Statistical analysis

To compare the distribution of included patients between the preintervention and the postintervention period, a Wilcoxon rank-sum test was used for continuous data and χ² test for categorical data.

By modelling the ITS data using a segmented regression analysis, both the time trend of each period and the immediate intervention’s effect were investigated.⁴⁰ In the segmented regression model, the estimated effects were expressed as incidence rate ratio (IRR). The incidence rate (IR) was defined as the number of persistent PIVs divided by the number of initially identified PIVs at T₀. The IRR quantified the relative increase or decrease of the IR as a result of intervention or time.

The model is specified as⁴¹:

Y_t = β₀ + β₁ time_t + β₂ intervention_t + β₃ time after intervention_t + ε_t

Y_t is the value of the dependent variable (IR) in month t.

Time is a continuous variable indicating time in months at period t, whereby time is centred at the intervention.

Intervention is an indicator for time t occurring before or after implementation.

Time after intervention is a continuous variable counting the number of months after the intervention at time t.

β₀ estimates the preintervention IR of persistent PIVs at the beginning of the time series.

β₁ estimates the preintervention trend.

β₂ estimates the immediate change in the level of the IR after implementation of the intervention.

β₃ estimates the change in the trend after implementation.

ε_t is an estimate of the random error.

To calculate a power-based sample size, a mean number of 16 recommendations per day for IVOS was considered. To detect an expected decrease of 50% in the primary outcome with a power of 95%, 12 data points (using days as data points) in each period were required. To ensure that a stable estimate of the baseline underlying secular trend would be obtained, 12 days spread over 4 months from December to March were analysed for each year in the preintervention period (from December 2015 to March 2019), totalling up to 48 data points. For the postintervention period, the same 12 days, supplemented by 12 extra days (to estimate a more reliable effect of time), were analysed from December 2019 to July 2020, totalling up to 24 data points.

The segmented regression analysis was performed using SAS software (V.9.4). Estimated effects with 95% CIs were calculated. P<0.05 was considered statistically significant.

For the postintervention period, descriptive analyses were performed. The positive predictive value (PPV), that is, the probability of alerts leading to recommendations, was measured as the ratio of the total number of pharmacists’ recommendations to the total number of alerts generated.