A prediction model to identify hospitalised, older adults with reduced physical performance

A prospective cohort study was conducted in a short-stay unit in a Danish emergency department (ED) from December 2014 to May 2015 [23]. In Denmark a short-stay unit provides targeted care for 48–72 h, followed by patient discharge or transfer to an in-patient unit. All participants were enrolled consecutively and assessed within the first 48 h of admission and again at a follow-up home visit that took place a minimum 14 days after the date of admission.

We recruited patients ages 65 years or older who were admitted to the short-stay unit, who resided in the hospital’s catchment area, and who were admitted with a medical diagnosis (rather than a surgical or psychiatric diagnosis). Common medical diagnoses included infection, thromboembolic disease, musculoskeletal disease, and cardiovascular disease, but not patients with obvious signs of stroke or ST-elevation myocardial infarctions. Patients were enrolled in the study if they demonstrated reduced physical performance within the first 48 h of admission, specifically if they performed ≤8 repetitions in 30s–CST. We assumed that older adults who performed >8 repetitions in the 30s–CST were without significant risk of losing functional mobility, and hence the rationale for their exclusion from the study. Additional inclusion criteria included patient ability to sit on a chair independently within the first 48 h of admission, patient orientation to time and place, and patient ability to speak and understand Danish. Patients who could not walk at their baseline health were excluded.

The sole study outcome measurement was the 30s–CST. Older adults with a 30s–CST ≤ 8 were classified as having reduced physical performance, whereas those with a 30s–CST > 8 were considered to have non-reduced physical performance. The cut-off point was chosen based on evidence that community-dwelling older adults scoring ≤8 in the 30s–CST are at risk of losing functional mobility. This cut-off point was deemed to have acceptable validity and reliability [13, 24].

The following self-reported information was collected in the process of evaluating potential predictors of persistent, reduced physical performance: age, gender, and mobility (climbing stairs and walking 400 m) [25]. Patients were asked if they had experienced difficulties with activities of daily living (ADL) within the last 2 weeks before the admission [26], if they had experienced falls, if they had participated in moderate physical activity (excluding ADLs) that was strenuous enough to increase work of breathing and pulse, how they perceived their health [27], and finally if they had used a walking aid before admission [28]. Additional potential predictors included the number of prescribed medications (taken from medical records) and physical performance as assessed by the 30s–CST and the De Morton Mobility Index (DEMMI) [29].

The 30s–CST assesses lower-body strength and has moderate inter-rater reliability for acute, admitted ‘medical’ patients. A floor effect at the time of admission makes the test only moderately feasible in an acute care setting, but on the other hand the simplicity of the test facilitates its use in a busy, short stay unit [15]. The 30s–CST was performed by counting the number of times in a 30 s interval that a patient can stand from a sitting position with their hands crossed against their chest [30]. A Minimum Importance Change (MIC) on 2.9–2.6 stands has been determined for the 30s–CST [31].

DEMMI assesses mobility and balance through 15 hierarchical items and provides a score between 0 and 100 [29]. DEMMI is a valid and reliable measurement of these parameters for both hospitalised and community-dwelling older adults [14, 32‐34]. A Minimal Detectable Change MDC₉₀ of 9.0 and a Minimal Clinically Important Difference (MCID) of 10.0 has been determined for DEMMI [32].

Information on living arrangement, education, acute diagnosis, destination after ED (home or another department), and contact with social services before hospitalisation was collected either as self-reported information or from medical records, and used as demographic factors. Cognitive performance was tested using the Orientation–Memory–Concentration Test (OMC) [35].

On weekday mornings, a physiotherapist recruited and tested patients for eligibility. Included patients provided written consent for study enrolment. In the 30s–CST assessment, patients who were unable to stand with their hands crossed against their chest scored 0. Patients who completed the task in a practice test, but were unable to stand in the actual test scored 1. To avoid fatigue after the 30s–CST test, we collected self-reported information before testing patients with the DEMMI. The DEMMI protocol was followed, except for the ‘sit to stand no arms’ (DEMMI item 6), as this had been demonstrated in the 30s–CST. After data collection, there was no further contact between the patient and the physiotherapist. The health staff had no access to study data and treatment was unaffected by study participation.

To inoculate post-discharge physical performance assessments from bias, a second physiotherapist, who did not perform the initial assessment, was selected to perform the follow-up assessment. If the patient was unable to participate at the originally scheduled post-hospital assessment then a later visit was scheduled soon thereafter. At the follow-up assessment, the 30s–CST and DEMMI were conducted with a ten-minute break between tests.

The sample size was calculated on the following assumption: for a multivariate analysis of potential predictors n = 50 + 8×, where x is the number of independent variables [36]. Since we anticipated a 20% dropout rate, a total of 156 patients were required as a precondition to including 10 potential predictors.

Potential predictors were classified into the following five domains: 1) demographic: age and gender; 2) self-reported mobility: walking 400 m, climbing a flight of stairs, walking aid use before admission, and falls; 3) self-reported habitual physical status: physical activity, self-rated health, and difficulties with ADLS in the 2 weeks before admission; 4) polypharmacy: number of prescribed medications; and 5) presenting physical performance: the 30s–CST and the DEMMI at the time of admission.

For the univariate logistic regression analysis, all continuous variables were dichotomised, with the exception of age, which was classified into 5 levels given the known association between age and physical performance. Cut-off points for continuous variables were based on Receiver Operating Characteristics (ROC) for the study data and on a literature review. In the literature we found a relationship between the ability to rise a maximum of five times and the risk of sarcopenia [37] and polypharmacy, as defined by ≥10 drugs associated with physical performance [38]. We found no recommended cut-off points for DEMMI. However, for semi-independent community-dwelling seniors, a score of 76.5 (95% CI 73.1–79.9) had previously been reported [39]. The ROC analysis revealed cut-off points at 30s–CST = 5, polypharmacy = 16 and DEMMI = 57 (see Additional file 1). We used the cut-off points found in the literature, except for DEMMI, for which the ROC cut-off 57 was used on account of the fact that acutely hospitalised older patients have lower physical performance than community-dwelling older adults [40]. Factors on ordinal scales were dichotomised (without difficulty or with difficulty/not at all).

For the multivariate analysis, age and gender were preselected [13, 17, 39‐42]. The smallest numbers of events determined the permitted number of predictors [43]. Potential predictors were included in the multivariate analysis using the following data reduction: 1) potential predictors with a p value ≤0.20 in the univariate analysis were considered [44]; 2) if the predictors within a domain had a moderate (>0.50) correlation, the potential predictor with the highest odds ratio was selected; 3) the final selection of predictors was based on the odds ratios and the assumed ease of use in an ED setting.

The potential predictors were tested for interaction. The area under the curve (AUC) was used to identify the final model and the model was tested with Hosmer–Lemeshow and for internal validity by bootstrapping [45].

Beta coefficients were employed to calculate the total score. The prediction model’s performance was assessed by calculating the sensitivity/specificity and predictive values for older adults with continuous reduced physical performance upon follow-up. Moreover, we identified the number needed to treat/test (NNT).

Analyses were performed using STATA 14 (Stata Statistical Software, College Station, TX) in adherence with principles outlined in the guidelines for Strengthening the Reporting of Observational Studies in Epidemiology [46] and Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis [47].

The Regional Scientific Ethical Committees of Southern Denmark approved this study with a waiver (20.08.2014). As required by Danish legislation, written informed consent was obtained from participants to permit collection of information from medical records. The project was registered with the Danish Data Protection Agency (2008–58-0035) and in the ClinicalTrials.​gov Identifier: NCT02474277 (12.10.2014).

Altogether, 78 (67%) of the patients improved their 30s–CSTs from admission to follow-up, 35 (30%) had an unchanged 30s–CST, and 4 (3%) had a lower 30s–CST. Although most patients improved from admission to follow-up, 76 (65%) of patients demonstrated persisting reduced physical performance (30s–CST ≤ 8).

More than half of patients had a 30s–CST improvement of 5 (IQR 3–7.3). The improvement was substantial for a sub-set of 13 patients (19%): their 30s–CST was 0 at admission and 11 at follow-up (IQR 10–12).

For DEMMI, 88 (75%) of the patients demonstrated improvements, whereas 16 (14%) deteriorated. The median improvement was 18.5 points (IQR 10.3–32.5).

Univariate analysis revealed 10 potential predictors with a p value ≤0.20; these were selected for further analysis (Table 2). The correlation was >0.50 or in other words of moderate strength, within the study domains of self-reported physical performance and presenting physical performance (see Additional file 2). This left six potential predictor variables for further model development, namely climbing stairs, physical activity, self-rated health, walking aid use, polypharmacy, and the 30s–CST, in addition to the preselected variables of age and gender.

The final selection of predictors, based on the odds ratio and the anticipated applicability and feasibility of use in the ED setting, narrowed down potential predictors to walking aid use before hospitalisation (OR: 7.1) and the 30s–CST ≤ 5 (OR: 9.1).

No significant interactions were found between potential predictors and the outcome measurement. The AUC for the full model was 0.80 (95% CI: 0.72; 0.89). The multivariate analyses showed that walking aid use before hospitalisation had an OR of 4.4 and that a 30s–CST ≤ 5 had an OR of 5.8 (Table 2).

Table 3 presents the selected predictors and their beta coefficients. In this sample, a score > 1.8 upon admission was able to identify 78% of patients who continued to have a reduced physical performance 1 month after acute hospitalisation. Furthermore, using a score of >1.8 only 2.43 patients were needed to identify one patient with reduced physical performance at follow-up (number needed to test).

Our finding, that a majority of patients improved their physical performance from the time of admission to 1 month after admission, corroborates the findings from earlier studies that used the Short Physical Performance Battery (SPPB) and walking speed [17, 18, 22]. In our study 65% of patients showed reduced physical performance 1 month after admission, reinforcing the need to provide this group with targeted interventions, since frailty is associated with a loss of independence, increased community costs, and readmission [13, 18, 48].

The univariate logistic regression revealed ten potential predictors for reduced physical performance (p value ≤0.20). Besides the preselected variables of age and gender, the event rate allowed two potential predictors to be included in the multivariate analysis. We selected use of walking aid before hospitalisation and a 30s–CST ≤ 5, as they had the highest odds ratio and were judged the most feasible tools to use in a busy ED setting. Moreover, using walking aids as a predictor makes clinical sense, since community-dwelling older adults use walking aids to improve balance and mobility [49]. On the other hand, walking aids are risk factors for low mobility [50] and their use before hospitalisation thus implies physical limitations and a higher risk of losing physical ability. Walking aids were also included in Hoogerduijn et al.’s model for assessing the risk of functional decline in acutely hospitalised older adults [51]. The other predictors in that study were a preadmission need for assistance in instrumental activities of daily living, a need for assistance in travelling, and a lack of education after age 14 [51].

We found that gender, age, self-reported information on walking aid use, and a 30s–CST ≤ 5, correctly identified patients who had continued reduced physical performance following acute hospitalisation. Moreover, a score > 1.8 identified 78% of patients with continuous reduced physical performance with a NTT of 2.43 patients. Clinically, all predictors need to be considered, since in isolation none of the model’s variables have a score > 1.8. Our prediction model based on physical measures and self-reported information is the first of its kind. However, a study in primary care settings concerning community-dwelling older adults aged 65 or older has shown that for older adults with poor health the combination of physical performance measures and self-reported information is substantially better than either alone [52].

Existing screening tools to identify older adults who need a comprehensive geriatric assessment (CGA) have shown poor reliability in an acute setting [9]. Our prediction model supports the identification of older adults who could benefit from a CGA, where a functional assessment is an integral part [53]. For patients discharged to other units than the geriatrics unit, the identification of older adults with persistent reduced physical performance might give rise to a targeted hospital intervention such as routine patient mobilisation. Furthermore, this study supports the evidence from other studies that self-reported information and physical performance measures provide different and complementary information [8]. From admission to follow-up, 19% of the patients had a 30s–CST change from 0 to 11. Hence, if the prediction model had been solely based on physical performance then 19% of patients would have been misclassified. For every 2–3 patients tested, clinicians will identify one patient with reduced physical performance 1 month after hospitalisation. However, since the negative predictive value is only 63%, every third with a negative test will still be at risk (Table 3). The prediction model does not comprehensively identify all at-risk patients, which the clinicians should be aware of. Clinically, this prediction model is easily applied: age and gender are known, determining walking aid use before hospitalisation requires one simple question, and the 30s–CST is easy to execute.

The study strength lies in its ability to assess physical performance using a simple objective measurement in combination with self-reported information. We used the 30s–CST in the prediction model while recognising that the floor effect could affect the baseline assessment. This choice was related to the well-known improvement in physical performance measures from admission to discharge [17, 18, 22].

Although up to 48 h was permitted from the time of admission to the time of baseline assessment, in practice the timeframe was much shorter as assessments were performed routinely every weekday morning. It follows that the prediction model was less influenced by the cause of hospitalisation.

The cohort included patients discharged from the short stay unit as well as patients transferred to other wards; thus a different risk for deterioration due to varied length of stay. However, the number of patients with reduced physical performance at follow-up in both groups (discharged from short stay unit or transferred to other wards) was comparable. This lack of difference in deterioration can be explained by the tiredness older adults generally experience after an acute admission [53].

The 30s–CST was used as an outcome measure even though the cut-off point of ≤8 for the 30s–CST is only validated for use in active, community-dwelling, older adults. We did so since the follow-up visit was performed in the older adult’s home.

The binary stratification of the outcome measure might have resulted in a misclassification of some patients, due to the variation in patient performance [54]. We chose this dichotomisation since it is used in current literature [13, 24] and since it reflects recommendations made in Denmark and elsewhere for screening programs for community-dwelling, older adults.

We have described the predictor selection in detail, making the selection process easily reproducible in other settings. We managed to reach our pre-calculated sample size, but the study is weakened by a lower event rate than expected, which in turn restricted the number of predictors that were included in the model. Thus, before clinical implementation we recommend that the model’s external validity is verified through larger studies using a different population. Moreover, the prediction model can only be generalised to older ‘medical’ patients who are mentally fit and show reduced physical performance upon admission.

Patients who were not assessed for eligibility can be seen as introducing a selection bias. However, 55% of patients were excluded based on organisational limitations, such as the day of admission, since patients were only recruited on weekdays. Of note, patients admitted on Sundays were included if they fell within the 48-h limit for enrolment. Patients who refused to participate generally offered two reasons; either they felt the project was irrelevant to them or they did not have the energy to participate.

The follow-up visits were completed at a median of 34 days (IQR 27–40) after admission, although the initial intention was to perform follow-up 14 days after admission. Delays in the follow-up assessment were due to patient preference, patient schedules, and the fact that some patients had not been discharged at the time of planned follow-up. We assume that the delay in follow-up was beneficial for this particular study, since it can be assumed that physical performance would have stabilised over a longer interval of time.