Osteoporosis International

, Volume 15, Issue 4, pp 274–280

Combination of bone mineral density and upper femur geometry improves the prediction of hip fracture


    • Department of Medical TechnologyUniversity of Oulu
  • Juha Partanen
    • Department of SurgeryUniversity of Oulu
  • Pekka Jalovaara
    • Department of SurgeryUniversity of Oulu
  • Timo Jämsä
    • Department of Medical TechnologyUniversity of Oulu
Original Article

DOI: 10.1007/s00198-003-1556-3

Cite this article as:
Pulkkinen, P., Partanen, J., Jalovaara, P. et al. Osteoporos Int (2004) 15: 274. doi:10.1007/s00198-003-1556-3


Bone mineral density (BMD) measured by dual-energy X-ray absorptiometry (DXA) is the main determinant of the clinical evaluation of hip fracture risk. However, it has been shown that BMD is not the only predictive factor for hip fracture, but that bone geometry is also important. We studied whether the combination of bone geometry and BMD could further improve the determination of hip fracture risk and fracture type. Seventy-four postmenopausal females (mean age 74 years) with a non-pathologic cervical or trochanteric hip fracture without previous hip fracture or hip surgery constituted the study group. Forty-nine had a cervical fracture (mean age 73 years) and 25 had a trochanteric fracture (mean age 76 years). The control group consisted of 40 age-matched females (mean age 74 years). The geometrical parameters were defined from plain anteroposterior radiographs, and the potential sources of inaccuracy were eliminated as far as possible by using a standardized patient position and calibrated dimension measurements with digital image analysis. BMD was measured at the femoral neck (FEBMD), Ward’s triangle (WABMD), and the trochanter (TRBMD). Stepwise linear regression analysis showed that the best predictor of hip fracture was the combination of medial calcar femoral cortex width (CFC), TRBMD, neck/shaft angle (NSA), and WABMD (r=0.72, r2=0.52, P<0.001). The area under the receiver operating characteristic curve (ROC) for this model was 0.93, while the area under ROC for TRBMD alone was 0.81. At a specificity of 80%, sensitivity improved from 52.5% to 92.5% with this combination compared with TRBMD alone. The combined predictors of cervical and trochanteric fracture differed, being NSA, CFC, TRBMD, and WABMD for cervical and TRBMD and femoral shaft cortical thickness for trochanteric fracture. In addition, we found a statistically significant correlation between FEBMD and femoral shaft and femoral neck cortex width (r=0.40, P<0.01 and r=0.30, P<0.01, respectively). The results confirm that the combination of BMD and radiological measures of upper femur geometry improve the assessment of the risk of hip fracture and fracture type compared to BMD alone, and that bone geometry plays an important role in the evaluation of bone strength.


Cervical fractureGeometryOsteoporosisRisk analysisTrochanteric fracture


Hip fractures, which are the most serious age-related osteoporotic fractures [1, 2], cause enormous medical and social costs, especially in developed countries. Therefore, the predictors of hip fracture are being extensively investigated. Falling mechanism, impaired mobility, and a variety of lifetime factors have been established as risk factors for hip fracture [3, 4]. Several studies have shown that bone mass and density correlate with the risk of osteoporotic fracture [5, 6, 7, 8]. Therefore, the measurement of bone mineral density (BMD) using dual-energy X-ray absorptiometry (DXA) is considered the main diagnostic method for the evaluation of fracture risk.

Bone geometry has also been shown to be important for the evaluation of the fracture risk [9, 10, 11]. Biomechanically, mechanical properties of bone at the tissue level are determined by both bone material quality and bone structure [12, 13], i.e. the strength of bone is influenced by both the material of which the structure is composed and the distribution and organization of this material. Therefore, knowledge of both material and geometry is required for an adequate evaluation of the structural integrity of bone. In the femoral neck area, strength properties have been shown to be determined not only by external geometry, but also by the thickness of the cortical shell [14]. The large variation in the mechanical behavior of bone is associated with geometrical variation [15]. Glüer et al. [9] have suggested that simple measurements based on pelvic radiographs predict fractures equally well as hip bone density.

There has been increasing interest in the combined impact of BMD and bone geometry on hip fracture risk. Different combinations of BMD, age, anthropometry and structural parameters have been used in the predictive models [9, 11, 16, 17, 18, 19, 20, 21, 22]. However, only a few studies have included cortical thickness, which is biomechanically important [14] and has also been shown to be a risk factor [9, 10]. It is also important to observe that the predictors of cervical and trochanteric fractures might be different. Some authors have focused on different fracture types, but the results are still somewhat confusing [9, 11, 20, 23]. Therefore, fracture type should also be considered.

The aim of the current study was to find out whether the combination of bone geometry and BMD can improve the prediction of hip fracture occurrence and fracture type. Geometrical measurements were performed by digitized analysis of standardized, calibrated plain radiographs.

Materials and methods

This study is based on a series described previously [4, 10].

Study subjects

The study subjects consisted of 102 consecutive postmenopausal females with non-pathologic cervical or trochanteric hip fractures without previous hip fractures or hip surgery who were admitted to Oulu University Hospital during the years 1998 and 1999. The inclusion criteria and ambulation at fracture were assessed by a trained nurse on admission. Twenty-eight patients aged over 84 years were excluded because of failure to find suitable controls with regard to age. Thus, 74 patients constituted the final study group (fracture group: mean age 74.2 years, range 53–84 years). Forty-nine patients had a cervical fracture (cervical group: mean age 73.1 years, range 53–84 years) and 25 a trochanteric fracture (trochanteric group: mean age 76.3 years, range 61–84 years).

The control group comprised 40 females (mean age 73.7 years, range 63–84) selected from among the clients who had bone densitometry done in a private clinic during the years 1998–1999. The exclusion criteria for the controls were hip fracture, any metabolic bone disease, and treatment with sex hormones, calcitonin, or bisphosphonates.

Written informed consent was obtained from all patients and controls, and the study protocol was approved by the institutional ethical committee.

Bone densitometric measurement

BMD of the upper femur was measured with two similar, equally tested and calibrated scanners (Lunar DPX, Lunar Radiation, Madison, Wisc., USA), using similar measurement routines. Before the measurements, a control phantom was scanned daily, and the same measurement program was used in both scanners. The coefficient of variation (CV) of the femoral neck in vivo reported by the manufacturer was 0.6–1.7%. The measurement of the patients was performed 2–4 days after the fracture.

Three parts of the hip (non-fracture side of the fracture patients and left side of the controls) were measured at the sites of the femoral neck (FEBMD), Ward’s triangle (WABMD), and the trochanter (TRBMD).


Anterioposterior pelvic roentgenograms of the fracture patients were taken within a few days postoperatively, and those of the controls were taken using the same X-ray equipment. A standard position was used in all cases: supine with the pelvis and both legs extended forward and the big toes touching each other, resulting in slight internal rotation of the femur. The beam was centered on the midline of symphysis pubis, and the focus-to-film distance was always 1 m. The 43 cm×38 cm cassette was placed under the patient. A calibration scale was fixed at the level of the greater trochanter of the uninvolved hip for calibrating the dimension measurements.

The X-rays were digitized with a CCD camera (Dage MTI 72E, Dage-MTI, Michigan City, Ind., USA) on a light table (Northern Light Desktop Illuminator, Imaging Research, Inc., Ontario, Canada), using an objective Canon CI-TV 16 mm lens (Canon, Tokyo, Japan), and digitally stored in a PC computer. The images were calibrated using the calibration scale, and the dimensions were measured with a digital image analysis system MCID M4 with the software version 3.0, revision 1.1 (Imaging Research, Inc.). Several dimensions of the uninjured hip were measured (Fig. 1): (1) upper femur dimensions: hip axis length (HAL), femoral neck axis length (measured in two ways: FNALa and FNALb), acetabular width (AW), femoral head diameter (HD), femoral neck diameter (ND), trochanteric width (TW), and femoral shaft diameter (FSD: measured at 3 cm below the center of trochanter minor); (2) cortical thickness: femoral neck cortex width (FNC), medial calcar femoral cortex width (CFC), and femoral shaft cortex width (FSC: measured at 3 cm below the center of trochanter minor); (3) femoral neck/shaft angle (NSA). All measurements were made by a single observer. The reproducibility was 0.9, 1.5, 2.5, 2.5, 3.3, 1.5, 0.7, 1.2, 1.1, 5.2 and 9.9% for FNALa, FNALb, FSD, TW, AW, ND, HD, HAL, NSA, FSC and FNC, respectively [10].
Fig. 1

Definition of the parameters measured from the anterioposterior roentgenograms of the upper femur. A-H hip axis length (HAL); A-B and A-C femoral neck axis length (FNALa and FNALb, respectively, measured in two ways); B-H acetabular width (AW); D-DD femoral head diameter (HD); E-EE femoral neck diameter (ND); F-FF trochanteric width (TW); G-GG femoral shaft diameter (FSD); I femoral neck cortex width (FNC); J medial calcar femoral cortex (CFC) width; K femoral shaft cortex width (FSC); P neck/shaft angle (NSA); O shows the calibration scale and N is a 3-cm bar generated with the software

Statistical analysis

Statistical analyses were performed with the SPSS statistical software (version 11.5.1, SPSS, Chicago, Ill., USA). Pearson’s linear correlation coefficients were calculated between BMD and the geometrical parameters. Student’s t-test was used to compare the fracture group with the control group and to compare the cervical and trochanteric fracture groups separately with the control group. A stepwise multiple linear regression analysis was performed to identify the best explanatory variables for fracture and for cervical and trochanteric fractures separately. The criterion for stepwise analysis was to continue iteration until the limit of P=0.05 was reached. The analysis gives a model for the combination of predictors that best explains the dependent variable in question. The sensitivity and specificity of each model were tested by calculating the area under the receiver operating characteristic curve (ROC), using the SPSS statistical software. Statistical comparison between the areas under the ROC curves was also performed [24].


Pearson’s linear correlation coefficients between BMD (measured at the three sites) and geometrical parameters were calculated. There were no strong correlations, the best found between FSC and FEBMD (r=0.40, P<0.01). FSC also correlated between WABMD (r=0.33, P<0.01) and TRBMD (r=0.34, P<0.01). Slight but statistically significant correlations were emerged between FNC and FEBMD (r=0.30, P<0.01) and WABMD (r=0.20, P<0.05) and TRBMD (r=0.22, P<0.05), respectively. TW correlated with TRBMD (r=0.20, P<0.05).

The mean geometrical parameters and BMD values of the study groups are shown in Table 1. FSC and CFC were significantly lower in all study groups compared with the controls (P<0.001). FNC and TW were also significantly lower in both the cervical group (P<0.001) and the trochanteric group (P<0.01) compared with the controls. FSD was remarkable lower only in the cervical fracture patients (P<0.001). NSA was significantly greater in the cervical group than in the controls (P<0.001), but there was no remarkable difference in the trochanteric group. BMD was significantly lower in all fracture groups than in the controls, with the exception of WABMD in the cervical group.
Table 1

Radiological parameters of the upper femur geometry and BMD. For abbreviations, see Materials and methods

































All fractures









































































































































*P<0.05, **P<0.01, ***P<0.001 when compared with the controls. +P<0.05, ++P<0.01, +++P<0.001 when compared with the trochanteric fractures. The results for all fractures versus controls and cervical fractures versus trochanteric fractures have been published previously [10]

The best explanatory variables for the different hip fractures are shown in Table 2. For any fracture, the best combination of explanatory variables was CFC, TRBMD, NSA, and WABMD (r2=0.52) in this order. The area under ROC was 0.93 for this combination, while the area under ROC for TRBMD alone was 0.81 (Fig. 2), which difference was statistically significant (P<0.001). At a specificity of 80%, sensitivity was improved from 52.5% to 92.5% by the model compared with TRBMD alone.
Table 2

The best explanatory variables for fractures, trochanteric fractures and cervical fractures by linear regression analysis. SE standard error

All fractures

Cervical fractures

Trochanteric fractures






2. FSC



















Area under ROC (SE) for model

0.93 (0.03)

0.95 (0.02)

0.94 (0.04)

Area under ROC (SE) for TRBMD alone

0.81 (0.04)

0.74 (0.06)

0.94 (0.04)

Area under ROC (SE) for FEBMD alone

0.81 (0.04)

0.74 (0.06)

0.97 (0.02)

Area under ROC (SE) for WABMD alone

0.74 (0.05)

0.65 (0.06)

0.94 (0.03)

Fig. 2

ROC curves for the regression models (left) and for BMD alone (right)

The best combination for cervical fractures was NSA, CFC, TRBMD, and WABMD (r2=0.53). The area under ROC was 0.95 for this model, while the area under ROC for FEBMD or TRBMD alone was 0.74 (Fig. 2), which was a statistically significant difference (P<0.001). At a specificity of 80%, sensitivity was improved from 47.5% to 92.6% by the model compared with FEBMD or TRBMD alone.

The best explanatory variables for trochanteric fractures were TRBMD and FSC (r2=0.61). The area under ROC was 0.94 for both the model and TRBMD alone. A slightly higher value of 0.97 was achieved with FEBMD alone, which difference was, however, not significant. At a specificity of 80%, sensitivity was 94.4% with the model and with TRBMD alone.


We evaluated the combination of BMD and bone geometry in the prediction of osteoporotic hip fracture risk and fracture type. The results showed a significant improvement in the assessment of the hip fracture risk and fracture type compared with BMD measurements alone. The study was performed by defining the geometrical parameters of the proximal femur from calibrated plain pelvic radiographs. BMD of the upper femur was measured with standard DXA.

The geometrical studies of upper femur have typically been based on direct measurements of non-calibrated radiographs or DXA images [11, 16, 17, 18, 19, 23]. However, when using radiographs or fan beam DXA, there is always some variation in the magnification of the images, which has to be considered. In the study of Karlsson et al. [17], DXA scans were more accurate than roentgenograms in evaluating the upper femur geometry because of the fixed position of the leg in their DXA measurements. However, even when they did not standardize the patient’s position for radiographs, they found that the hip fracture patients had a more valgus NSA than the controls as measured on radiographs; this difference was not revealed by the DXA measurements. With their pelvic radiographs, Glüer et al. [9] used a standardized protocol, and corrected for magnifications caused by the scanning process. Here, we also measured the geometrical parameters from standardized, calibrated plain pelvic radiographs, using computerized image analysis, which has been shown to be both precise and accurate [10].

We found statistically significant correlations between BMD and both FSC and FNC (P<0.01), which are in good agreement with the previous findings [25]. This correlation indicates that BMD measurement by DXA, which is a projectional method, is biased by the thickness of the cortical shell, i.e. the thicker the cortex, the higher the values for DXA-based BMD. This is a limitation of DXA that might be corrected, but according to Woodhead et al. [26], DXA has only limited value in determining cortical width.

In the present study, we focused on the differences between the patients with different fracture types, to define which parameters are relevant to identifying the fracture type. We found that NSA was significantly greater and FSD smaller in the cervical group, while no differences in these parameters were observed in the trochanteric group compared with the controls. Our finding is in good accordance with other studies, where NSA has been found to be a valuable indicator for the evaluation of hip fracture risk, especially in the case of femoral neck fracture [9, 11, 18, 19]. In some studies, HAL has also been shown to be remarkable for hip fracture [18, 20, 21], while some other studies have shown HAL not to be directly associated with the increased risk of hip fracture [17, 19]. Here, we were unable to detect any significant changes in HAL between either the cervical or the trochanteric group and the controls. These confusing results on HAL might be explained by magnification errors in the fan beam DXA [27] or differences in hip rotation and abduction [23]. Our finding on HAL by the calibrated measurement system is supported by the studies of Pocock et al. [27] and Michelotti and Clark [23]. In general, the differences in the reported results may be affected by the difference in the patient positioning in radiography and in DXA-based measurements.

The effect of cortical thickness on hip fracture has been evaluated in only a few studies. It seems to be an important factor for fracture type. Michelotti and Clark [23] found a difference in femoral cortex thickness between the fracture group and controls, as was observed also here. Cortical thickness might compensate for the variation in the shaft diameter or axis length based on Wolff’s law. This might be associated directly with inertia, which is an important factor for the mechanical strength of bone. However, as we have shown previously [10], cortical thickness correlates strongly positively with FSD, but no correlation was found between cortical thickness and HAL. Therefore, it seems that this compensatory mechanism does not properly strengthen the cortex in hip fracture patients.

Since there were a number of statistically significant parameters that may affect the occurrence of hip fractures, we performed stepwise multiple regression analysis to identify the best of them. The previously presented regression models for predicting osteoporotic fracture risk often include, in addition to geometrical parameters, age, weight and body mass index [11, 16, 19]. We focused exclusively on the geometrical properties of bone and excluded age, weight, and body mass index as predictive factors, to find out which combination of geometrical parameters and BMD is best able to predict the hip fracture risk and fracture type.

The best combination of explanatory variables for any hip fracture was CFC, TRBMD, NSA, and WABMD. This means that cortical thickness, bone density, and neck-shaft angle are all important for hip strength. The model includes two different BMDs, TRBMD and WABMD. They were measured from different parts of the upper femur and seem to be independently related to the fracture risk. The sensitivity of the combined model was better than in the study of Crabtree et al. [16] when it was compared with BMD alone. However, our model included both geometry and BMD, while their model was based on hip strength analysis, age, and body mass index. Our findings indicate that the combined model of bone geometry and BMD is highly sensitive and specific and the parameters are relevant to the prediction of fracture risk.

The model of cervical fracture included the combination of NSA, CFC, TRBMD, and WABMD as the best explanatory variables. Here again, the combination was more sensitive and specific than BMD alone. The model underlines the importance of NSA and cortical thickness in the assessment of the cervical fracture risk. For NSA, the results are in good agreement with the study of Alonso et al. [19], where an increase of 1 SD in the neck-shaft angle was associated with an odds ratio of 3.48 for hip fracture in women.

The best explanatory combination for trochanteric fractures was TRBMD and FSC. Site-matched BMD would seem to be an especially reliable indicator of trochanteric fractures. This is also supported by the findings of Gnudi et al. [11]. They reported site-matched BMD to be the best predictor of both cervical and trochanteric fractures. The authors concluded that the proximal femur geometry plays an important role only in neck fractures, but they did not include cortical thickness into their study. This might explain the small difference in the results. However, FSC did not improve significantly the model of trochanteric fractures. The area under ROC was 0.94 both with and without FSC, which shows site-matched BMD to be the most important predictor of trochanteric fractures. A slightly but not significantly higher value of 0.97 was achieved with FEBMD. This may be explained by the high correlation between cortical thickness and FEBMD. The stepwise analysis might have excluded FEBMD because of this high cross-correlation.

Typically, the quantity, quality, and shape of bone are changed continuously by bone turnover. Bone modeling occurs during growth and after bone trauma, but also in conditions of altered mechanical loading [28]. Modeling allows the development of normal growth but also modulation and repair mechanisms when the loading conditions change. Thus, individuals with greater NSA have lower cortical thickness than those with small NSA [29]. Partanen et al. [10] also reported a negative correlation between NSA and cortical dimensions, which might indicate that bone tends to increase its strength by adapting to the loading conditions. This indicates that human bone has not adapted to the loading conditions of the falling direction, which means that the rate of repair cannot keep up the normal structure of bone, and fractures may occur [30]. In general, bones are adapted for use, not for fracture [31]. This might be one potential factor to explain that no compensation mechanism was found for cortical thickness, there was no correlation between HAL and cortical thickness, and FSD and cortical thickness correlated positively.

We conclude that the combination of BMD and radiological upper femur geometric properties significantly improve the assessment of hip fracture risk when compared to BMD alone, and that bone geometry plays an important role in the evaluation of bone strength. The combination of cortical thickness, TRBMD, neck shaft angle, and WABMD was the best predictor of fracture. The predictors of different fracture types differed clearly, TRBMD being the most important predictor of trochanteric fracture, while the combination of neck shaft angle, cortical thickness, TRBMD, and WABMD was the best predictor of cervical fracture. Thus, the prediction of different fracture types should be evaluated separately. The predictors used in the current study can be defined reliably from standardized, calibrated plain radiographs, and the method might be useful in the clinical evaluation of the hip fracture risk and could be applied to the screening of patient groups with a high risk of hip fracture.


The University of Oulu Relief Fund is acknowledged for a grant to Pasi Pulkkinen.

Copyright information

© International Osteoporosis Foundation and National Osteoporosis Foundation 2004