Participants were recruited from general medical wards at The Northern Hospital, Victoria, Australia. Consecutive participants were screened by a recruiting officer and were eligible to participate if 65 years or older and were assessed within 48 hours of admission. Patients were excluded if they had a planned hospital stay of less than 48 hours, severe dysphasia, documented contra-indications to mobilization, were isolated for infection, or if death was imminent. All eligible participants were invited to participate. Consent was obtained within 48 hours of hospital admission. For patients deemed incompetent to consent, this was obtained from the 'person responsible' or next-of-kin. Interpreters were employed when required.

Participants were assessed at the bedside every 48 hours during hospitalisation or on the Monday following a weekend. Baseline measurements included age, sex, place of residence prior to admission, primary language, gait aid use prior to hospitalisation, Mini Mental State Examination (MMSE) 18, Charlson Comorbidity Index 19, APACHE11 Severity of Illness Scale 20, the Barthel Index (BI) 821, Hierarchical Assessment of Balance and Mobility (HABAM) 5 and the new mobility items. The BI and HABAM were selected for a head-to-head comparison with the new mobility instrument. The BI is widely used as a self report measure of independence in activities of daily living in the acute hospital setting 11 and, prior to this study, the HABAM was identified as having the most desirable properties of existing mobility instruments 3. Each of these outcome measures are described in further detail below.

At each assessment a researcher administered the BI and the MMSE. As close as possible to this assessment, the patient was assessed on the mobility items by the principal researcher, who was blind to BI scores. The HABAM items were a subset of these mobility items.

Mobility items were administered in the order of bed, chair, balance and walking activities to maximise patient safety, confidence and ease of testing. Familiarisation trials were not provided to minimise fatigue and time required to administer the test. At each test the therapist and patient independently rated the patient's current mobility compared with admission mobility on a 5 point scale (much worse, bit worse, same, bit better, much better). This provided a reference standard for important change in mobility.

The APACHE 11 is a severity of illness scale with a score range from 0 to 71, where higher scores represent increasing severity of illness during the first 24 hours of hospital admission. The Charlson Index classifies comorbid conditions according to risk of mortality. One year mortality rates in a medical population have been reported to be 12%, 26%, 52% and 85% for Charlson scores of 0, 1–2, 3–4 and greater than 5 respectively 19.

The modified BI is an ordinal scale that provides a total score between 0 and 100 where higher scores indicate greater independence in activities of daily living 21. The HABAM is an interval level mobility instrument that provides a score between 0 and 26 5 where higher scores indicate increasing levels of independent mobility and was designed for application in an older acute medical population. The MMSE is reported to be a valid and reliable measure of patient cognition 18. It provides a score between 0 and 30 points where increasing scores indicate better cognitive ability.

The complete set of 51 mobility items were pilot tested for two weeks to remove items with practical limitations, a process that included patient and assessor interview about the mobility tests. The remaining 42 items were then tested on a large sample by the principal researcher. After completion, items with practical limitations were removed and Rasch analysis conducted.

Data analyses were performed using SPSS version 12.0 22 and RUMM2020 23. The Rasch partial credit model was employed to identify misfitting and redundant items and to identify a hierarchy of mobility items ranked from easiest to hardest. Participants were divided into 3 class intervals (ie, 3 groups of patients at different levels of mobility). Item misfit was considered if the chi-square or F statistic probability value was less than the Bonferroni-adjusted a value for multiple testing or the fit residuals were greater than ± 2.

Item residuals from Rasch analysis were also examined as a finding of no association between residuals for individual items has been argued as evidence of local item independence 24. High positive correlation between residuals provides evidence of local item dependence and high negative correlations is thought to indicate multidimensionality.

Differential item functioning (DIF) analysis 25 was planned for age, gender, time of assessment, cognitive status (MMSE) and whether an interpreter was required. DIF was considered significant if the chi-square probability value was lower than the Bonferroni-adjusted p value. A priori, these factors were considered potential confounders to item functioning.

Item response thresholds were also studied to investigate the existence of disordered thresholds, that is, response patterns on the rating scale that are not in the expected order. The person separation index (PSI) was reported to provide an indication of the internal consistency (reliability) of the scale by examining the ability of the instrument to discriminate among respondents.

Sample size for Rasch analysis was based on recommendations by Linacre et al 26. These authors recommend a sample size of 64 – 144 to provide 95% confidence +/- 0.5 logits. Baseline and 48 hour assessments during a 3–4 month period were expected to provide more than 200 assessments. In the absence of DIF by time, all available assessments would be included for Rasch analysis as recommended by Wright 27 and Chang and Chang 28.

An inter-rater reliability study was conducted on a subset of patients who reported no fatigue from the first assessment. After the first assessment and a 10 minute rest, the mobility assessment was repeated by a physiotherapist blind to the outcomes of the first test. Test order of assessing physiotherapists was randomised. Power calculations were performed based on recommendations by Walter et al 29. The Minimal Detectable Change at 90% confidence (MDC₉₀) and accompanying 95% confidence intervals were estimated 30.

Correlation coefficients and associated 95% confidence intervals were calculated to investigate the convergent validity of DEMMI scores with the BI (a measure of a related construct) and HABAM (a measure of the same construct), and discriminant validity with the MMSE, Charlson Index and APACHE 11 (measures of different constructs). To investigate known-groups validity, an independent t test was performed on DEMMI scores of patients discharged to home compared to inpatient rehabilitation.

The MCID was calculated for DEMMI, HABAM and BI as the mean change score for patients who rated themselves 'much better' at discharge (criterion based method). The MCID was also calculated using distribution based method recommended by Norman et al31.

The Effect Size Index (distribution method)(ESI) and Guyatt's Responsiveness Index (criterion method)(GRI), were selected a priori to calculate measurement responsiveness of the DEMMI, HABAM and BI. Inferential 95% confidence bands were calculated to enable statistical comparison of responsiveness estimates as recommended by Tryon 32.

The time required to administer the DEMMI was rounded to the nearest 30 seconds and was recorded using a stop watch.

Pilot testing on 15 consecutive older general medical patients identified 9 items for removal based on practical limitations (Table 1).

Figure 2 shows that of the 388 new hospital admissions screened for inclusion, 219 were eligible, 104 were recruited and 89 performed at least one mobility assessment. Three patients were readmitted during the study period and were included twice as new hospital admissions. Table 2 shows the admission characteristics for the 86 patients included in this study. There were no adverse events as a result of the mobility assessments. A further 8 items were removed due to practical limitations that were identified following further testing and the 4 walking items were rescored to 2 items to limit local item dependence (an assumption of Rasch analysis)(Table 1).

Following item testing and Rasch analysis, 32 items were reduced to 17 (Table 1). DIF by time was not identified for the 17 items and therefore Rasch analysis was performed on data from hospital admission and subsequent 48 hour assessments. Rescoring three items (lie to sit, sit to stand and walking distance) produced ordered thresholds for all items.

Data for the 17 mobility items fitted the Rasch model (item-trait χ² = 41.17, df = 34, p = 0.19). The t test procedure 2433 identified that the percentage of individual t tests outside the acceptable range was only 4.23%. (95% CI 1.0% to 7.0%). This provides further evidence of the unidimensionality of the 17 mobility items.

Examination of the residual correlation matrix indicated negative correlations of greater than 0.3 between sit unsupported and bridge (r = -0.55), standing on toes and stand on one leg eyes closed (r = -0.58) and tandem standing eyes closed and walking distance (r = 0.35). However, these findings were not supported by high fit residuals for any of these items. A positive correlation of greater than 0.30 was only identified between the roll and lie to sit (r = +0.37) items. Although this result indicates the possibility of some response dependency between these mobility tasks, both items were retained as each provides important clinical information regarding patient mobility and care needs during acute hospitalisation. In addition, examination of the admission only dataset indicated a lower correlation of +0.21.

Person separation was 0.92, indicating the test could discriminate 5.8 strata of ability.

Correlation between independent assessor DEMMI interval scores was high (Pearson's r = 0.94, 95% CI 0.86 to 0.98). The mean scores for assessors 1 and 2 were 57.19 (sd = 17.07) and 55.05 (sd = 13.77) points respectively. A paired t test indicated no systematic differences between assessors (p = 0.14). Using a pooled standard deviation of 15.51, the standard error of measurement (SEM) was 4.10 and the inter-rater reliability MDC₉₀ was 9.51 points (95% CI 5.04 to 13.32) on the 100 point DEMMI interval scale. This indicates that a patient needs to improve or deteriorate by 10 points or more for a clinician to be 90% confident that a true change in patient condition has occurred. A paired t test indicated no systematic difference between the first and second assessment scores (p = 0.77).

DEMMI scores had a significant and high correlation with HABAM and BI scores. This provides evidence of convergent validity for the DEMMI.

Discriminant validity for the DEMMI was evidenced by a low correlation with measures of other constructs (MMSE, APACHE 11 severity of illness and Charlson co-morbidity index scores).

An independent t test showed that patients who were discharged to inpatient rehabilitation had significantly lower DEMMI scores at acute hospital discharge than those discharged to home. Patients discharged to inpatient rehabilitation had a mean DEMMI score of 39.55 (sd = 9.41, 95% CI 33.72 to 45.38) and patients discharged to home had a mean DEMMI score of 59.61 (sd = 13.22, 95% CI 56.30 to 62.93). This provides evidence of known groups validity for the DEMMI.

There was no significant difference identified between the responsiveness of DEMMI and HABAM measurements or DEMMI and BI measurements using the ESI or GRI based on patient or therapist report of change.

By calculating the average change in DEMMI score for patients who reported to be 'much better' in their mobility between hospital admission and discharge, the MCID for the DEMMI was identified to be 8 points, that is, a change of 8 points or more is likely to represent a patient perceived important change in mobility. Using Norman et al.'s 31 distribution based method, the MCID was also calculated to be 8 points for the DEMMI.

Feedback from 15 clinicians was obtained following their administration of the DEMMI. Minor changes were made to the sit unsupported item and testing protocol and the final format of the DEMMI.

Figure 3 shows that of 344 new hospital admissions screened, 216 were eligible, 132 were recruited and 112 performed at least one mobility assessment. Six patients were readmitted during the study period and were included twice as new hospital admissions. Another six patients did not complete a hospital admission assessment. Table 2 shows the admission characteristics for the 106 patients included in this study. A total of 312 mobility assessments were performed using the 17 mobility items. Patients in the validation study did not differ from the instrument development sample on any baseline characteristic.

Prior to conducting Rasch analysis the jog item was removed. This item required clinical experience of medical conditions to determine whether testing should proceed. No participant was able to successfully complete the standing on one leg with eyes closed item in the validation study. Rasch analysis was therefore performed for the remaining 15 items.

In the validation study, the pooled dataset showed misfit to the Rasch model due to large sample size as there was no evidence of DIF by time or multidimensionality. Using the t test procedure 2433, multidimensionality was not identified. Four items (reaching for pen, backward walking, standing on toes and sit to stand no arms) had a positive correlation of 0.3 or greater and three items (walking distance, roll and lie-sit) had a negative correlation of 0.3 or greater with the first residual component. The t test procedure indicated the percentage of individual t tests outside the acceptable range was 4.88% (95% CI -2.0% to 7.0%). This provides further evidence of the unidimensionality of the 15 DEMMI items and therefore does not explain the misfit of the data to the Rasch model. No evidence of local item dependency was identified as there was an absence of correlations in the residuals above a magnitude of 0.3.

Data fitted the model at each assessment time point; baseline (χ² = 24.60, df = 30, p = 0.74), first 48 hour assessment (χ² = 36.37, df = 30, p = 0.20) and subsequent 48 hour assessments (χ² = 36.26, df = 28, p = 0.14). Given the similar findings across samples, analysis of hospital admission data is reported.

There were 106 hospital admission mobility assessments performed. The mobility items were well targeted for older acute medical patients. Figure 4 shows the average item difficulty and the person ability locations on the logit scale. There were only a few persons with a lower ability than the easiest item (to the left of the scale), or with a higher ability than the hardest item (to the right of the scale). None of the items showed misfit to the model and χ² and F statistic Bonferroni adjusted probability values for the 15 items were non significant. Person separation was 0.88, indicating the test could reliably identify 3.7 strata of ability.

Three items showed significant DIF by age but appear to be statistical artifacts due to a small number of participants in one of the three class intervals. Two items (lie to sit and walking independence) showed mild disordering of one threshold. However, inspection of item thresholds in the pooled dataset showed these items to be ordered and they were not rescored.

Figure 5 shows the item hierarchy for the 15 DEMMI items was consistent across independent samples. The final DEMMI is shown in Additional file 1 and its clinimetric properties reported in Table 3. The measurement properties of the DEMMI (Rasch, reliability, validity and MCID) were consistent with estimates obtained from the instrument development sample.