Predictors of clinically meaningful bone mineral density gains with romosozumab: An explainable machine leaning analysis of a real-world cohort

Abstract

Objective

To evaluate the real-world effectiveness of Romosozumab in postmenopausal women with severe osteoporosis and to identify baseline clinical and biochemical predictors of clinically meaningful bone mineral density (BMD) gains (≥10 %, used for exploratory classification) using an explainable machine-learning approach.

Methods

We conducted a retrospective, observational multicentre study across seven hospitals in Castilla-La Mancha, Spain. Postmenopausal women aged ≥50 years who initiated romosozumab between May 2023 and November 2024 for severe osteoporosis or high fracture risk were included. Lumbar-spine, femoral-neck, and total-hip BMD were assessed by dual-energy X-ray absorptiometry (DXA) at baseline and 12 months. Baseline biochemical variables included serum P1NP, CTX, PTH, vitamin D, calcium, phosphate, alkaline phosphatase, and creatinine. Predictors of a ≥ 10 % BMD gain were examined using elastic-net logistic regression combined with SHapley Additive exPlanations (SHAP) for model interpretability.

Results

Fifty-eight women were analysed (mean ± SD age 71.7 ± 10.0 years; BMI 26.1 ± 4.8 kg/m²; mean age at menopause 47.3 ± 6.0 years). Mean 12-month BMD increases were + 15,35 % at the lumbar spine, +12,42 % at the femoral neck, and + 8,62 % at the total hip. The proportion achieving a ≥ 10 % gain was 39 %, 38.1 %, and 31.7 %, respectively. SHAP analysis identified consistent predictors of response: lower baseline BMD, higher phosphate levels, and younger age at menopause were associated with greater gains, whereas elevated PTH and alkaline phosphatase predicted a reduced response. Patients who had not received corticosteroids or NSAIDs in the six months prior to treatment initiation, typically for pain or inflammation, also showed greater increases in BMD.

Conclusions

Romosozumab was effective and well-tolerated in routine clinical practice, yielding meaningful and site-specific gains in BMD. Explainable machine-learning analysis identified physiologically coherent and consistent clinical predictors of ≥10 % response.