Associations between post-operative rehabilitation of hip fracture and outcomes: national database analysis

The NHFD is a clinically-led, web-based audit of hip fracture care and secondary prevention in England, Wales, Northern Ireland and the Channel Islands. All 253 eligible hospitals who treat hip fractures are registered with NHFD, and 98% participate by regularly uploading case records in a standard dataset format. All patients aged 60 and over with a hip fracture coded as S72.0, S72.1, S72.2 in ICD-10, including subtrochanteric fractures, are submitted, regardless of the fracture mechanism. The NHFD is approved by the NHS England Health Research Authority Confidentiality Advisory Group to collect patient data without consent under Section 251 exemption, so no ethical approval was required for this study. We carried out the same data quality checks used by NHFD on the 2013–15 data we received [16].

To measure the effects of rehabilitation it would be most useful to have an activity type variable which described the type of rehabilitation delivered by specialist, mainly PT staff and the intensity e.g. number of sessions. As this is not part of the NHFD dataset we were forced to use proxies for rehabilitation activity or exposure. We used “mobilised on day of or day following surgery” as one measure of rehabilitation exposure. It is recognised that not all patients will be suitable for such mobilisation, and this field only seeks to capture how quickly individuals progress in initial physiotherapy. A patient would be described as ‘mobilised’ if they are able to sit or stand out of bed on the day of their return from operation, or on the following day. We used the variable “assessed by PT on day of or day after surgery” as another measure of rehabilitation exposure. Unfortunately these two variables are optional so some data is missing.

We used demographic variables known to be associated with hip fracture outcomes, including age, sex, and delay from admission to operation. The Abbreviated Mental Test Score (AMTS), which is assessed pre-operatively, is a simple and robust 10 item screening tool for acute patients. In AMTS poor cognitive function = 0 and good cognitive function = 10. The American Society of Anaesthesiologists (ASA) score is a preoperative risk score based on the presence of co-morbidities at the time of surgery. An ASA score > 2 is associated with increased risk additional to that of classification of operation and duration of surgery.

The first outcome variable was 30-day mobility score on a scale from 1 to 5, which counter-intuitively is higher for less mobile patients in NHFD. A new single mobility score has recently (2015) replaced all four measures previously used by the NHFD. In order to replace missing values of the single mobility score in older data, we used the translational matrix shown in Additional file 1: Table S1 in the online Supplementary Data file.

The second outcome variable was discharge destination from acute orthopaedic ward in combination with the location from which the patient was admitted. For the preoperative setting we used “admitted from”. The method for combining the two variables is shown in Additional file 1: Table S2. Discharge destination models were fitted using the 30-day discharge destination combined with preoperative setting expressed on an integer scale from 1 to 4, with higher integers indicating better outcomes (1 = “Dead”, 2 = “Worse”, 3 = “Same”, 4 = “Better”).

Ordinal logistic regression models can be used where the outcome variable is nominal or ordinal i.e. with more than two levels [17, 18]. In mathematical notation:

We tested all the variables of interest in the NHFD dataset to check whether they were statistically associated with discharge destination (χ² tests and likelihood ratio tests). These variables were used to fit the ordinal logistic regression models. We used the predict command in Stata 14 to compute the ordinal-logistic odds ratio. We computed the c-statistic for a predictive score to measure the performance of the model.

Propensity score methods can be used for confounding control in non-experimental research [19]. We also carried out a propensity scoring or matching analysis, which is often used to obtain a better estimate of effect size in real-world examples. Propensity score models do not rely on modelling the outcome, but on a model of the treatment given the confounders [20, 21].

The mobility outcome was mobility at 30 days post-operatively. The Mobility models were fitted by using complete cases only. The treatment or exposure variable was mobilisation on the day of or day after surgery. Confounders were gender, age at discharge, hours elapsed between admission and surgery, pre-operation AMTS, ASA grade, and pre-fracture mobility score (1 to 5 or missing).

We defined a propensity score for mobilisation treatment with respect to confounders, using a logistic regression model, then defined weights to estimate the average treatment effect (ATE), [21] equal to the reciprocal of the fitted probability of mobilisation for the mobilised and non-mobilised patients respectively. More information on the propensity scores, weights and balance checks is provided in the online Additional file 1: Figures S3-S5. Having defined the propensity weights, we used Poisson regression models, with Huber variances, to regress the 30-day mobility score with respect to the binary treatment variable (mobilisation), with two mean mobility scores for mobilised and unmobilised patients, respectively. The first model was unadjusted/unweighted, and the second model was weighted using the ATE weights. For both models, we estimated the ATE using the Stata 14 add-on package scenttest, [22] which carries out a scenario t-test between the mobilised and unmobilised scenarios. We carried out balance checks for the propensity score using the Stata add-on package somersd [23].

Discharge destination models were fitted as for Mobility. We defined a propensity score for mobilisation treatment with respect to confounders, using a multinomial logistic regression model. For each patient, we defined three propensity scores, one per mobilisation group, as the fitted probability of the patient being in the mobilisation group. We then defined weights for estimation of the ATE [21]. Therefore we were directly standardising the sample to a hypothetical target population, with the same distribution of covariates as the sample, but with no association between treatment-propensity and treatment. The unadjusted Somers’ D suggested that our ATE weights had balanced the propensity score. We then used Poisson regression models and the Stata packages as for the mobility outcome.

Most of the baseline characteristics of the complete cases in the Mobility model population after data translation are shown in Table 1. For further background, Additional file 1: Table S3 in the online Supplementary Data shows the mobility characteristics of the whole (both complete and incomplete cases) Mobility model outcome population before data translation. For comparison, Additional file 1: Table S4 shows the baseline characteristics of mobility variables before and after mobility data translation. Additional file 1: Table S5 shows all the baseline characteristics of Mobility Model population after data translation, for complete cases only.

Χ² tests (Additional file 1: Table S1) showed patients’ pre-fracture mobility levels were most strongly associated with mobility 30 days after discharge, and AMTS, ASA grade and patients’ age were also strongly associated with post fracture mobility. Gender and whether patients received PT assessment were weakly associated with post fracture mobility.

Table 2 shows the univariate and multivariate ordinal logistic regression analyses of the mobility model (as noted in the Methods a higher score indicates worse mobility). After adjustment, mobilisation on the day or the day following surgery is associated with a lower mobility score i.e. better mobility function 30 days after discharge. In the univariate analysis, mobilisation by a PT day or the day following surgery was more protective from a high score than mobilisation by other health professionals. However, in the adjusted/multivariate analysis, patients mobilised by a PT were not protected from poorer mobility compared to mobilisation by other professionals. Similarly, in both analyses, patients who received a PT assessment were not protected from poorer mobility 30 days after discharge, compared with those who did not receive an assessment.

We assessed discrimination of the ordinal logistic model in the same way as a traditional logistic model, using area under the receiver operating characteristic curve [24]. We fitted three different models but there was no difference in discrimination between them (area under the receiver operating characteristic curve = 0.755).

To further investigate associations of early mobilisation we then carried out a propensity-adjusted analysis comparing 30-day mobility between mobilised and unmobilised patients i.e. a two-category outcome (by grouping the PT and non-PT mobilisation). The Somers’ Ds of the propensity score suggested that our ATE weights had balanced the propensity score between the treatment groups. We used the propensity weights to estimate the ATE, equal to the difference between the two scenario means, with PT and Other grouped together. The results in Additional file 1: Figure S2 show that the mobilisation-propensity score predicts mobilisation, with a Somers’ D of 0.298. The ATE-adjusted scenario means are a little closer together than the unadjusted scenario means. The outcome is better (lower mobility score) in mobilised (=Yes) patients than in unmobilised patients (=No), whether the mobilisation is done by a PT or by somebody else, and the difference persists (slightly attenuated) after propensity weighting. The unadjusted results suggest that patients mobilised by a non-PT have a lower propensity score i.e. are better mobilised than patients mobilised by a PT, but the propensity-adjusted results suggest that this crude difference may easily be due to differences in the patients. This might happen if more vulnerable patients were allocated mostly to PTs and less vulnerable patients were allocated mostly to non-PTs. More detail about propensity scores and weights is provided in the online Additional file 1: Figure S3 is a forest plot of confounder categories vs Somers’ D score which shows that ATE-weighted propensity scores for confounders are greatly reduced.

To further investigate the associations with who performs early mobilisation we then carried out a propensity-adjusted analysis comparing 30-day mobility score between mobilised (mobilised by PT or mobilised by other professionals) and unmobilised patients i.e. using three categories to separate the effects of PTs and other staff (Fig. 1). The results in Fig. 1 show that the mobilisation-propensity score predicts mobilisation, with a Somers’ D of 0.298, as for the two-scenario analysis. The ATE-adjusted scenario means are a little closer together than the unadjusted scenario means, but not a lot. The outcome is better in mobilised patients than in unmobilised patients, whether the mobilisation is done by a PT or by somebody else. And the difference persists (slightly attenuated) after propensity weighting. However, it does not seem to make much difference whether the mobilisation is done by a PT or by somebody else, at least after propensity weighting.

Additional file 1 Table S2 shows the baseline characteristics of the discharge destination model population. We again used propensity weights to estimate the ATEs, equal to the differences between the three scenario means (fully mobilised by PT, fully mobilised by others and fully non-mobilised) (Fig. 2). For this outcome a higher score indicates a better outcome. The mobilisation-propensity score predicts discharge destination, with a Somers’ D of 0.332. Figure 2 shows that the ATE-adjusted scenario means are a little closer together than the unadjusted scenario means. The outcome is better in mobilised than unmobilised patients, whether done by a PT or by somebody else, and the difference persists (slightly attenuated) after propensity weighting. The results are quite similar to the 30 day Mobility model, suggesting that this is a reliable association.

In summary, we investigated the associations of early mobilisation, as a proxy for rehabilitation, mainly delivered by PTs, with two outcome variables, both before and after adjustment for confounding variables. In both in the 30 day mobility and discharge destination outcome models, we confirmed that patients mobilised early had a better outcome than those not mobilised early, whether they were mobilised by PTs or by non-PTs. The unadjusted analyses suggest that patients mobilised by non-PTs were less immobile at 30 days, but after adjustment, patients mobilised by non-PTs did as well as patients mobilised by PTs. There are a number of possible explanations for this: there may be no actual difference, PTs could mobilise more vulnerable patients with a worse prognosis, or non-PTs such as nurses may deliver more rehabilitation sessions. However we also showed that 57% of patients are currently being mobilised early by a PT and only 3% by “Others”, presumably nurses. If our findings are confirmed, PTs’ roles in early rehabilitation should be reconsidered, with a view to redeploying them to more effective rehabilitation activity.

There are few published cohort studies which have analysed rehabilitation exposures and outcomes, and in most instances they only measured effects of transfers to rehabilitation facilities. For example, in an Austrian cohort of 281 patients, [25] patients who were transferred to a nearby acute geriatric hospital for rehabilitation had significantly higher functional outcome scores. A study from the Scottish Hip Fracture Audit of 2708 patients aimed to determine whether place of residence is associated with a difference in access to comprehensive rehabilitation, rather than the effects of rehabilitation exposure [15]. It found that previously mobile care home patients were less likely to return to pre-admission levels of function at 120 days post fracture.

We used two methods and two outcomes and obtained similar results from both, suggesting that our findings are robust. However, both the models could be applied to only the sub-population of patients with complete exposure and outcome data.

Unmeasured confounding is another limitation. Although we adjusted for five confounders, it is possible that other patient factors influenced outcomes. For example, a US study was assessed the relationship between self-reported disease burden and functional improvement during and after inpatient rehabilitation [26]. Compared with patients without chronic conditions, those who had a stroke had significantly worse self-care, transfer and locomotion ratings.

The biggest limitation in our analysis is that we were forced to use binary variables for early mobilisation or PT assessment as a proxy for rehabilitation, because there is inadequate data about the process of rehabilitation in the NHFD. There is likely to be great variation in the amount of early mobilisation rehabilitation, and there is little published data on the dose-response relationship. The hip fracture guideline from the UK’s National Institute for Health & Care Excellence found only one, small randomised controlled trial with 60 patients of early versus delayed mobilisation, and three randomised studies with a total of 288 patients using three different types of PT programmes, which were not comparable and so could not be meta-analysed [27]. The American Academy of Orthopaedic Surgeons guidelines for the management of hip fractures in the elderly recommend intensive supervised occupational therapy and physiotherapy in hospital and home, [28] but the studies it quotes as evidence carried out extensive home rehabilitation programmes, [29] for example an intervention group received a median of 4.5 home visits, (PT median 3, minimum 0– maximum 7 visits; occupational therapist median 1.5, minimum 0–maximum 8 visits; and 11 patients were visited by a nurse) [30]. Real world rehabilitation may not have such a high dose.