Clinical Decision Support

What Is Clinical Decision Support?

A clinical decision support (CDS) system is any tool that helps a clinician make better decisions at the point of care. Before AI, CDS was simple — drug allergy alerts, dosage calculators, and formulary lookups built into the electronic health record. Useful, but limited. Clinicians saw so many alerts that they developed "alert fatigue" and started clicking past everything.

AI-powered CDS is different. Instead of rigid if-then rules, it uses machine learning to analyse patterns across thousands or millions of patient records and surface insights that would be impossible for a human to derive manually. Think of it as the difference between a smoke detector (rule-based: if smoke, then alarm) and a fire investigator (pattern-based: analysing heat signatures, airflow, materials, and history to predict where a fire is most likely to start).

Emergency Triage: ESI in the Western World

When a patient arrives at an emergency department in the United States, Canada, Australia, or much of Europe, they are assessed using the Emergency Severity Index (ESI) — a five-level triage system that sorts patients by acuity and expected resource needs.

ESI is a well-validated system, but it relies heavily on the experience and judgement of the triage nurse. A 2023 study in *Annals of Emergency Medicine* found that inter-rater agreement for ESI levels 2-4 was only 60-70%. Two experienced triage nurses can look at the same patient and assign different levels.

How AI Improves Triage

AI triage models ingest data available at the point of arrival — vital signs (heart rate, blood pressure, oxygen saturation, temperature), chief complaint text, age, medical history from the EHR — and predict clinical outcomes: likelihood of ICU admission, need for emergency surgery, or risk of deterioration within 24 hours.

The AI does not replace the triage nurse. It provides a "second opinion" — flagging patients who look ESI 3 on the surface but whose vital sign patterns suggest ESI 2 risk. At hospitals like Mount Sinai and Mayo Clinic, AI-assisted triage has reduced undertriage rates (patients assigned a level that is too low) by 15-25%.

> Look at data/triage-scenarios.json for the ESI triage practice cases used in the sandbox exercises.

EHR Integration: The Clinical Workflow

In Western healthcare, the electronic health record is the centre of gravity. Everything flows through it — orders, results, notes, billing, prescriptions, referrals. The two dominant EHR systems are:

Epic — Used by over 250 million patients in the US, plus major health systems in the UK, Netherlands, Denmark, and Australia. Epic's market dominance means that any AI tool that wants to reach US clinicians must integrate with Epic.

Oracle Health (formerly Cerner) — The second-largest EHR vendor, now owned by Oracle. Strong presence in federal healthcare (US Department of Veterans Affairs, Department of Defense) and the UK's NHS.

CDS Hooks: The Standard for AI Integration

How does an AI recommendation actually reach the doctor at the right moment? Through a standard called CDS Hooks — a lightweight API specification that triggers AI-powered suggestions at specific points in the clinical workflow.

Think of CDS Hooks as doorbell buttons placed at key moments:

When a CDS Hook fires, the AI service receives structured patient data (via FHIR — Fast Healthcare Interoperability Resources — the standard data format for health records), runs its model, and returns a "card" with a recommendation that appears inline in the EHR. The doctor can accept, dismiss, or override it.

This is critical because context is everything in clinical AI. An alert that appears at the wrong time is noise. An alert that appears at the exact moment a doctor is about to sign a risky prescription is a safety net.

Clinical Pathways: Western Disease Profiles

AI-powered clinical pathways guide treatment decisions for common conditions. In Western populations, the leading causes of death and disability shape which pathways matter most.

Heart Disease (Leading Cause of Death in US, UK, EU, Australia)

AI assists at multiple points in the cardiac care pathway:

Stroke (Time-Critical Emergency)

The phrase "time is brain" captures why AI matters here. Every minute of untreated large vessel occlusion stroke destroys 1.9 million neurons.

Cancer Screening (Breast, Lung, Colorectal)

> Look at data/clinical-pathways.json for the disease-specific clinical pathway templates used in the sandbox exercises.

Real-World CDS Deployment: Sepsis Prediction at Scale

Sepsis — a life-threatening response to infection — kills over 350,000 Americans annually and is the leading cause of in-hospital death. Early detection improves survival by 20-30%, but sepsis is notoriously difficult to recognise because its early symptoms (elevated heart rate, low-grade fever, confusion) overlap with dozens of benign conditions.

Epic's Sepsis Prediction Model runs continuously in the background for every admitted patient. It analyses vital signs, lab results, nursing assessments, and medication orders in real time. When the model detects a pattern consistent with early sepsis, it generates a Best Practice Advisory (BPA) alert in the nurse's workflow.

The results have been mixed — and this is an important lesson. Early versions of sepsis AI had high false positive rates (alerting for patients who did not have sepsis), contributing to alert fatigue. The most successful implementations pair the AI alert with a structured clinical response: a nurse-driven assessment protocol ("sepsis huddle") that determines whether the alert warrants action. The AI gets the team's attention; the human applies judgement.

Key Takeaways