Thank you for the introduction, hello. I work as an assistant professor at the Department of Medical Informatics in Amsterdam, but my background is in clinical pharmacy. So the paper was already introduced, so let's start with the study. First, I have no conflicts of interest to declare. Some background. So, as you probably know, drug-drug interactions cause patient harm. Drug-drug interactions happen when two drugs are administered together, known to interact, which can cause increase or decrease of one or both drugs, causing toxicity or therapy failure. We also know from previous research that ICU patients are more prone to potential drug-drug interactions. This is because ICU patients are treated with a lot of medications at the same time. Clinical decision support systems help to prevent drug-drug interactions by providing alerts when two drugs known to interact are prescribed. However, these systems seem to not well work in the ICU, and this is because ICU environment differs from non-ICU wards. There is a lot of monitoring going on, and also it is often not possible to reframe from prescribing interacting drugs. Therefore, we see very high override rates of potential drug-drug interaction alerts in the ICU, up to 80%, and this is also very nicely explained by a theoretical framework proposed by Seyss et al. In this framework, you see that if a computerized decision support system is not aligned with the environment where it is deployed, it has low sensitivity, sorry, specificity. And in context of clinical decision support system, specificity means that the system gives an alert only when it is necessary. If it is not well aligned with the environment, it produces a lot of alerts with low yield, so false alerts, causing alert fatigue, high override rates, and ignoring the alerts. Therefore, in our study, we hypothesized that by tailoring the system to the ICU setting will increase specificity of these alerts, resulting in reduced alert fatigue, and therefore more attention to clinically relevant ones. And as a consequence, less exposure to clinically relevant drug-drug interaction will occur, and improve monitoring if one cannot reframe from prescribing such combinations. And how did we test this hypothesis? We conducted a trial in the Netherlands and included nine ICUs, a mix of academic and non-academic ones, and together they admit around 11,000 people yearly. We used a national knowledge database that contains information about which drugs interact, and this database is used in the Netherlands by all Dutch hospitals. However, the content of this database is not tailored to the ICU environment. All ICUs participating in our trial needed to use one and the same clinical decision support system, called medication interaction module, short of MIM, by the provider It's Medical. At the start of the study, five ICUs were already using this type of system, and four were not using any type of clinical decision support. It is important to stress two definitions to understand our study. First, a potential DDI refers to the administration of two drugs known to interact, and the potential refers to the uncertainty if such administration will lead to actual DDI. Clinically relevant PDDIs in the ICU, we call those high-risk drug combinations. And also, it is important to know that in most studies on PDDI frequency in the ICU, people tend to use drug prescriptions and not drug administration. We deliberately chose for drug administrations, because drug prescriptions not always get administered, and we really wanted to study exposure to these high-risk drug combinations. Before we started the trial, we completed two other phases, so the whole project consisted of three. In the first phase, we analyzed how often and what type of PDDIs occur in ICU. We analyzed around two million administration records of around 100,000 patients, using data from electronic health record systems. We identified around 200,000 potential drug-drug interactions, which pertained to 270 PDDI types. And to give you one example, a PDDI type, very well known, is a type called QTC prolonging drugs interactions. And in this type, many, many drugs can occur in different potential to prolong QTC. Using information from this phase one, we conducted DELFI study to define what is really important in the ICU, so which PDDIs are clinically relevant and which not. And we done this national study with a panel of experts of ICU physicians and pharmacists, and together, they analyzed 184 PDDI types and assessed that 36%, so almost 40%, was not clinically relevant in the ICU environment. So this, you can say, are low-yield alerts. Using this information, we conducted a cluster randomized step-by-step trial in phase three, and the intervention was tailoring the CDSS systems according to the DELFI study, meaning turning alerts on for high-risk drug combinations and turning low-yield alerts off. And this is how it looked. The trial lasted one year, and you see that these nine ICUs were started sequentially every month one, and the sequence was randomized, and the intervention was provided at ICU level, targeting the physician, because in the Netherlands, only the physician can prescribe drugs. And to assess the effects in the analysis, we compared high-risk drug combinations in the control period, so the control wedge, to the intervention period. Our primary outcome was the number of administered high-risk drug combinations per thousand drug administration per patient, and secondary ICU length of stay, and the proportion of appropriately monitored high-risk drug combinations. Our statistical analysis were all intentioned to treat basis ones. We used generalized linear mixed effect model to assess the effect, and we created three models, one unadjusted, one adjusted for variables that differ significantly between intervention and control groups, and one adjusted for a number of prognostic factors, which might influence the number of drug-drug interactions. The results, here you see a summary of patients' characteristics. We included around 10,000 patients, divided between control and intervention group, and then looking at differences, we see two that are significantly different. There were slightly more people with medical admission type, and slightly more people having COPD. And although these differences are small, they were significant, and therefore we corrected for those differences in our M1 model. And finally, here are their results. In all models, we see a significant reduction in the number of high-risk drug combinations, and to summarize, it is a reduction between 12 to 14 percent. Regarding the ICU length of stay, we see a reduction of 6 to 10 percent. And regarding the appropriate monitoring, we see an increase of 9 percent of appropriate monitoring when one could not refrain from administering high-risk drug combinations. A couple of strengths and limitations to stress. This is a large multicenter study with a strong design, randomized, a class-randomized step with trial. We measured actual drug administrations as opposed to prescriptions. We conducted a Delphi study to assess the high-risk, low-risk of drug combinations, and besides measuring exposure, we also measured monitoring and ICU length of stay. Regarding limitations, also some other factors may influence effectiveness of the systems. For example, usability, how the alerts look, and when they pop up. But because all the ICUs participating in the trial were using the same CDSS, this limitation is limited. Also, the number of ICUs, the number of clusters was small. It was sufficient for the power and to see a significant effect. However, not big enough to, for example, study differences between academic and non-academic ICUs. Also, we did not assess patient harm associated with this high-risk drug combinations. So, main takeaways from the study. This is a first study evaluating tailoring CDSS to the ICU environment on the administration of high-risk drug combinations, the monitoring and ICU length of stay. Tailoring CDSS alerts for PDDIs improves CDSS effectivity. I could say less is more. For practice, of course, this study also shows that the one size does not fit all. You can't have one CDSS system for the whole hospital. And we also believe that our list of this high-risk drug combinations is transferable to other systems and also to ICUs outside Netherlands because the frequency of PDDIs is more or less comparable between countries. And this concept of tailoring CDSS to a specific environment is also broadly applicable. You can think about pediatric, neonatology, oncology wards. And also, we hope that this study inspires hospitals to take a more critical approach on all the alerts physicians, nurses are getting. In terms of research, of course, more studies are needed to assess tailoring of CDSS to specific environments. It would be good to have a measurement of adverse drug events related to PDDIs. And also, there are many possibilities to improve the PDDI alerts even further. Of course, I did not do this study by myself. Especially, I would like to stress the contribution of Tinka Bakker. She was the PhD student on this project and successfully defended her thesis on 7 December last year. And, of course, there were many other people involved. And I think a really multidisciplinary project involving medical informatics, intensivists, pharmacists, and methodology experts. And also, of course, the study was not possible without the contribution of ICUs and many collaborating partners. And after the talk, I, of course, have time to take your questions. And you can email me or scan this QR code to learn more about the study. Thank you.

clinical decision support systems

drug-drug interaction alerts

ICU alert fatigue

customized CDSS

hospital ward adaptation