Clinical results

Recruitment

Forty-one NHS Trusts screened 1551 adults with a traumatic shoulder dislocation, from whom 1069 were not randomised. Between 14 November 2018 and 14 March 2022 we randomised 482 of into the study, 101% of target (Figure 1). One Trust was unable to randomise any participants prior to closure of the trial.

FIGURE 1 CONSORT flow diagram. (Figure not reproduced here; see the original article.)

Baseline characteristics

Two-thirds of participants were male (n = 317/482; 66%), 36% of participants had a sport-related injury (n = 172/482) for their dislocated shoulder and 16% reported a concurrent injury (n = 96). The mean age of participants was 45 years (SD = 20) (Table 5).

The group allocated to additional physiotherapy had a slightly higher baseline score than the ARTISAN-only group. Figure 2 shows this graphically, where it can also be seen that while the distributions of the two groups are similar, the overwhelming majority of participants have increased from their baseline OSIS score. It is also clear that participants with lower OSIS scores at baseline achieve larger increases in function partly due to regression to the mean.

FIGURE 2 Oxford Shoulder Instability Score waterfall plot. Participants are ranked by baseline score (dot) and are connected to their 6-month score by a line. Line colour represents allocation group and line type (solid or dashed) represents age group at randomisation. (Figure not reproduced here; see the original article.)

Age cut point

Figure 3 shows the probability density plots of the participant age. Two distributions were fitted, centred at 24 and 56 years old, with the point where probability of membership in both distributions was 50% at 33 years of age. As this was within 10 years of the randomisation stratum (age of 40), no further sensitivity analyses exploring the effects of age were done.

FIGURE 3 Probability density plots of the participant age. (Figure not reproduced here; see the original article.)

Intervention delivery

During the ARTISAN study, 96 therapists delivered the core ARTISAN session across the 41 study sites. The number of physiotherapists at each site ranged from one to seven, with a low number of physiotherapists randomising and treating more than three participants in the study.

Compliance

Source physiotherapy data were requested from sites for all 482 participants; only four records were unobtainable, meaning data were available for 478 participants. Table 6 shows the number of participants attending additional physiotherapy sessions. Of those receiving the ARTISAN-only intervention, 15% had extra physiotherapy, whereas 83% of participants in the group allocated to additional physiotherapy received at least one additional physio session.

Table 7 also shows patients’ clinical treatment status at 6 months in both groups.

In the ARTISAN-only group, 42 participants opted to have extra physio sessions, 194 received the ARTISAN-only session and 2 participants were referred to physiotherapy by clinicians. These two participants’ treatments were not compliant with the intervention protocol and they were considered to have switched allocation groups for the per-protocol analysis. There were two participants whose data were missing.

In the group allocated to ARTISAN plus additional physiotherapy, 24 participants were confirmed by site records not to have attended any physiotherapy appointments. Seventeen participants attended an additional physiotherapy session but were scheduled to have more and did not attend these sessions. The remaining 185 participants either had been discharged (i.e. completed the physiotherapy course) or were still receiving ongoing treatment. Finally, there were two participants for whom data could not be retrieved in this group.

The treatment patient pathways are demonstrated in the participant treatment flow diagrams for each respective allocation group in Figures 4 and 5.

FIGURE 4 Diagram of participant pathway for the ARTISAN-only group. a Participants are compliant with protocol. b Participants are not compliant with protocol (i.e. have switched treatment groups). (Figure not reproduced here; see the original article.)

FIGURE 5 Diagram of physiotherapy pathways for the ARTISAN Plus group. (Figure not reproduced here; see the original article.)

Therapist effect

As the study is focused on the effectiveness of physiotherapy, there were concerns at the beginning of the study that there may be a physiotherapist effect which could strongly affect the estimate of efficacy of the ARTISAN intervention. Hence, the study captured the number of therapists who delivered the core (pre-randomisation) session.

There were 96 therapists who delivered the core ARTISAN session across the 41 study sites. The median number of therapists at each site was 2 (range 1–7). The number of participants randomised by each therapist ranged from 1 to 21; of the 96 therapists, only 42 (44%) randomised more than three participants.

At the interim analysis point, we estimated the therapist effect by calculating the ICC using the 3-month follow-up data with a multilevel model. A chi-squared test comparing the likelihood of the full model against the model without the physiotherapist effect was also conducted. The ICC was estimated to be 0.0201 (95% CI 0 to 0.601) (participants n = 138; physiotherapists n = 67). The addition of physiotherapy effects did not improve the model (χ² = −2.27e-13, df = 1; p = 1).

Due to the low number of physiotherapists randomising more than three participants, the DMC asked for a sensitivity analysis with those physiotherapists randomising three or fewer participants removed. Again, the addition of therapist effects did not improve the model (χ² = 1.14e-13, df = 1; p = 0.5). The ICC for was estimated to be 0.0019 (95% CI 0 to 0.483) (participants n = 94; physiotherapists n = 23).

The DMC endorsed the recommendation that the sample size did not need to be inflated for therapist effects, but requested that the analysis was repeated using the 6-month follow-up data at the end of the study to check whether the interim analysis was correct. Again, 12 of the 41 sites had a single physiotherapist performing the randomisation of the participants into the ARTISAN study, and many physiotherapists still did not randomise more than three participants.

Repeating the interim analysis using the 6-month follow-up data showed that the physiotherapist effect was small and the ICC was not statistically significant. The model was adjusted for the dominant arm injured, age group and physiotherapist, but did not include the allocation group (354 participants and 80 physiotherapists). An ICC value of 0.026 was observed (95% CI 0 to 0.106). Again, comparing the models with and without the physiotherapy effects had χ² = 0.638 and p = 0.424. Again, this showed that including physiotherapy effects did not improve the model fit.

Physiotherapy-delivered intervention content

Details of the core (pre-randomisation) ARTISAN session are given in Table 8. A total of 460 participants were given the core session as per protocol; 22 participants had details missing. The mean times taken to deliver the ARTISAN session were similar in both interventions.

Table 8 also contains the self-reported grade of the physiotherapists who delivered the additional physiotherapy sessions.

Tables 9–11 show the physiotherapy exercises prescribed at home and during the session in the additional physiotherapy group.

Follow-up

Follow-up at 6 months post randomisation was less than expected, with 354/482 (73%) returning their questionnaires. Comparing the baseline demographics of the two groups, participants returning their questionnaires were more likely to be female (n = 138/354, 39% returned vs. n = 27/128, 21% lost to follow-up), aged 40 and over (n = 227/354, 64% returned vs. n = 34/128, 27% lost to follow-up) and to have had their dislocation via a non-sporting-related incident (n = 246/354, 70% returned vs. n = 64/128, 50% lost to follow-up). However, mean baseline OSIS scores were similar (mean scores: 19.1 returned vs. 19.1 lost) and there was no correlation with whether or not the participant’s dominant arm was injured (n = 199/354, 56% returned vs. n = 71/128, 55% lost).

Despite a larger than anticipated loss to follow-up, the baseline characteristics predominantly remained similar between intervention groups. For example, the percentage of the younger population (39 years and under) who returned their questionnaires was 36% in both intervention groups; similarly for gender, females comprised 39% of those who returned questionnaires in the ARTISAN only group and 40% in the ARTISAN Plus group. The imbalance between the interventions for mechanism of injury also remained the same as the baseline figures. This provides reassurance that the mechanism of missingness was not related to the trial interventions or processes. Therefore, we do not expect a large risk of bias due to the participants lost to follow-up.

Primary outcome

The primary outcome was the OSIS. The possible scores range from 0 (worst function) to 48 (best function). The trial was designed to detect a difference of 4 points between allocation groups at 6 months post randomisation. Descriptive statistics for the OSIS at all time points are shown in Table 12 and Figure 6.

FIGURE 6 Oxford Shoulder Instability Score box plot: median and interquartile range (IQR), min and max represented by line. (Figure not reproduced here; see the original article.)

The mean difference (unadjusted) at the 6-month follow-up showed that the ARTISAN Plus group had a better outcome and was statistically significant. The ARTISAN Plus group had a positive mean difference of 2.2 (95% CI 0.1 to 4.3). The point estimate was just over half of our target difference of 4.0.

The adjusted intention-to-treat model was the primary analysis for the study. A mixed-effects model was fitted with the following (pre-specified) variables as fixed effects: allocation group, whether the participant’s dominant arm was injured, age group, and baseline score. Site of randomisation was fitted as a random effect. The model coefficients are reported in Table 13. The mean adjusted difference was 1.5 points (95% CI −0.3 to 3.5) in favour of the ARTISAN Plus group. This shows that there was no significant difference between the intervention groups. The limits of the 95% CI excluded our target difference of 4.0. Age group was found to be significant, with participants aged 40 years and over performing worse by −2.7 points. The reason why the unadjusted result was different to the adjusted result was that the baseline score was a significant predictor of the participants’ outcome. Since this was not accounted for in the unadjusted analysis, the slight imbalance in scores at baseline caused a different result to that of the adjusted model.

Secondary outcomes

QuickDASH

Table 14 and Figure 7show the results of the QuickDASH outcomes at each time point. The results at the 6-month primary end point show a mean difference of −1.7 (95% CI −5.4 to 2.0) in favour of the ARTISAN Plus group; therefore the unadjusted results show that the effect of the intervention is not statistically significant.

FIGURE 7 QuickDASH score box plot: median and IQR, min and max represented by line. (Figure not reproduced here; see the original article.)

QuickDASH adjusted model results

Table 15 presents the adjusted results for the QuickDASH at the primary end point of 6 months. The results also agree with the previous unadjusted result; there was no significant difference observed between the interventions. However, the ARTISAN Plus group had slightly higher levels of disability than the ARTISAN-only group. Participant age group was a statistically significant variable, with the older age group having 7.5 points more disability than the younger group. The models for the early end points also followed this pattern.

Quality of life (EuroQol 5-dimension score)

Table 16 and Figure 8 show the unadjusted results for the EQ-5D-5L at each time point.

FIGURE 8 Quality-of-life EQ-5D-5L box plot: median and IQR, min and max represented by line. (Figure not reproduced here; see the original article.)

EuroQol 5-dimension score adjusted model results

Initially, the adjusted model did not converge. However, replacing the site random effect as a fixed effect showed that the outcome at 6 months did not differ significantly between the intervention groups. The other follow-up points also showed no significant between-group differences (Table 17).

Complications and serious adverse events

Table 18 reports the number of complications (pre-specified AEs and SAEs) reported by participants at any follow-up. There were 27 reported SAEs during the trial that were unrelated to the TASD or TASD treatment, 13 in the ARTISAN-only group and 14 in the ARTISAN Plus additional sessions group (Table 19).

Temporal effects

It was pre-specified that temporal effects would be assessed to observe whether participants in one treatment group had faster recovery than those in the other.

The model results for each of the outcome measures (OSIS, QuickDASH and EQ-5D) had shown that time point is a significant predictor in determining the outcome of participants; in other words, people do make statistically significant improvements at each time point. However, when analysing the interactions, there were no significant results between the time points and treatment groups (no interaction term in any model crossed the significance threshold for the p-value of 0.05). This means that the recovery trajectories are non-significant between the allocation groups.

Sensitivity and subgroup analyses

Sensitivity analyses: predefined per-protocol analysis

There were two participants randomised to the ARTISAN-only allocation group for whom the treating clinician prescribed further physio sessions. These participants were included in the intention-to-treat primary analysis. Also, four participants were incorrectly randomised into the study as they had no radiological evidence of their shoulder dislocation.

To assess the effects of the treatment group switching and inclusion of ineligible participants, the treatment-switching participants were removed and those randomised in error were added to their randomisation groups. It was found that this did not affect the primary analysis model. The adjusted mean difference was 1.4 points for the ARTISAN Plus physio group (95% CI −0.5 to 3.3; p = 0.150).

Sensitivity analysis: post hoc ‘as treated’ analysis

During the analysis, it was noted that there were a large proportion of participants who did not attend any physiotherapy. Hence, we did a post hoc ‘as-treated’ analysis. In this scenario, we considered participants who received at least one physiotherapy session (as defined by the recruitment site), regardless of their initial allocation, to be part of the ‘additional physiotherapy’ programme. Participants who did not have any additional physiotherapy were considered as the ARTISAN-only arm.

We found that 238 participants had additional physiotherapy and 239 had no additional physiotherapy after randomisation. We fitted a random-effects model with the same covariates as in the primary analysis. The results in Table 20 show that the models were similar and there was no significant difference between the two intervention groups.

Sensitivity analysis: post hoc per protocol (physiotherapy received)

In this scenario, we only considered the intervention to be the receipt or not of additional physiotherapy sessions. That is, we compared those participants in the ARTISAN-only group who did not have a further physiotherapy session with participants in the ARTISAN Plus group who received at least one further session of physiotherapy.

Of the 478 participants who had their physiotherapy details confirmed by recruitment site, 200 participants allocated to the ARTISAN-only group received only the core ARTISAN session, and 201 participants in the additional physiotherapy arm received at least one additional physiotherapy visit. Repeating the primary analysis with these participants yielded a model which did not converge. However, including site as a fixed effect showed no significant differences between the two interventions, and broadly similar model coefficients, as shown in Table 21.

Imputation of missing data

As the study had a sizeable amount of missing follow-up data, the missing scores for the primary outcome were imputed using MICE in R. The results showed no significant differences in intervention effects. The mean difference was 0.9 (95% CI −0.8 to 2.6; p = 0.309) in favour of the ARTISAN Plus physiotherapy programme.

Pre-specified subgroups: dominant arm

The study pre-specified a subgroup analysis to explore whether there was evidence of differences in the intervention effects between participants who injured their dominant arm and those who injured the non-dominant arm. This was explored by adding an interaction term between the allocation group and dominant arm injury term in the intention-to-treat model (Table 22). The results show that there were no significant differences, nor was the model largely altered. The overall estimates of the treatment effects are shown in Figure 9.

FIGURE 9 Interaction plot between arm of injured shoulder and intervention for OSIS. (Figure not reproduced here; see the original article.)

Pre-specified subgroups: age group

The study also pre-specified a subgroup analysis to explore whether there was evidence of differences in the intervention effects between the two participant age groups. Again, this was explored by adding an interaction term between the allocation and age groups (Table 23 and Figure 10).

FIGURE 10 Interaction plot between age group and intervention for OSIS. (Figure not reproduced here; see the original article.)

The interaction model for age group showed that older participants who were in the ARTISAN Plus arm were overall better by 0.64 points compared to younger counterparts in the ARTISAN-only arm. This is not a statistically significant or clinically meaningful effect. The results for the primary outcome split by age group are also presented in Table 24.

Impact of coronavirus disease

A sensitivity analysis was planned to investigate whether the COVID pandemic had an impact on the ARTISAN trial patients who were recruited during the COVID pandemic where the trial and physiotherapy sessions were paused for a period of time.

The results for the primary outcome (OSIS at 6 months) showed that the participants recruited prior to COVID and those recruited after the trial had restarted had no difference in their outcomes. The mean OSIS score of pre-COVID patients was 36.8 and the mean score for patients after COVID was 37.8, a difference of 1 point in favour of the patients recruited post COVID (95% CI −1.0 to 3.1).

The results after adjusting for the effect of the interventions and the randomisation variables showed a difference of 0.7 points (95% CI −1.3 to 2.6; p = 0.490). As this is not statistically or clinically significant, we conclude that COVID did not impact on the outcome of the trial.