Study I & II
The required number of patients to include in the longitudinal study was based on a power calculation for Study I, assuming a correct classification rate (CCR) of 75%
with a CI 95% of 65- 85%. A CCR of 75% was chosen as this was in line with the accuracy found in the original PREP2 study.23 Allowing for a 20% drop-out, it was decided to include at least 90 patients.89
STATA 16 was used for data analysis. Data were visually inspected with histo-grams, boxplots, qq-plots and dotplots to determine the distribution of normality.
Continuous baseline characteristics, stroke details, baseline and follow-up scores were summarized by mean, standard deviation (SD), min, and max when normally distributed; otherwise by median, interquartile range (IQR), min, and max.
Demographic and clinical characteristics of the patients who were unavailable for the three-month follow-up were compared with those available to determine if the difference was statistically significant. The unpaired t-test or the Wilcoxon rank sum test was used for continuous data and the Chi2 test for dichotomous data.
Specific for Study I
Improvement in UL impairment on FMA and UL function on ARAT from baseline to follow-up was examined. As FMA and ARAT are ordinal scales and data were non-normally distributed, within-group difference on the two scales from inclusion to follow-up was tested with the nonparametric Wilcoxon signed rank test.
The overall accuracy of the PREP2 was quantified by comparing the agreement between predicted and achieved ARAT categories using the CCR.89 The CCR, along with sensitivity and specificity, were calculated for each of the four categories.
Also, CCR was calculated separately for patients with a SAFE score < 5 or ≥ 5 to differentiate between patients with either severe UL impairment at baseline, who had MEP status obtained, and patients with relatively mild UL impairment at base-line, who did not need to have MEP status obtained.
To examine if prediction accuracy of PREP2 obtained two weeks after stroke could be improved, a classification and regression tree (CART) analysis was carried out.89 CART analysis produces a decision tree without the user determining which vari-ables to include or their order in the tree.94,95 The CART analysis was based on the components of PREP2: SAFE score, age, NIHSS score, and MEP status. For patients with a SAFE ≥ 5, MEP+ status was assumed in the analysis.
Specific for Study II
Accelerometer data were displayed for the whole group and in line with a recent study also in three categories, each reflecting a range of scores on FMA at base-line.62 The category "Severe" comprised the FMA scores of 0-22, "Moderate" 23-50, and the category "Mild" the scores 51-66.62
Prediction of use ratio
Several regression models were created. The first model, Model 1, was a linear regression model to assess the strength of the (unadjusted) association between baseline FMA score and UL use ratio at three months. In Model 2, a multiple regression model, the association between FMA at baseline and use ratio was adjusted for other secondary variables chosen a priori, based either on the results of previous studies or clinical reasoning. The independent variables and their dis-tribution were assessed (their dispersion, frequency disdis-tributions). Moreover, the relationship between the independent variables, one at a time, was assessed.
Secondary variables chosen a priori were: MEP status (MEP present/ not pres-ent). Neglect (dichotomized into present/ not prespres-ent). Dominant UL affected was included as previous research has demonstrated that dominant side affected may result in better UL stroke recovery.18,61,96 Twopd (affected/ not affected), as previous research has shown this was a predictor for future UL function.82 The FIM score, reflecting the need for assistance in daily life activities, was entered as a continuous variable from 18 - 126 (max). Gender, as older women use their dominant hand more in daily life compared with older men.97 Severity of pain, a continuous score of 0 - 10.
In Model 3, the contribution of the biomarker MEP was assessed by removing MEP status from model 2 and comparing the fit of the model with and without MEP. Furthermore, the contributions of the individual predictive variables were examined. Finally, to assess the strength of each potential predictor univariate regression between each of the predictor variables and use ratio was performed.
All necessary assumptions for generalized linear models, including linearity, equal-ity of variance, and normalequal-ity of errors were visually inspected for all models and found adequate. Presence of multi-linearity was examined by the Variance Infla-tion Factor for each independent variable. Using a conservative approach, VIF below 3 were accepted.98 Multi-linearity was not present.
The ability of the models to predict use ratio was assessed by the size of the ad-justed R2. The contribution of each individual predictor in the model was assessed from the significance level, size of p-value and the size of the β-coefficient includ-ing the 95% CI.99
To assess the ability of the models to predict future use ratio for an individual pa-tient, the 95% prediction interval (PI) for the regression line was calculated based on the SD for the adjusted R2 (PI = ± 1.96 * SD). The PI is an estimate of the in-terval in which a future observation of UL use ratio will fall, with 95% probability, given what has already been observed.
Normal and non-normal use ratio
Use ratio was dichotomized into normal and non-normal using a threshold based on an established reference value from a study with 74 community-dwelling adults.63 In the reference population the mean use ratio was 0.95± SD 0.06, range 0.79-1.1.63 In the present study the lower limit of the PI interval for the reference value was calculated (0.95- 1.96* 0.06=0.83) and used as a conservative threshold for normal use ratio. According to this, patients with a use ratio above or equal to 0.83 were classified as having a normal use ratio, and patients with a use ratio below 0.83 as having a non-normal use ratio.
The association between the use ratio and each of the variables FMA, MEP status, neglect, dominant UL affected, twopd and FIM were visually inspected followed by a multivariate logistic regression. To maintain adequate power for the statistical analysis the events per variable rule, which calls for at least ten outcomes for each variable in the regression model, was compiled with.100,101 A receiver-operating curve (ROC) of the logistic model was graphically displayed, and a two-way con-tingence table was used to identify the cut point with the highest sensitivity and specificity values.
Study III
The interview transcripts were imported to the qualitative research software pro-gram NVivo12. The pilot focus group interview was considered to add interesting aspects to the topic and data from this interview were analysed along with data from the succeeding three interviews.
A thematic content analysis of the interviews was performed.91,102 The analysis was both a deductive and an inductive process.91,102 Deductive, as the CFIR frame-work was used as the aim was to answer specified pre-defined question regarding barriers and facilitators (theory-based coding). Inductive, as to let the material talk because attitudes towards UL prediction algorithms have not previously been explored, and knowledge of how to implement algorithms into the clinic setting is scarce (data-based coding). First, meaning units were identified and the four inter-views were individually open-coded in NVivo. Second, the interinter-views were com-pared for similarities and divergences and subthemes were established. Finally, information gained from all four interviews was synthesized.
The coding and interpretation of results were continuously discussed with co-authors. This triangulation between authors with different perspectives and posi-tions will increase the understanding of complex phenomena.103 Several perspec-tives appeared repeatedly in all four interviews, indicating data saturation.