Internationally, hospital discharge data are used widely to determine the incidence of hospitalization for injurious falls. However, a major limitation of this is that patients who have been readmitted or transferred from one hospital to another, and in some instances even within the same hospital, for treatment of the same injury may by recorded more than once. Hospital discharge datasets usually lack unique personal identifiers to identify individual patients and information on the date of injury that would facilitate identification of multiple admissions for the same injury.

In the absence of personal identification and date of injury information, some studies have simply assumed that each discharge refers to a single episode of injury, resulting in an overestimation of the true hospitalized falls incidence rate.^1,2 Others have used various case definitions based on variables readily available in most hospital discharge datasets, such as “readmission within 28 days”, “transfer”, “single day admission”, and “admission to non-acute hospitals”, to identify incident cases.^3–8 The aim of this study was to assess the validity and estimate the effects of these case definitions on the calculation of hospitalized falls incidence rates—that is, single admissions, as well as (in the case of multiple admissions) the first hospital admission for the same injury.

METHODS

Information on hospitalized injurious falls cases was obtained from the 2000–01 Inpatient Statistics Collection which covers all inpatient separations/discharges from public hospitals in New South Wales (NSW). Hospitals in NSW are required to submit details for every inpatient episode of care. An episode of care ends by discharge, transfer, death, or by the patient becoming a different type of patient within the same hospital.⁹ Cases selected for this study included all public hospital admissions in NSW of patients aged 50 years with an ICD 10 external cause/mechanism of injury indicating a fall (W00-W10) and a primary diagnosis of injury (S00-T98).¹⁰

An internal probabilistic data linkage method was adopted, using the record linkage package LinkageWiz,¹¹ to identify multiple separations for injurious falls for each patient. The strength of the probabilistic linkage procedure is that it takes into account spelling errors and missing values in each field used for matching.¹² Matching variables used in the process included the phonetic encoding of family names, middle name and given name as well as the date of birth (day, month, year), sex, address of residence (street number, street name, postcode), country of birth, language spoken at home, as well as the principal diagnosis and the external cause of injury.

The rate of first admission was examined by age, sex, and injury type. Using the first admissions identified by the record linkage as the “gold standard”, we also examined the sensitivity and specificity of other classification approaches routinely used to estimate the number of first admissions of hospitalized falls. These include the exclusion of transfers, readmissions within 28 days, admissions to non-acute hospitals, and “day only” admissions.

The incidence of injurious falls leading to hospitalisation was computed for each approach using the number of NSW residents aged 50 years and over. Ninety five percent confidence intervals were computed for various proportions. All analyses were carried out using SAS (version 8.1).¹³

RESULTS

There were 22 250 separations for injurious falls in people aged 50 years and older in NSW between the 1 July 2000 and 30 June 2001. Of these cases, 20 883 (93.9%, 95% CI 93.5 to 94.2) cases were identified, using data linkage, as incident cases for hospitalized injurious falls.

The proportion of incident cases of fall related injuries leading to hospitalization did not vary significantly by age and sex. Similarly, although vertebral fractures, fractures of other parts of the femur, and fractures of skull and facial bones had the lowest proportions of incident cases compared with other fracture types, the difference was not statistically significant.

When comparing the performance of different approaches to identifying incident cases, compared to the data linkage, the “transfer from” variable performed best in terms of sensitivity and specificity (table 1). A combination of both “readmission within 28 days” and “transfer from” improved the specificity but reduced the sensitivity. A case definition based on the exclusion of admissions to non-acute hospitals had a relatively high sensitivity but a low sensitivity and the reverse was true for day only admissions.

When examining the impact of these approaches on the incidence of hospitalized injurious falls, the exclusion of non-acute admissions approach provided the best estimate of incidence compared to that provided by data linkage (table 1).

DISCUSSION

This is the first study to examine the validity of various approaches to identifying first hospital admissions for injurious falls. Using data linkage methods, 93.9% of hospital separations for injurious falls were found to correspond to first admissions/incident cases. A similar proportion was found in two other studies using different methods.^14,15

In the absence of more specific variables such as a unique personal identifier and the date of injury, our results suggest that case identification approaches based on readmission/transfer variables, particularly when used in combination, may provide the most accurate identification of incident falls leading to hospitalization. However, the incidence rate of hospitalized injurious falls when using this approach was 9.4% lower than that based on data record linkage, raising questions about its utility in estimating incident cases.

According to coding guidelines, the “readmission within 28 days” variable should be used to indicate a readmission for the same problem/condition.¹⁶ However, it is possible that some hospitals/coders may be using this even when a previous admission is related to other injuries or health conditions. The identification method based on “day only admissions” had a very low sensitivity and is simply not suitable for estimating incident cases of hospitalized injurious falls. This is reflected in the lower incidence rate found when using this method compared to the “gold standard”.

Although the inclusion of “acute admissions only” had a low specificity, it yielded a high sensitivity and a comparable incidence rate (only 2% higher) to that resulting from the use of probabilistic data linkage. Although sensitivity is important in this context, as it is more indicative of “true” first admissions/incident cases, it is equally important to achieve a high specificity and avoid false positives, as various attributes might change between first and subsequent admissions.

The unsatisfactory validity (specificity in particular), in relation to data record linkage, of the variables readily available in most hospital datasets indicates that alternative methods to identify incident cases of injurious falls admitted to hospitals are needed. In the state of California, only first admissions for an injury are assigned an external cause which means that by selecting records with an external injury code, only incident cases are selected and multiple counting of cases is avoided.¹⁷ However, this method is limited when the aim is to measure hospital bed use and the economic impact of injury on the delivery of health care.

As probabilistic record linkage is far from being practical in identifying repeat admissions, the use of a unique patient identifier coupled with the date of injury remains the most valid and accurate method of identifying incident cases of injurious falls as well as other conditions.¹⁴ Unique patient identifiers have the potential to not only identify incident cases for various hospitalized health conditions but to also ensure the safety of patient care by enabling access to patient encounter information across the continuum of care.^18,19 Routine recording of the “date of injury” would further simplify the process of determining whether these re-admissions are for the same injury or for a new injury which might have occurred at a later date. This is particularly relevant to falls in older people where more than one fall can occur in a given year.

Although efforts have been made to achieve an accurate linked dataset, data record linkage involves trade-offs between the number of false positives and false negatives and the inclusion of such cases in the final dataset is inevitable. However, previous studies have shown that probabilistic record linkage techniques, similar to those used in this study, result in high quality outcomes and that the process results in no more than a 1% error rate.²⁰ This is supported by the findings of another study, which used a unique personal identifier and date of injury, and found the same proportion of incident cases of hospitalized injurious falls as our study.¹⁴

In conclusion, all the approaches using variables readily available in most hospital discharge datasets to estimate incident cases of hospitalised injurious falls have relatively low specificity, raising questions about their use. This emphasizes the importance of the introduction of a unique patient identifier and the date of injury in providing a more accurate picture of incident cases of injury related hospitalizations, including those resulting from falls.