Development and external validation of competing risk models for major osteoporotic fracture and hip fracture in an ethnically diverse UK population of 15.6 million adults

Technical Summary: Development and External Validation of Competing Risk Models for Major Osteoporotic and Hip Fracture

Problem Statement
Fragility fractures represent a significant global cause of mortality, disability, and loss of independence, with incidence expected to rise as populations age. While fracture-risk prediction models (such as FRAX and QFracture) guide clinical decisions, existing tools incompletely capture growing populations living with learning disabilities, chronic diseases, HIV, and cancer survivorship. Furthermore, current tools often fail to explicitly account for competing mortality (the risk of death before a fracture occurs), which can lead to overestimation of 10-year fracture risks, particularly in older adults and those with multimorbidity. There is a need to determine whether explicitly incorporating these specific clinical characteristics and utilizing a competing-risk framework improves prediction accuracy and clinical utility beyond established predictors.

Methodology
The study utilized a large-scale, population-based design involving 15.6 million adults aged 30–100 years from the UK. The data were derived from linked primary care records (QResearch and CPRD), hospital admissions, and mortality records.

• Cohorts:
  • Derivation Cohort: 9,465,613 participants from QResearch (England).
  • Internal Validation: 3,155,655 held-out participants from QResearch.
  • External Validation: 3,011,531 participants from the Clinical Practice Research Datalink (CPRD) covering England and Wales.
• Model Development:
  • Model A: Re-estimated the established QFracture predictor architecture using cause-specific Cox models that explicitly account for competing mortality.
  • Model B: Built upon Model A by incorporating additional predictors: learning disability, Down syndrome, multiple sclerosis, visual impairment, HIV/AIDS, selected cancers (blood, cervical, colorectal, lung, prostate, pancreatic, liver), and cancer-directed endocrine therapies (aromatase inhibitors and androgen-deprivation therapy).
  • Model C: A variant of Model B excluding ethnicity to assess the impact of ethnicity omission on fairness and calibration.
• Statistical Approach:
  • Sex-specific models were developed for two outcomes: major osteoporotic fracture (MOF) and hip fracture.
  • Competing risks were modeled to estimate the cumulative incidence of fracture while accounting for the probability of death.
  • Performance was evaluated using discrimination (C-statistics), calibration (observed-to-expected ratios, calibration intercepts/slopes), Brier scores, decision-curve analysis (net benefit), and reclassification metrics.
  • Missing data were handled via multiple imputation, and non-linear/age-dependent associations were modeled using fractional polynomials and interaction terms.

Key Contributions

1. Inclusive Predictor Architecture: The study explicitly integrates risk factors often omitted from standard tools, including learning disabilities, Down syndrome, multiple sclerosis, visual impairment, HIV/AIDS, and specific cancer treatments, within a general-population model.
2. Competing Risk Framework: The models utilize cause-specific Cox models to distinguish improvements arising from contemporary estimation and competing-risk modeling from those contributed by additional predictors.
3. Ethnicity Analysis: The study provides empirical evidence on the effects of omitting ethnicity from prediction models, challenging the notion that removing protected characteristics automatically ensures fairness.
4. Scale and Validation: The analysis covers an unprecedented scale (15.6 million adults) with rigorous internal and external validation across different databases and time periods.

Results

• Discrimination: All models demonstrated high discrimination. C-statistics for Model B ranged from 0.766 to 0.900 across validation cohorts and outcomes. Discrimination was generally similar between Model A, Model B, and the original QFracture, though Model B showed slight numerical improvements in some comparisons.
• Calibration: Models A and B showed substantially better overall calibration than QFracture in both validation cohorts. QFracture consistently overpredicted risk (observed-to-expected ratios ranging from 0.499 to 0.758), whereas Models A and B showed ratios closer to unity (0.950–1.122).
• Clinical Utility: Decision-curve analysis indicated that Models A and B provided higher net benefit over parts of the prespecified threshold range compared to QFracture, particularly for women and hip fracture prediction.
• Reclassification: Model B reclassified individuals to higher risk categories at younger ages compared to QFracture. For instance, individuals with learning disability reached the illustrative 10% MOF risk threshold approximately 10 years earlier than those without. Those reclassified upward by Model B had higher observed 10-year risks (7.93% for MOF) compared to those reclassified downward (6.50%).
• Ethnicity Impact: Omitting ethnicity (Model C) resulted in similar discrimination but materially worse calibration, specifically leading to overprediction of risk in non-White ethnic groups.
• Predictor Effects: New predictors such as aromatase inhibitors (women) and androgen-deprivation therapy (men) showed strong associations with fracture risk (approx. 2-fold and 3.6-fold adjusted hazards for hip fracture, respectively).

Significance and Claims
The authors claim that the principal advancement of Model B is the simultaneous inclusion of diverse clinical populations (learning disability, HIV, cancer survivors) and the explicit accounting for competing mortality within a single population-based model. The study concludes that:

1. Contemporary Estimation: The largest improvement in population-level calibration and clinical utility arose from the use of a competing-risk framework and contemporary data estimation, rather than solely from the addition of new predictors.
2. Targeted Identification: The addition of specific predictors (e.g., learning disability, cancer therapies) allows for the identification of high-risk individuals at younger ages than age and conventional risk factors alone would suggest, potentially prompting earlier bone-health reviews.
3. Ethnicity and Fairness: Simply omitting ethnicity from prediction models does not guarantee equitable predictions; in this context, it led to calibration errors and overprediction in non-White groups. The inclusion of ethnicity improved calibration, though the authors note that ongoing evaluation of performance across population groups is required to ensure fairness.
4. Evolution of Tools: The findings support the development of prediction models that evolve alongside changes in survival rates, treatments, and population demographics, rather than remaining fixed to historical cohorts.

The paper does not claim to establish new intervention thresholds but suggests that these models provide more accurate and inclusive risk estimates to support clinical decision-making regarding bone-density testing and preventive interventions.