Clinical decision support (CDS) is software that puts the right information, for the right patient, in front of a clinician at the moment they act. A clinical decision support system (CDSS) is the same job with “system” in the name — not a different product.
If you meant AI-powered CDS (how to implement it, bias, black box) → clinical decision support AI. If you meant CAD / AI diagnosis → AI for medical diagnosis. If you meant AI in radiology → artificial intelligence in radiology. If you meant EHR interop → interoperability in EHR.
PYCAD builds imaging AI that can feed a CDS workflow (a mark on a scan); it is not a CDSS vendor.
Three parts
A CDSS is not one box. It is a library, a reasoner, and a way to speak.
- Knowledge base. The library: drug–drug rules, dosing formulas, guideline order sets, allergy tables. It goes stale unless someone feeds it.
- Inference engine. The reasoner. New order in the EHR, run it against the library. If-then is still most of what ships: if penicillin allergy and amoxicillin, then stop.
- Communication. How the answer reaches the clinician: a blocking allergy pop-up, a quiet flag on a lab, a graph, a pre-built pneumonia order set. A brilliant rule that nobody sees is a unused rule.
Clinicians override safety alerts somewhere in the 49–96% range in the literature. That is alert fatigue, not a reason to skip the engine. It is a reason to stop shouting.
A prescription, worked
A patient already on four drugs. The physician types a fifth. The engine cross-checks the new drug against the med list, allergies, and last eGFR, and flags a pair that can damage the kidney in this person. The alert names both drugs, the risk, and a safer alternative from the guideline — not a red “Warning” with no next step. The physician still signs. The system made the pause cheap.
Passive vs active, and what wakes it
Passive is a library in the chart: a link to the diabetes guideline when that problem list item is open. Optional. Quiet.
Active is a shoulder-tap: the allergy fire, the overdose, the duplicate CT from last Tuesday. It interrupts because the cost of missing it is high.
A well-built CDSS stays quiet until a trigger. Common ones:
- Prescribe — interaction, allergy, duplicate therapy, a dose that ignores weight or kidney function.
- Order a test — last week’s same CT; a more useful panel for this differential.
- Enter a diagnosis — the standard workup for new hypertension, not a blank order pad.
- Vitals — a drop in pressure plus a fever that should make someone think sepsis.
Knowledge-based vs not
| Knowledge-based | Non-knowledge-based | |
|---|---|---|
| How | If-then rules a human wrote | A model that found a pattern in a pile of charts |
| Typical job | Allergy, interaction, guideline reminder | Risk score, ranked differential, a mark on a scan |
| You can audit | Yes — point at the rule | Only if the model can show its work |
| Goes stale when | Nobody updates the library | The catchment drifts and nobody re-fits |
Hospitals run both. The rule engine is the safety net. The model is the extra signal. How to implement the AI side — bias, black box, a pilot — is clinical decision support AI.
Imaging as a second reader
A CDSS can sit on a chest CT and highlight a likely PE so the radiologist looks there first. That is a second reader inside a CDS workflow, not a diagnosis product. The CAD / AI-diagnosis job is AI for medical diagnosis. What AI in radiology is, as a field, is artificial intelligence in radiology. This page does not become either.
FAQ
How does a CDSS talk to an EHR?
HL7 v2 and FHIR, same as any other clinical app: pull the meds and allergies, push the alert back into the chart the clinician already has open. The languages and the pipe are interoperability in EHR. This page is what the engine does once it can see the chart.
Is a CDSS just another name for the EHR?
No. The EHR is the chart. The CDSS is the assistant that reads the chart and speaks at the point of an order. You can have an EHR with almost no decision support. You cannot have useful CDS with nowhere to read the meds from.
Will it replace doctors?
No. It flags, reminds, and pre-fills. The call, the conversation, and the signature stay human.
What usually goes wrong?
Alert fatigue first — too many low-value pops, and the critical one is ignored. Then a stale knowledge base, an integration that never quite lands in the EHR, a bill nobody budgeted, and staff who were shown a demo instead of a workflow. Calibrate the noise. Update the library. Train on the overrides.
PYCAD builds imaging AI that can feed a CDS workflow (a mark on a scan); it is not a CDSS vendor. Case studies.