Skip to main content

Clinical Decision Support Systems

Clinical decision support systems aim to provide the right information, to the right person, at the right time, in the right format. Most fail at one or more of these criteria. The history of CDS is a cautionary tale about the gap between technically correct and clinically useful.

finished certainty: likely importance: 7

In the early 1970s, Ted Shortliffe at Stanford developed MYCIN, an expert system for diagnosing bacterial infections and recommending antibiotics. In blinded evaluations, MYCIN’s recommendations were rated as appropriate by infectious disease experts 65% of the time — equal to or better than the attending physicians. MYCIN was never deployed clinically. The technology worked; the context didn’t.1

This pattern — technically capable systems that fail to change clinical practice — has repeated with remarkable consistency across five decades of clinical decision support (CDS) research. The technical challenges of building predictive models have been substantially addressed. The human, organizational, and regulatory challenges of deploying them remain formidable.

Taxonomy of CDS

Clinical decision support encompasses a broad range of interventions, from simple to complex:

Knowledge-based CDS uses explicit rules derived from clinical guidelines. Drug-drug interaction alerts, allergy warnings, dose-range checks, and evidence-based order sets are the most common forms. These are embedded in essentially every modern EHR system and represent the overwhelming majority of deployed CDS.2

Non-knowledge-based CDS uses data-driven models — machine learning, statistical models, neural networks — to generate predictions or recommendations. These are technically more sophisticated but less widely deployed. Examples include sepsis prediction models, deterioration early warning systems, and diagnostic support tools.

Passive CDS provides information without requiring clinician action — dashboards, risk scores displayed in the chart, relevant literature links. Active CDS interrupts the clinician’s workflow with alerts, recommendations, or required actions (e.g., “hard stops” that prevent order entry without acknowledgment).

The Alert Fatigue Problem

Alert fatigue is the single biggest problem in clinical decision support, and it is largely self-inflicted.

The logic seems sound: if a drug interaction is dangerous, alert the prescriber. But the implementation is catastrophic. A typical hospitalist receives dozens to hundreds of CDS alerts per day. Most are clinically irrelevant — low-severity drug interactions, duplicate alerts for medications already reviewed, warnings about conditions the patient doesn’t have. The signal-to-noise ratio is abysmal.3

The result is that clinicians develop “alert blindness” — they click through alerts without reading them, including the rare alert that carries genuinely critical information. This is not a technology failure; it’s a design failure. CDS systems were built to be comprehensive (alert on everything that might matter) rather than specific (alert only on things that do matter in this specific clinical context).

ML-Based CDS

The promise of machine learning CDS is precision: rather than applying the same rules to every patient, ML models can generate patient-specific predictions that account for individual risk factors, current clinical state, and treatment context.

Sepsis prediction is the most prominent deployed ML-CDS system — and also the most cautionary tale. Epic’s sepsis prediction model was deployed across hundreds of hospitals. Wong et al. (2021) evaluated it at a large academic medical center and found a positive predictive value of just 12%, with two-thirds of sepsis cases missed entirely. The model performed far worse than its marketing suggested.4

More promising examples exist. Google DeepMind’s acute kidney injury prediction system (Tomašev et al., 2019) predicted AKI up to 48 hours in advance with clinically useful accuracy. The eCART early warning system for clinical deterioration showed improved outcomes in prospective evaluation.

Diagnostic support using large language models is the newest frontier. Google’s Med-PaLM 2 achieved expert-level performance on medical question answering. GPT-4 has passed the USMLE with scores well above the passing threshold. Whether these capabilities translate to useful diagnostic support in clinical practice — where the challenge is not answering exam questions but integrating ambiguous findings with patient preferences and clinical context — remains undemonstrated.5

The Five Rights of CDS

The AMIA (American Medical Informatics Association) framework for effective CDS articulates “five rights”: the right information, to the right person, in the right format, through the right channel, at the right time in the workflow.

Most failed CDS systems violate one or more:

  • Wrong information: Alerts that are technically correct but clinically irrelevant (low-severity interactions, alerts for conditions the patient doesn’t have).

  • Wrong person: Sending alerts to the prescriber when the pharmacist is the appropriate decision-maker, or vice versa.

  • Wrong format: Dense text alerts that require reading in a time-pressured environment. Effective CDS should be glanceable.

  • Wrong channel: Pop-up alerts that interrupt workflow for non-urgent information. Many alerts would be better as passive indicators (e.g., an icon in the patient header) rather than active interruptions.

  • Wrong time: Alerts that fire at order entry when the decision was made hours earlier during rounding. By the time the alert appears, the clinician has already committed to the plan cognitively.

The five rights framework is simple and widely cited but rarely operationalized. Building CDS that satisfies all five requires deep understanding of clinical workflows — which requires direct observation of clinicians in practice, co-design with end users, and iterative deployment. Most CDS development occurs in informatics departments with limited clinical workflow integration.

The Implementation Science Gap

The evidence base for CDS effectiveness is mixed. Bright et al. (201214ya) conducted a systematic review and found that CDS improved clinical practice in roughly 60% of studies — but the improvements were often modest, and the studies were predominantly from a small number of institutions with strong informatics infrastructure.

Kawamoto et al. (200521ya) identified four features of CDS systems that were independently associated with improved clinical practice: 1. Automatic provision (not requiring clinician initiative) 2. Integration into clinical workflow (not standalone systems) 3. Actionable recommendations (not just information) 4. Delivery at the point of decision-making

Systems with all four features were significantly more likely to improve clinical outcomes than systems lacking any of them. Twenty years later, many deployed CDS systems still lack one or more.

These findings have been replicated repeatedly over two decades, yet adoption remains inconsistent.6

The gap between CDS research and CDS implementation is a case study in implementation science — the discipline of studying how evidence-based interventions are adopted (or not) in real-world practice. The barriers are organizational (institutional inertia, competing priorities), technical (EHR integration, interoperability), human (alert fatigue, trust, workflow disruption), and regulatory (liability, FDA oversight of clinical software).

Where This Is Going

The integration of LLMs into CDS is the most transformative near-term development. Rather than rule-based alerts or risk scores, imagine a system that can synthesize a patient’s entire chart, relevant literature, and institutional guidelines into a concise, context-aware recommendation — delivered in natural language, at the right moment in the workflow.

Ambient clinical intelligence — systems that listen to patient-provider conversations and automatically generate documentation, flag relevant guidelines, and suggest orders — is being deployed by Nuance/Microsoft, Abridge, and others. This represents a shift from CDS-as-alert to CDS-as-ambient-assistant.

Whether this new generation of CDS will succeed where previous generations failed depends on whether it addresses the fundamental problems: workflow integration, clinician trust, false positive rates, and the five rights. Better technology alone has never been sufficient. If the new systems are just better at generating alerts that clinicians ignore, we’ll have the same problem with fancier tools.

Further Reading