Clinical decision support (CDS)
Also known as: CDS, clinical decision support system, CDSS
Clinical decision support (CDS) is technology that gives clinicians relevant information or recommendations at the moment they make decisions: a warning about a dangerous drug interaction, a reminder that a patient is due for screening, or an AI alert predicting sepsis.
Good CDS prevents errors and improves care. While its potential is huge, poorly designed CDS creates 'alert fatigue,' where clinicians ignore so many warnings that the important ones get lost.
Clinical decision support (CDS), explained at three levels
At home
Think of a car's lane-departure warning. It does not drive for you, but it beeps when you drift. Clinical decision support does that for clinicians. Software in the health record checks what is being ordered or recorded against rules and data, and warns about a dangerous drug combination, reminds about an overdue screening, or flags a patient who may be getting sicker.
At work
Clinical decision support is technology that gives clinicians patient-specific information or recommendations within their workflow, such as drug interaction checks, order sets, reminders, and risk predictions. For health organizations, it is a main tool for standardizing care, preventing errors, and meeting quality targets, but badly tuned CDS wastes clinician time and teaches people to ignore warnings.
At home
Most CDS runs inside the health record and follows a trigger-and-check pattern. When your doctor enters a prescription, the system compares it with your allergy list, other medicines, kidney function, and age, using a drug knowledge database. If it finds a problem, it shows a warning. Other types are quieter: order sets bundle the usual tests for a condition, and dashboards flag patients at risk. Some use fixed rules; others use machine learning predictions.
At work
CDS combines a trigger (an order, a chart being opened, a new lab result), a knowledge source (rules, drug databases, or a predictive model), and an intervention (alert, order set, documentation template, dashboard). The 'five rights' framework says it must deliver the right information, to the right person, in the right format, through the right channel, at the right time. CDS Hooks lets external services return recommendation cards to the EHR at defined moments, such as when an order is signed.
At home
Each alert competes for attention, and most are low value: clinicians override the large majority of drug alerts. That trains people to click past warnings, including the rare important ones, a problem called alert fatigue. Predictive models add a different risk: a sepsis score trained at one hospital may miss most cases at another. While adding alerts feels safe, in my opinion removing low-value alerts is often a bigger safety improvement than adding new ones.
At work
Govern CDS with a committee that owns each rule and model, tracks firing and override rates, and retires what does not help. For predictive models, validate locally before go-live, monitor for drift and performance across patient groups, and run silent-mode trials first. Some CDS software qualifies as a medical device under Health Canada and FDA rules, depending on whether clinicians can independently review the basis for a recommendation. Define accountability when recommendations are followed or ignored.
Real-world examples
At home
Aisha's doctor tried to prescribe an antibiotic, and the system warned it could interact with Aisha's blood thinner. The doctor chose a different antibiotic, avoiding a risky bleeding complication.
At work
Fatima is a chief medical information officer at an Ontario hospital. Her team reviewed drug alerts and found that 95% were overridden. They turned off the lowest-value alerts, and override rates on the remaining critical warnings dropped noticeably.
How it compares
| feature | Rule-based CDS | AI-based CDS |
|---|---|---|
| Logic | If-then rules written by experts | Predictions learned from data |
| Transparency | Easy to explain | Often harder to explain |
| Example | Drug allergy warning | Sepsis or deterioration prediction |
| Main risk | Too many irrelevant alerts | Poor accuracy in new settings |
Rules are predictable but blunt. AI is adaptive but needs validation.
For analysts
Every new alert costs clinician attention. I'd require each CDS rule or model to have a named owner, a measured benefit, and a review date, and retire the ones that don't earn their interruption.
History & origin
Early computer systems such as MYCIN, developed at Stanford in the 1970s to recommend antibiotics, showed computers could reason about medicine but were never used clinically. The Regenstrief Institute and Intermountain Healthcare built CDS into records systems in the 1970s and 1980s. As electronic health records spread in the 2000s and 2010s, drug-interaction alerts and order sets became standard. Studies found clinicians overrode most alerts, sometimes over 90% of them. In 2021, a JAMA Internal Medicine study of a widely used proprietary sepsis prediction model found it missed about two-thirds of sepsis cases while alerting on many patients who didn't have it. In my view, that study changed the conversation: AI in clinical settings needs independent validation, not vendor assurances.