced effect observed in the LS low-TG subgroup (OR = 0.79). SUA demonstrated consistent protective effects (LS/LH: OR = 0.95-0.99, all p < 0.05), potentially independent of disease stage. Interaction analyses revealed that neither lipid levels nor disease stage significantly modified the effects of BMI and SUA (all interaction p > 0.4).

Conclusion: This study developed clinical prediction models with excellent discriminative ability and substantial clinical utility for male patients with AS. These models offer rheumatologists an efficient tool to rapidly assess individual risks of low BMD, facilitating early diagnostic decision-making and enabling personalized interventions tailored to anatomical site-specific osteoporosis risks.

Keywords: ankylosing spondylitis; dynamic nomogram; early prevention; low bone mineral density; prediction model.

Figure 1

Subgroup analysis of BMI’s protective effects against low BMD in male AS patients: Integrated forest plot displays LS and LH data stratified by lipid profiles (TC/TG tertiles). Protective trends were observed in most subgroups (LS: significant in low/high-TC [OR = 0.86/0.77] and low-TG [OR = 0.79]; LH: significant in low/mid-TC [OR = 0.77/0.74] and all TG subgroups [OR = 0.70–0.83]), though nonsignificant in mid-TC for LS (p = 0.108). Nonsignificant interaction terms (all p > 0.4) suggest lipid-level-independent protective mechanisms of BMI.

Figure 2

Subgroup analysis of SUA effects on low BMD: Unified forest plot presents LS and LH data across AS disease stages (early-stage: LS OR = 0.99/LH OR = 0.95; advanced-stage: LS/LH OR = 0.99). Interaction p-values (0.8) suggest stage-independent protection.

Figure 3

Low BMD nomograms for (A) LS and (B) LH in male AS patients, with interactive red dots for variable input (e.g., age at onset, BMI) and real-time display of total points/predicted probability (%). Example: 20-year-old male with BMI 26.35 kg/m², serum calcium 2.89 mmol/L, SUA 457 μmol/L, and hip involvement (LS: 355 points → 88.9% risk; LH: 241 points → 74.5% risk).

Figure 4

Web-based nomogram for LS low BMD prediction in male AS patients, demonstrating real-time risk probability calculation through interactive input of clinical parameters (left panel) with automated graphical output display. Operational example: Users adjust sliders for variables including age at onset, BMI, and serum biomarkers to generate instant probability estimates visualized along the scoring continuum.

Figure 5

ROC analysis of low BMD prediction models for (A) LS and (B) LH, with training (blue) and validation (red) cohort performance relative to reference (dashed line, AUC = 0.5), displaying sensitivity-specificity relationships and AUC (95% CI) values.

Figure 6

Calibration curves for (A) LS training cohort, (B) LS validation cohort, (C) LH training cohort, and (D) LH validation cohort. The dashed diagonal line (ideal) represents perfect prediction, the solid black line (bias-corrected) shows the adjusted calibration, and the apparent predictions are depicted by the blue solid line (LS model) and red solid line (LH model). The convergence of these curves demonstrates calibration performance: proximity between apparent (blue/red) and ideal lines reflects prediction accuracy, whereas agreement between apparent and bias-corrected (black) lines indicates model stability.

Figure 7

Decision curve analysis for (A) LS training, (B) LS validation, (C) LH training, and (D) LH validation cohorts, showing the nomogram models’ net benefit (blue/red solid lines) versus reference strategies: “Treat All” (diagonal solid line, indicating treat-all approach with inherent over-treatment) and “Treat None” (horizontal solid line, representing no-intervention strategy). The models’ curves exceed both reference lines across most threshold probability ranges, demonstrating significant net benefit advantages in clinically relevant probability intervals.

Figure 8

Clinical impact curves for (A) lumbar spine training cohort, (B) lumbar spine validation cohort, (C) left hip training cohort, and (D) left hip validation cohort. Axes: x = High Risk Threshold; y = Number high risk (out of 1,000) (estimated individuals classified as high-risk per 1,000 patients). Curves: predicted events by lumbar spine (blue) and left hip (red) nomograms; observed events (black).