World Health Organization fracture risk assessment tool in the assessment of fractures after falls in hospital

This study was conducted at Niigata University Hospital, an 810-bed academic teaching hospital in the city of Niigata. There are 23 clinical departments and the service area of the hospital as a tertiary care hospital covers all districts in Niigata Prefecture, which has a population of 2,400,000. All patients who had been admitted to the hospital during the period from April 2006 to March 2009 and who were aged from 40 to 90 years at admission were studied. During that period, 20,973 patients were admitted to the hospital, but 653 patients were excluded from the study because of missing data. Finally, data were obtained for a total of 20,320 patients aged from 40 to 90 years (median, 65.0 years; 25th percentile, 56.0 years; 75th percentile, 74.0 years). The patients included 9,738 females and 10,582 males, and 4,949 (24.4%) of the patients required acute admission. The dataset was randomly divided into two datasets of the same sizes by a person blinded to our study. One dataset was used for receiver operating curve (ROC) analysis to determine cut-off values (development dataset) and the other was used for validation of the analysis (test dataset).

Various risk assessment tools for prediction of inpatient falls have been developed, but only the STRATIFY tool and the Morse Falls Scale [18] have been subjected to prospective validation in several cohorts with appropriate tests of predictive validity [7, 8]. The STRATIFY tool showed high sensitivity and specificity in the original study, and its simplicity has facilitated its wide use in clinical practice [19]. We therefore used this tool in our study to assess patients' risk for falls. However, systematic review of the STRATIFY tool revealed that the tool may not be optimal for identifying individuals at high risk for falls and that population and setting affect performance of the tool [10]. The risk factors of fractures related to osteoporosis are age, prior fragility fracture, parental history of hip fracture, smoking, use of systemic corticosteroids, excess alcohol intake and rheumatoid arthritis [15]. By integrating these risk factors, WHO proposed the FRAX™ tool to compute ten-year probability of osteoporotic fracture. The FRAX™ tool has the advantage that it can be used without information on BMD and is adjusted for ethnic differences. Actually, it is used to determine thresholds for therapeutic intervention in a Japanese setting [14]. We therefore used the tool in this study to assess the risk for fracture.

Information on patients' background such as age, gender, body weight, height, history of bone fractures, smoking history, alcohol consumption, prescription of drugs, coexisting illness, admission day and discharge day was obtained from the hospital information system. Information on risk factors for falls was obtained from medical charts of the patients and fall assessment records completed by attending nurses at admission [8]. They included history of falls, gait instability, agitated confusion, urinary incontinence or frequency, visual impairment, lower limb weakness and prescription of 'culprit' drugs. The assessment was performed again when fall events occurred. Therefore, STRATIFY score [17] and FRAX™ score [14, 16] were calculated at admission and when the fall events occurred. STRATIFY score was calculated on the basis of the original method [17], and FRAX™ score was based on the ten-year probability of major osteoporotic fracture according to body mass index.

The clinical outcome we studied was fallers with or without fractures rather than falls [9, 20]. Data on fall events were obtained from online incident reports, records of x-ray order entry and medical charts of the patients. The incident reports concerning fall events were documented by the attending nurse and other medical staff, and the reports contained data on time, location, injury sustained and potential causative factor for falls. Peripheral fractures verified by radiological findings were included, but vertebral fractures were excluded from the study [20].

Fall events and fracture after falls were analyzed by two different methods. The first method is the traditional chi-square test and multiple logistic analysis, which has been used for analysis of inpatient falls. For multiple logistic analysis, significant risk factors were selected by using the stepwise selection method and by the forced entery method. The second method is survival analysis in which time between admission and the event (falls or fractures after falls) during the hospital stay was considered as survival time. Discharge of the patient without fall events was considered as censoring. The Kaplan-Meier method was used for the analysis, and the logrank test was used to examine whether each risk factor was significantly associated with events. Multivariate Cox's proportional hazards regression model with time-dependent covariates was used to examine the risk factors that were most significantly associated with events among the various risk factors. Since values of several risk factors such as history of falls varied over time, these factors were included in the model by defining them as time-dependent covariates. Significant factors were selected by the stepwise selection method and by the forced entry method. A cut-off value to distinguish patients at risk from patients not at risk was determined on the basis of results of ROC analysis of data in the development dataset. The value corresponding to the nearest point of the ROC curve to the top left-hand corner was chosen as a cut-off value to distinguish patients at risk from patients not at risk [21]. Sensitivity, specificity and area under the ROC curve (AUC) were calculated. Distribution of continuous data was shown by medians (25-percentiles, 75-percentiles). All statistical analyses were performed using SPSS Statistics 17.0 (SPSS Japan Inc., Tokyo, Japan), and a p-value less than 0.05 was considered significant.

All data were analyzed anonymously. The Ethical Committee of Niigata University School of Medicine gave ethical approval.

The numbers of patients who experienced more than one fall during admission were 308 (3.1%) in the development dataset and 345 (3.4%) in the test dataset. Eight (2.6%) of the fallers in the development dataset and 10 (2.9%) of the fallers in the test dataset suffered peripheral fractures after falls. The ages of the patients who suffered fractures after falls were 78.0 (64.0, 83.0) years for the development dataset and 75.0 (64.5, 76.5) years for the test dataset. All but one of the patients who suffered peripheral fractures broke bones during the first fall.

Univariate analyses (chi-square test and logrank test) revealed that various known risk factors including dichotomized STRATIFY score were significantly associated with fallers (Table 1). The cut-off value of the STRATIFY score to predict falls was determined to be a value of 2 based on results of ROC analysis of data in the development dataset. Proportions of high-risk patients based on the cut-off value of STRATIFY score were 26.5% in the development dataset and 26.8% in the test dataset. Sensitivity and specificity of the cut-off value to predict falls were 0.648 and 0.749 in the development dataset and 0.678 and 0.749 in the test dataset, respectively (Table 2). LOS was dichotomized at 14 days, which was the median value for all patients. Only the factor 'visual impairment' was not significant. When the risk factors as well as STRATIFY score were entered simultaneously into the multiple logisitc model, factors shown in Table 3A were significantly associated. On the other hand, multivariate Cox's regression analysis showed that the factors in Table 3B were significantly associated with falls. In both datasets, the factor 'history of falls', factor 'gait instability' and STRATIFY score were significantly associated with falls in both analyses.

The chi-square test revealed that the factor 'lower limb weakness' and the FRAX™ score were significantly associated with fracture after falls in both datasets (Table 4). In addition, dichotomized STRATIFY score was a significant factor in the development dataset and dichotomized LOS was a significant factor in the test dataset. The cut-off value of the FRAX™ score to predict fracture after falls was determined to be as a value of 10 based on results of ROC analysis of data in the development dataset. Proportions of high-risk patients based on the cut-off value of FRAX™ score were 34.1% in the development dataset and 33.9% in the test dataset. Sensitivity and specificity of the cut-off value were 0.750 and 0.659 in the development dataset and 0.700 and 0.661 in the test dataset, respectively (Table 2B). On the other hand, the logrank test revealed that only dichotomized FRAX™ score was significantly associated with fractures after falls both in the development dataset and the test dataset (Table 4, Figure 1). When the risk factors shown in Table 4 were entered into the multiple logisitc regression model, only FRAX™ score was selected as a factor significantly associated with fractures after falls by the stepwise selection method both in the development dataset and the test dataset (Table 5A). Similarly, multivariate Cox's regression analysis showed that only FRAX™ score was a significant factor (Table 5B). Next, we analyzed how the combination of STRATIFY and FRAX™ tools would work for prediction of fracture after falls. Both scores were forcibly entered into the multiple logisitc model (Table 5C) and multivariate Cox's regression model (Table 5D), but only FRAX™ score was significantly associated with fracture after falls in Cox's regression analysis.

The performance of the two risk tools in a much more selected population was evaluated. We performed multiple logisitc analysis and multivariate Cox's regression analysis by entering the risk factors as well as the two risk scores simultaneously into the models for patients more than 65 years old, who were at higher risk for both falls and fractures. The results were almost the same as the results of analysis for all subjects (Table 6). The factor 'history of falls', the factor 'gait instability' and STRATIFY score were significantly associated with falls both in multiple logistic regression analysis (Table 6A) and multivariate Cox's regression analysis (Table 6B). FRAX™ was significantly associated with fracture after falls in the development dataset analyzed by multivariate Cox's regression analysis and in the test dataset analyzed by both methods.