A plasma protein-based risk score to predict hip fractures

Abstract

As there are effective treatments to reduce hip fractures, identification of patients at high risk of hip fracture is important to inform efficient intervention strategies. To obtain a new tool for hip fracture prediction, we developed a protein-based risk score in the Cardiovascular Health Study using an aptamer-based proteomic platform. The proteomic risk score predicted incident hip fractures and improved hip fracture discrimination in two Trøndelag Health Study validation cohorts using the same aptamer-based platform. When transferred to an antibody-based proteomic platform in a UK Biobank validation cohort, the proteomic risk score was strongly associated with hip fractures (hazard ratio per s.d. increase, 1.64; 95% confidence interval 1.53-1.77). The proteomic risk score, but not available polygenic risk scores for fractures or bone mineral density, improved the C-index beyond the fracture risk assessment tool (FRAX), which integrates information from clinical risk factors (C-index, FRAX 0.735 versus FRAX + proteomic risk score 0.776). The developed proteomic risk score constitutes a new tool for stratifying patients according to hip fracture risk; however, its improvement in hip fracture discrimination is modest and its clinical utility beyond FRAX with information on femoral neck bone mineral density remains to be determined.

Fig. 1. Overall design of the present study.

A proteomic risk score was developed in the CHS (proteomics determined using the aptamer-based SomaScan 5K platform; 3,171 participants, 456 incident hip fractures, 39% men, mean age 74.4 years; Fig. 1 and Supplementary Table 1). The developed proteomic risk score was validated in two independent Trøndelag Health Study (HUNT) cohorts (proteomics determined using SomaScan 5K platform (3,259 participants, 187 incident hip fractures, 61% men, mean age 64.5 years) or 7K platform (1,988 participants, 155 incident hip fractures, 54% men, mean age 63.9 years)). In addition, the proteomic risk score was also validated in a subset of the UK Biobank (proteomics determined using the Olink double antibody proximity extension platform; all participants: 50,876 participants, 686 incident hip fractures, 46% men, mean age 57.0 years; randomly selected participants: 44,817 participants, 504 incident hip fractures, 46% men, mean age 56.7 years). Pink, cohorts (CHS and two HUNT cohorts) analyzed by aptamer-based SomaScan platform. Blue, cohort (UK Biobank) analyzed by the double antibody proximity extension Olink platform.

Fig. 2. Association between the proteomic risk score and incident hip fractures.

a, Association between the proteomic risk score and incident hip fractures in three separate validation cohorts. Base models are adjusted for age, sex and cohort-specific factors. Association with incident hip fractures is determined by Cox proportional regression models. Data are given as HRs and 95% CIs per s.d. higher risk score. The HUNT-SomaScan-5K cohort includes n = 3,259 participants and 187 incident hip fracture cases. The HUNT-SomaScan-7K cohort includes n = 1,988 participants and 155 incident hip fracture cases. The UK Biobank-Olink cohort includes n = 50,876 participants and 686 incident hip fracture cases. The results from the proteomic risk score were combined using fixed effects inverse-variance weighted meta-analysis with a total of n = 56,123 participants and 1,028 incident hip fracture cases. b, Associations between seven total population percentile bins of the proteomic risk score and risk of incident hip fractures in the UK Biobank. Association with incident hip fractures is determined by Cox proportional regression models adjusted for age, sex, proteomic batch, ethnicity and UK Biobank center (50,876 participants and 686 incident hip fracture cases). Data are given as HRs and 95% CIs with the 40–60% bin as reference group.

Extended Data Fig. 1. Kaplan–Meier curves of hip fractures according to proteomic risk score quartiles in UK the Biobank.

Data are given as point estimates of survival rate with 95% confidence intervals. (A) All participants with available proteomic analyses in the UK Biobank were included (50,876 participants and 686 incident hip fracture cases). (B) The randomly selected participants with available proteomic analyses in the UK Biobank were included (44,817 participants and 504 incident hip fracture cases).

Extended Data Fig. 2. Hip fracture survival probability according to proteomic risk score quartiles adjusted for age, sex, and cohort specific factors in the UK Biobank.

Data are given as point estimates of survival rate with 95% confidence intervals. (A) All participants with available proteomic analyses in the UK Biobank were included (50,876 participants and 686 incident hip fracture cases). (B) The randomly selected participants with available proteomic analyses in the UK Biobank were included (44,817 participants and 504 incident hip fracture cases).

Extended Data Fig. 3. Receiver operating characteristic curves and fracture discrimination (AUC from logistic regression models) using proteomic risk score or PRS-FN-BMD beyond FRAX-CRF in the UK Biobank.

Genetic analyses, proteomic analyses and eBMD were required. (A) All available participants (in total 49,087 participants and 663 incident hip fractures). (B) Randomly selected participants (in total 43,286 participants and 487 incident hip fractures). The base model was also adjusted for sex, proteomic batch, ethnicity, and UK Biobank centre. PRS-FN-BMD = Weighted polygenic risk score based on independent GWAS significant signals for femoral neck bone mineral density (FN-BMD) derived from Estrada et al.. FRAX-CRF = FRAX score for estimation of incident hip fracture risk using all available clinical risk factors in the UK Biobank. The sensitivity (=true positive rate) and specificity (=1-false positive rate) are presented on the Y-axis and X-axis, respectively. *AUC for the model including FRAX-CRF + proteomic risk score (blue line) was significantly larger than the AUC for the model including only FRAX-CRF (red line). Two-sided DeLong’s test gave (A) for all available participants P = 1.4 × 10⁻⁷ and (B) for the randomly selected participants P = 1.2 × 10⁻³  (ref. ).