COMMENT|ONLINE FIRST

The efficiency of computerised clinical decision support systems

Andre Carlos Kajdacsy-Balla Amaral, Brian H Cuthbertson

Lancet Published: January 20, 2024

DOI: https://doi.org/10.1016/S0140-6736(23)02839-8

The rapid increase in the uptake of electronic medical records over the past decade is leading to an increased deployment of computerised clinical decision support systems (CDSSs). These range from simple pop-up alerts about serious drug allergiesto more sophisticated tools incorporating clinical prediction rules, prompting clinicians to deliver evidence-based processes of care and discouraging non-indicated care.¹ Although the opportunity for improving processes of care exists in CDSSs, in one systematic review in 2020 of 115 studies, CDSSs produced an average absolute improvement in processes of care of only slightly more than 5%.¹

These frustrating results from such promising technology can be explained by several factors, such as interface, workflow, acceptability, relevance, and timeliness of intervention, among others.² Chief among these factors is alert fatigue, when clinicians ignore alerts because of their excessive and intrusive nature. In one study, clinicians had to review 123 alerts to prevent a single adverse event.³ Several solutions exist for decreasing alert fatigue, such as streamlining the list of drug interactions, learning from past overridden alerts, and focusing such alerts on medications less commonly used.⁴

In The Lancet, Tinka Bakker and colleagues⁵ used a stepped-wedge cluster trial to test the effectiveness of a streamlined CDSS to decrease the number of high-risk drug combinations in nine intensive care units (ICUs) in the Netherlands. A panel of experts categorised several drug combinations into high-risk or low-risk groups and deployed the CDSS only for those in the high-risk group. This decreased the number of potential combinations to 62% of the original list. Compared with the control group using older or no CDSS, it was found that tailoring alerts to the ICU led to a 12% decrease (p=0·0008) in the number of high-risk drug combinations per 1000 drug administrations per patient. However, it is important to note that this finding is of debatable clinical significance, given that the difference in median ICU length of stay was only 3 h between groups (decreasing from 1·10 to 0·97 days) and could well be explained by residual confounding and not the intervention itself.

Although these results are exciting because they were shown on a large scale across several centres, they merely confirm what the literature had already identified: that yes, alerts work, but in the current era of increasing computing power and artificial intelligence, the challenge is not to show that simplistic CDSSs work, but to create systems that are smarter, more user friendly, more interactive, and that have a higher positive predictive value for alerts in terms of clinically relevant outcomes. These systems would not substitute for clinicians, but rather become a useful and non-intrusive tool to augment clinical work. Although there is some work in this area, such systems have yet to be developed.6, 7, 8

Using the main finding from the study by Bakker and colleagues,⁵ it is possible to see how such a system would function. The most frequently reduced high-risk drug combination in this study was the combination of two QT-prolonging agents, which decreased from 28·7% to 21·4% of admissions. Although much attention is paid in ICUs to heart-rate corrected QT interval, there are sparse data on how clinicians should make informed decisions⁹ and debatable evidence that it is actually an important clinical problem for most patients in the ICU, most likely representing a marker of severity of illness.¹⁰

Instead of alerting clinicians about every possible instance of a combination of QT-prolonging drugs, a modern system should integrate baseline heart-rate corrected QT intervals from a 12-lead electrocardiogram; the continuously monitored heart-rate corrected QT interval from such monitors; the dose of the drugs being used (since for most drugs the relationship between dose and QT interval is linear);¹¹ and known risk factors for the development of arrhythmias, such as hypokalaemia, hypomagnesemia, age, gender, bradycardia, neurological conditions, and renal dysfunction. Such alerts would only be brought to the clinician's attention if the integration of all these factors was clinically important or if the continuously monitored heart-rate corrected QT interval increased beyond a set limit where the risk of Torsades de Pointes was high.¹² Furthermore, it would not simply alert the clinician that the patient is taking QT-prolonging drugs, but would present data on the known effects on the heart-rate corrected QT interval of each drug and suggest alternatives that might be less likely to prolong it (for example, quetiapine is unlikely to increase the heart-rate corrected QT interval).¹³ With the addition of large language models, the system would also have read the clinical knowledge to understand why a specific drug is being used and automatically override alerts for drugs that are either essential or provide the clinician with a better alternative and rationale for them.

This method seems complex, but systems should be tailored to the needs of the environment. Patients in the ICU are the most complex patients in medicine, and this is an obvious reason why recommendations from generic or simplistic systems are frequently overridden in patients in the ICU. For example, recommendations from a CDSS to tailor antimicrobial therapies were accepted in only 38% of patients in an ICU because of the need for a more integrated evaluation that was undertaken by the clinician.¹⁴ It is necessary to foster and fund future work using the robust methods from the study by Bakker and colleagues,⁵ but using more advanced and intelligent CDSSs.