What is clinical decision support? Types, examples and AI in 2026

The US definition is set by the Assistant Secretary for Technology Policy and Office of the National Coordinator for Health IT (ASTP/ONC). Its current wording, revised in April 2026, describes CDS as a digital tool that provides timely and person-specific information, intelligently filtered or presented at appropriate times, to enhance patient outcomes and quality of care ^[1]. Two words do most of the work: person-specific, which separates CDS from a textbook, and timely, which separates it from a report you read next week.

Older sources quote a broader formulation from the 2007 AMIA roadmap by Osheroff and colleagues, which said CDS provides clinicians, staff, patients or other individuals with knowledge and person-specific information to enhance health and health care ^[2]. The older wording matters because it names the recipients: CDS is not only for physicians. A pharmacist's interaction check, a nurse's falls-risk prompt and a patient's medication reminder are all decision support. AHRQ's shorter description adds where it usually happens: CDS provides timely information, usually at the point of care, to help inform decisions about a patient's care ^[18].

You will also see CDSS, for clinical decision support system. The two terms are used loosely and often interchangeably; the distinction worth keeping is that CDS is the intervention — the alert, the order set, the answer — and the CDSS is the software that produces it ^[4]. One CDSS can deliver many different CDS interventions, and one intervention can draw on several systems.

Researchers also split CDS by how it reaches a recommendation. Knowledge-based systems run rules — IF-THEN statements written by people from guidelines and evidence. Non-knowledge-based systems use artificial intelligence, machine learning or statistical pattern recognition to derive the relationship from data instead ^[4]. A drug-interaction alert is knowledge-based; a deterioration score trained on vital signs is not. Large language models that reason over retrieved papers sit in the second family, but behave differently from a classic risk model: their output is prose and argument rather than a number.

The distinction is practical, not academic. A rule can be read, so its basis is inspectable by design; its weakness is maintenance, because every rule has to be written and kept current by someone. A trained model scales without that hand-maintenance but can be a black box, and depends on the quality of the data it learned from ^[4]. Most of the regulatory and governance debate about AI in CDS is about closing that inspectability gap.

CDS Hooks lets the EHR call an external CDS service when the clinician reaches a defined point in the workflow — a hook — and display what comes back as cards. The current published release is 2.0.1 (STU 2, technical correction 1), published on 12 March 2025 and built on FHIR R4; a 3.0.0 normative ballot was posted in December 2025 but is not the published version ^[9].

The hook library includes patient-view (the chart is opened), order-select and order-sign (an order is chosen and signed), order-dispatch, encounter-start, encounter-discharge, appointment-book, medication-refill, allergyintolerance-create and problem-list-item-create; the older medication-prescribe and order-review hooks are deprecated ^[9]. Each card must carry a summary under 140 characters, an indicator of info, warning or critical, and a source; it can add suggestions the clinician can accept, and links that launch a SMART on FHIR app for the full workflow. CDS Hooks shares its underlying framework with SMART on FHIR, which is how a card hands off to a richer app ^[9].

Not all CDS is wired into the chart. Reference platforms, pathway apps and AI assistants are often used beside the EHR — on a second screen, a phone or a browser extension — which is the right channel for a question that needs reading and reasoning rather than a two-second interrupt. The trade-off is that the clinician has to bring the context: the Five Rights still apply, they are simply met by the user rather than by the integration.

The most cited review is Kawamoto and colleagues in the BMJ in 2005. Across 70 randomised controlled trials, CDS significantly improved clinical practice in 68% of them ^[5]. More useful than the headline is what predicted success. Four features were independently associated with improvement, and of the 32 systems that had all four, 30 — 94% — improved practice ^[5].

Fifteen years later, Kwan and colleagues pooled 122 trials covering more than 1.2 million patients. Computerised CDS increased the proportion of patients receiving the desired care by 5.8% (95% CI 4.0% to 7.6%), with wide variation between trials (I² 76%) and a top quartile of 10% to 62% ^[6]. For clinical endpoints the picture was much thinner: across 30 trials, the median improvement in patients reaching guideline targets was 0.3% ^[6].

The honest reading is that CDS reliably nudges process — more patients get the recommended test, drug or dose — and that turning a process gain into an outcome gain depends on whether the recommended care itself moves outcomes, and on whether the support is designed well enough to be acted on rather than dismissed.

A 2006 review in JAMIA found that clinicians override drug-safety alerts in 49% to 96% of cases ^[7]. A 2020 systematic review of 23 studies found average override rates from 46.2% to 96.2% ^[8], and one overview estimates that up to 95% of CDSS alerts are inconsequential ^[4]. When nearly every alert is ignorable, the important one is ignored too. Overrides are not always wrong — many are clinically justified — but a high override rate is a sign that the right-information and right-person tests are failing.

The evidence base for large language models in clinical care is still young. A 2025 JAMA systematic review of 519 studies evaluating LLMs in health care found that only 5% used real patient care data ^[14]. And the first randomised trial of an LLM as a diagnostic aid found that giving 50 physicians access to the model did not significantly improve their diagnostic reasoning score (76% with the model, 74% without), even though the model alone scored 16 percentage points higher than the physicians without it ^[15]. The lesson is the one the Kawamoto review taught about rules: the tool's raw performance matters less than how its output reaches, and is used by, the clinician.

That is why the question to ask of any AI CDS is not only whether it is accurate, but whether you can check its basis yourself. The FDA's January 2026 guidance makes the clinician's ability to independently review the basis for a recommendation one of the four tests that keep CDS software outside medical-device regulation ^[13][17]. In US certified health IT, ONC's HTI-1 rule added a Decision Support Interventions criterion, part of the Base EHR definition from 1 January 2025, requiring 13 source attributes for evidence-based interventions and 31 for predictive ones, plus risk management covering validity, reliability, robustness, fairness, intelligibility, safety, security and privacy ^[10][11]. Those source-attribute and risk-management requirements are proposed for removal in the HTI-5 proposed rule of December 2025; as of October 2026 no final rule has been published, so they remain in the regulations ^[12].