Interaction effects of significant risk factors on low bone mineral density in ankylosing spondylitis

Research subjects

A total of 249 patients from July 2012 to November 2018 recruited at the China-Japan Friendship Hospital were used as the training set. Another independently cohort including 140 patients from March 2019 to August 2022 were used as the validation set. This study employs continuous enrollment to mitigate selection bias. In this research, all included patients met the modified New York criteria for AS (Van der Linden, Valkenburg & Cats, 1984). Exclusion criteria: (1) the patient had undergone hip replacement surgery; (2) Suffering from other autoimmune diseases, including but not limited to inflammatory bowel disease, psoriasis, ichthyosis, etc; (3) The patient has serious basic diseases, such as severe malnutrition, liver and kidney failure, etc; (4) The patient did not agree to participate in this study. All patients included in this study were evaluated by doctors to see if they met the inclusion criteria and signed the informed consent. The study has been reviewed and approved by ethics committees, including the research ethics committee of China-Japan Friendship Hospital (approval No. 2017-67) and the ethics committee of cedar Sinai Medical Center (approval No. pro00048849), and was conducted in accordance with the declaration of Helsinki.

Data collection/measurements

The questionnaires were administered by investigators with experience in epidemiological research. Patients’ individual data, including gender, age, body mass index (BMI), current medication status, smoking, alcohol-related conditions, family history, onset age, and sports activities, were collected. The functional status, disease activity and severity in patients with AS were obtained by filling in the Bath Ankylosing Spondylitis Disease Activity Index (BASDAI) and Bath Ankylosing Spondylitis Functional Index (BASFI) questionnaires (Zochling, 2011).

All patients were evaluated by the same rheumatologist. The evaluation scope included physical examination, modified-Schober score, chest expansion score, and the Bath Ankylosing Spondylitis Metrology Index (BASMI) (Calin et al., 1994; Song et al., 2009; Jenkinson et al., 1994). The visual analog scale (VAS, 0–10cm) was used to assess the night pain and the patient global assessment (PGA) (Sieper et al., 2009).

The New York classification criteria was used to grade the degree of sacroiliac joint, and the classification criteria were normal (0) to most serious (4). The diagnosis of AS was unilateral grade 3, unilateral grade 4 or bilateral grade 2 (Van der Linden, Valkenburg & Cats, 1984). A lateral radiograph of the cervical, thoracic and lumbar spine were collected, and the modified ankylosing spondylitis score (mSASSS) was used to evaluate the AS-related changes. The scoring standard of mSASSS are as follow the previously published study (van der Heijde et al., 2019). The mSASSS was scored by a musculoskeletal radiologist and a cross-trained rheumatologist. Cohen’s kappa coefficient was used to analyze the consistency of the two doctors’ scores on the study subjects. When Cohen’s kappa (κ) coefficients were >0.85, it indicates that the consistency between researchers is good.

Hip involvement was evaluated by an experienced rheumatologist, including restricted range of motion, pain and Bath Ankylosing Spondylitis Radiology Hip Index (BASRI-hip). The BASRI-hip scoring method was refer to previous published study (Konsta et al., 2023). Radiographic hip joint involvement was defined by at least 1 score in the BASRI-hip scoring system (MacKay et al., 2000).

The disease activity score (ASDAS) for ankylosis was calculated using the formula in order to better assess the patient’s disease activity (Deodhar et al., 2022). The calculation of ASDAS-CRP refers to previous literature (Ørnbjerg et al., 2022). To evaluate the disease activity, the disease activity score (ASDAS) of ankylosis was calculated using the ASDAS-CRP formula, and the use of formula refers to published literature (Zochling, 2011).

At the time of patient enrollment, blood samples from AS patients were collected and analyzed using standard laboratory techniques, and BMD was measured. Before serum samples were collected, patients fasted overnight (at least 8 h). Indicators reflecting inflammation, including ESR and CRP, were collected. Other laboratory indicators closely related to AS, such as HLA-B27, were also recorded. BMD (g/cm2) was measured using dual-energy X-ray absorptiometry (DXA). The detection range of BMD includes: lumbar spine (L1-L4), femoral neck, and total hip. In this study, patients with a measured bone density T-scores < −1 at either site were defined as low BMD (Cabrera et al., 2018). In addition, the low BMD group was also divided into osteoporosis and osteopenia. Osteopenia and osteoporosis are defined according to the World Health Organization standards (Kanis, 1994).

Statistical analysis

Continuous variables are represented using the median, range, and/or mean, along with the standard deviation (SD) where appropriate. Categorical variables are presented as frequencies or percentages. To ensure data integrity, we endeavor to employ mean or median imputation whenever possible to address potential data missing issues. In cases of relatively symmetrical data distributions without conspicuous outliers, we typically opt for mean imputation. Conversely, when data distributions exhibit pronounced skewness or contain outliers, we are inclined to utilize median imputation. To assess differences between groups, appropriate statistical tests such as the chi-square test, t-test, and rank sum test were employed. The correlation coefficient was calculated using Spearman analysis. Logistic regression analysis was performed to identify significant risk factors associated with low BMD, both before and after adjusting for confounding factors. The effect sizes are reported as odds ratios (ORs) with corresponding 95% confidence intervals (CIs). We select variables based on the results of logistic regression and clinical relevance, ensuring that the inclusion of each feature is a reasonable explanation, and establish predictive model based on these variables. A nomogram for low BMD across different sites was developed, and its accuracy was evaluated using the concordance index. Akaike information criterion (AIC), Bayesian information criterion (BIC), and likelihood ratio (LR) test are used for model calibration, while net reclassification improvement (NRI), integrated discrimination improvement (IDI), and area under the receiver operating characteristic (AUROC) are employed for model discrimination. Clinical utility is assessed using decision curve analysis (DCA), and the model is further validated using validation set data. All statistical analyses were carried out using STATA software, special edition (version 14.0, STATA Corp, TX). The nomogram was constructed using the R language (version 3.5.2). By calculating the variance inflation factor (VIF), we determine whether there is collinearity among the factors. A VIF below 10 indicates no significant collinearity and can be used to create a nomogram (Cheng et al., 2022). The nomogram associates each variable with its corresponding score, and the sum of scores for all variables is defined as the total score. By drawing a vertical line from the axis of the total score, the estimated probability of occurrence can be obtained. Statistical significance was considered at a threshold of P < 0.05.

Patient characteristics

A total of 249 patients from the training set participated in this study, including 194 males and 55 females. The average age was 34 ± 11 years, the onset age was 24 ± 10 years, and the diagnostic duration was 6 ± 5 years. Among them, 132 patients (53.0%) had normal BMD, and 117 patients (47.0%) had low BMD. BMI was significantly different between the normal BMD group and the low BMD group (P < 0.05), while other baseline characteristics were similar (all P > 0.05). The characteristics of study patients in the training set are presented in Table 1.

The disease-related variables and BMD of the study patients are summarized in Table 2. BASMI, chest expansion, radiographic average grade of the sacroiliac joint, hip involvement, and ASDAS-CRP showed a significant difference between the normal BMD and the low BMD groups (all P < 0.05). The baseline characteristics and disease-related variables in validation set were shown in Table S1.

Prevalence of low BMD in different sites

Among all patients in the training set with low BMD, included 105 patients (89.7%) had osteopenia and 12 patients (10.3%) had osteoporosis. The prevalence of low BMD in different sites is shown in Table S2. The lumbar spine was the most common site for low BMD (29.3%), followed by the femoral neck (26.5%) and total hip (24.9%). For patients with hip involvement, the total hip was the most common site for low BMD (34.5%). For patients without hip involvement, the lumbar spine was the most common site for low BMD (16.7%).

The association of BMD and mSASSS in AS patients with syndesmophytes

For AS patients in training set with syndesmophytes, increase in mSASSS was significantly associated with higher anteroposterior lumbar spine BMD (r_s = 0.201, P = 0.024) but not with femoral neck or total hip BMD (r_s = −0.156, P = 0.081; r_s = −0.146, P = 0.102, respectively) (Table S3).

Identification of risk factors for low BMD in the femoral neck and total hip

We further studied the effect-size estimates of multiple examined factors in association with the risk of low BMD before and after adjusting for confounding factors for femoral neck and total hip BMD (Table 3). Based on univariate logistic regression analysis several factors associated with the development of low BMD in the femoral neck and total hip were found at a significance level of 5% (Table S4). After adjusting for age and gender, statistical significance was still existed in all factors. After multivariate adjustment, chest expansion, BASMI, mSASSS, BMI, average radiographic grades at the sacroiliac joint and hip involvement were recognized as risk factors for low BMD of the femoral neck (P < 0.05). Chest expansion, BASFI, mSASSS, ASDAS-CRP, diagnosis duration, BMI, night pain, average radiographic grades at the sacroiliac joint, PGA, and hip involvement were recognized as risk factors for low BMD of the total hip (P < 0.05).

Prediction accuracy assessment

Basic and full models were constructed to evaluate the predictive performance of important factors associated with low BMD (Table 4). The full model included all the variables investigated, however, the basic model included all variables except for the significant risk factors identified by regression analyses. Calibration and discriminant statistics were applied to evaluate the prediction performance of the femoral neck and total hip significance factors which were added in the basic model. The prediction accuracy of the full model was significantly higher than that of the basic model. As shown by the comprehensive discriminant improvement, there were significant differences between the two models in predicting the performance of low BMD in the femoral neck and total hip (P < 0.001). For both the femoral neck and total hip, decision curve analysis suggested that our full model had superiority over the basic model for the fact that more clinical net benefits were obtained in a rather wide range of threshold probabilities when using full models than those when using the basic model (Fig. 1).

We applied the model to the validation set comprising 140 patients. The results demonstrated that the full model achieved an AUROC of 0.79 in the femoral neck prediction model and an AUROC of 0.80 in the total hip prediction model. The predictive performance of the validation set in Table S5.

Interaction explorations

Since the occurrence of low BMD in AS patients is a complex process, the influence of any risk factor may be small when evaluated alone, but it may be more obvious when other risk factors are combined. To obtain more accurate information, combined with the outcomes of clinical and logistic regression analysis, we divided the variables that were relevant for the low BMD into groups and further explored the risk factors affecting low BMD in the femoral neck and total hip (Table 5).

Hip involvement, mSASSS, and the average radiographic grade of the sacroiliac joint were found to be significant risk factors associated with low BMD in the femoral neck and total hip. When the average radiological grade of the sacroiliac joint exceeds grade 3 or hip involvement, the presence of syndesmophytes (defined as at least one vertebral corner mSASSS score ≥ 2 (van der Heijde et al., 2019)) further increased the risk of low BMD in the femoral neck and total hip (both P < 0.05).

Notably, hip involvement not only interacted with the presence of syndesmophytes but also with the average radiological grade of the sacroiliac joint. When the average radiological grade of the sacroiliac joint exceeds grade 3, hip involvement increased the risk of low BMD in the femoral neck and total hip (both P < 0.05). This finding implies that for AS patients with hip involvement, the combination of severe sacroiliitis or syndesmophytes significantly increase the risk of low BMD at the femoral neck and total hip.

Prediction model

Finally, a nomogram was constructed to predict the risk of low BMD in AS patients based on significant factors that were identified in the femoral neck and total hip (Fig. 2). Nomogram’s important factors were analyzed by positive logistic regression at a significance level of 5%. For a specific patient, each indicator has specific values, mapped onto the “Points” scale to obtain individual scores per indicator. Summing all scores gives a total. Locate this total on the “Total Points” and map to the “Risk” scale to determine patient low BMD risk. For example, assuming a female (20 points) AS patient with an onset age of 20 (10 points), BMI of 18 (82 points), BASFI of 5 (64 points), chest expansion of 3 (30 points), sacroiliitis average of four scores (10 points), mSASSS scores of 18 (10 points) and hip involvement (50 points), the possibility of low BMD of total hip was estimated to be 70%.