Information for this study was obtained from the General Practice Research Database (GPRD). The GPRD comprises prospectively collected computerized medical records for more than 10 million patients under the care of general practitioners in the United Kingdom. It has been the source for numerous epidemiologic studies, and the accuracy and completeness of these data have been well-validated and documented.19
Previous studies of GPRD data have shown a high level of data validity with respect to the reporting of fracture.20
GPRD data have been linked to the national Hospital Episode Statistics in England for approximately 45% of all practices.21
In this study, data were linked from April 1997 until March 2008.
The study population has been described previously and is available under an open access license.2
In short, we identified all patients aged 18 years or older with at least one recorded diagnosis of MS during the period of GPRD (January 1987–August 2009) or Hospital Episode Statistics data collection (April 1997–March 2008). Patients with a history of MS before the start of valid data collection were excluded. Each patient with MS was matched by year of birth, sex, and practice to 6 control subjects. The index date of MS diagnosis was the date of the first record of MS, after start of valid data collection. Control patients were assigned the same index date as their matched case patient. Each patient was followed from index date to the end of data collection, the date of transfer of the patient out of the practice area, or the patient's death, whichever came first. For the development of the prediction model, we followed FRAX in the decision to exclude patients who were currently treated with osteoporosis medication or who had previously taken medicines against osteoporosis. Therefore, we excluded all patients who had ever used osteoporosis treatment at baseline, which included prescriptions of bisphosphonates, raloxifene, strontium ranelate, and parathyroid hormone.
All patients were followed up for the occurrence of fractures. The types of fracture were classified according to the International Classification of Diseases (ICD-10) categories. These included skull (S02), neck (S12), ribs (S22), pelvis (S32), shoulder (S42), forearm (S52), hand (S62), hip/femur (S72), ankle (S82), foot (S92), or unspecified fractures (T02, T08, T10, and T12). A clinical osteoporotic fracture was defined as a fracture of the radius/ulna, vertebrae, femur, hip, humerus, pelvis, or ribs.
The presence of risk factors was assessed by reviewing the computerized medical records at baseline. Potential risk factors included older age, female sex, low BMI, being a smoker, drinking alcohol, history of falling 3 months–1 year before, history of fracture, and history of chronic diseases (congestive heart failure, rheumatoid arthritis, cerebrovascular disease, inflammatory bowel disease, dementia, depression, epilepsy, and chronic obstructive pulmonary disease). In addition, evidence of fatigue, visual disturbances, vertigo, dizziness, imbalance, disturbance of sensation, spasticity, sexual dysfunction, paroxysmal symptoms, cognitive dysfunction, vitamin D deficiency, and proxy indicators of increased disability (home visits by the general practitioner, nursing care, patient receiving residential care/living in a care home, or patient using a wheelchair or walking aid) 6 months before the index date were considered. Further, prescriptions for PO or IV administered GCs, statins, antiarrhythmics, antidiabetics, antidepressants, antipsychotics, hypnotics/anxiolytics, asthma medication, anticonvulsants, hormone replacement therapy, vitamin D, levothyroxine, baclofen, opioids (potencies equivalent to tramadol or higher), nonsteroidal anti-inflammatory drugs, meprobamate, tizanidine, dantrolene, modafinil, methylphenidate, or amantadine 6 months before the index date were also considered potential risk factors.
Cox proportional hazards models were used to calculate the long-term risk of osteoporotic and hip fracture. The Cox model allows calculation of an individual's probability of fracture (i.e., survivor function) for each set of patient characteristics. For the analysis of long-term risk, we first fitted the regression model with all possible risk factors, which were determined at baseline. All characteristics, except age, were included as categorical variables in the regression models. For BMI, we used categories <20 kg/m2, 20–25 kg/m2 (reference), >25 kg/m2, and unknown. Smoking status was divided over current smoking, ex-smoking, never smoking (reference), and unknown smoking status. Alcohol consumption was similarly categorized as current, ex, never (reference), and unknown. All other variables addressing medical history and the prescription of drugs were separately added and were categorized as yes or no. We used the PHREG procedure in SAS with forward selection and a significance level of 0.05. This resulted in a list of variables that were possible candidates for our prediction model.
For osteoporotic fracture, the following variables came out of the forward selection. MS, sex, age, use of GCs in the prior 6 months, use of anticonvulsants in the prior 6 months, history of fracture, current smoking, and BMI >25 kg/m2. Because of clinical importance, we decided to add BMI <20 kg/m2, use of antidepressants in the prior 6 months, and history of falling 3 months–1 year before. For the prediction of hip fracture, forward selection resulted in the variables MS, age, use of antidiabetics, baclofen, and amantadine in the prior 6 months, history of fatigue in the prior 6 months, BMI <20 kg/m2 and BMI >25 kg/m2. In this case, we considered sex, use of GCs in the prior 6 months, use of antidepressants in the prior 6 months, history of falling 3 months–1 year before, history of fracture, and current smoking important variables to add. However, because there were only 104 hip fractures among the patients with MS and control subjects, we were restricted to 10 predictors. Therefore, we dropped the use of antidiabetics, baclofen, and amantadine in the prior 6 months. We also dropped a history of fracture, because this variable led to β of 0. Next, we investigated possible statistical interactions of these selected variables with MS. To account for multiple comparisons, we applied a Bonferroni correction. None of the interaction terms was subsequently added to the model.
The β coefficients in the final Cox model were converted into integer risk scores. The value of each integer was calculated as the rounded sum of the Cox model predictor scores, multiplied by 10. The 5- and 10-year risks of fracture were then estimated using these scores, conditional on patient survival. Various methods were used to test the fitting of the Cox models, including a test of the proportional hazards assumption. We also compared the observed 5-year probability of fracture (based on the Kaplan-Meier estimate) with the probability predicted by the Cox model. To assess the internal validity of the model further, the C-statistic was calculated, and we performed a 10-fold cross-validation. We applied the shrinkage factor that we found to the β coefficients of the model, and we adjusted the C-statistic for overestimation. We compared 10-year risks of osteoporotic and hip fracture as predicted by FRAX and by the present risk score.
As a sensitivity analysis, we also evaluated backward selection and stepwise selection instead of forward selection. Both methods resulted in exactly the same set of predictors, both for osteoporotic and for hip fracture. All data management and statistical analyses were conducted using SAS 9.1/9.2 software.
Standard protocol approvals, registrations, and patient consents.
The study was approved by the Independent Scientific Advisory Committee of the GPRD.