Clinical Decision Support Software

What this is

Clinical decision support (CDS) software takes patient information and gives clinicians suggestions, such as alerts, order sets, risk scores or diagnostic prompts. The FDA's rules for it center on software that supports health care professionals. This page explains how the FDA draws the regulatory line in the United States and what the research says about whether CDS helps. Evidence labels: established, mixed, promising but limited, insufficient evidence, and marketing claim. This is education, not legal advice, and we did not verify other countries' rules.

Where the FDA draws the line (United States)

The 21st Century Cures Act excluded some decision support from the legal definition of a medical device. FDA guidance, issued January 29, 2026, explains how it reads the four criteria. A tool must meet all four to be outside device regulation. [1]

Some consequences the FDA spells out:

• Software that processes or analyzes a medical image, including computer-aided detection and diagnosis systems, does not meet Criterion 1, so it stays a device. [1]
• Software that supports or provides recommendations to patients or caregivers, not to HCPs, meets the definition of a device. [1]
• Software intended for a critical, time-sensitive task or decision is not considered to meet Criterion 4, because an HCP is unlikely to have time to review the basis. [1]
• FDA says software meant to support time-critical decision making does not meet the definition of Non-Device CDS. [2]
• Failing one criterion does not always mean regulation. The FDA says some software that does not meet all four criteria is still outside device regulation or falls under enforcement discretion. [2]

The FDA explains why the fourth criterion matters: automation bias is the tendency to over-rely on a suggestion from an automated system, and it can lead to errors of commission or omission. [1] A separate US health IT rule from the Office of the National Coordinator defines "predictive decision support interventions," and the FDA notes that some of these may be devices and others not. [2]

What research shows about whether it helps

Regulatory status and effectiveness are separate questions. A tool can sit outside device rules and still work well, or poorly.

• Process of care: modest. A 2020 meta-analysis of 108 studies (122 trials, over 1.2 million patients) found CDS raised the share of patients receiving recommended care by 5.8 percentage points (95% CI 4.0 to 7.6), with large variation between studies. [3]
• Clinical results: small. In 30 trials that reported clinical endpoints such as blood pressure or lipid control, the median change in patients meeting targets was 0.3% (interquartile range -0.7% to 1.9%). [3] Label: mixed, with small average effects on outcomes.
• Older evidence agrees. A 2012 review of 148 randomized trials found improved process measures, but few trials measured clinical outcomes, costs or unintended harms. [4]

Where it goes wrong

Alert overload. A systematic review of 23 studies of drug alerts at the point of prescribing found average override rates between 46.2% and 96.2%. Between 29.4% and 100% of the overrides were judged appropriate, depending on alert type. [5] Many alerts are ignored, and some of the ignoring is sensible.

A widely used prediction tool that was not validated. An outside evaluation of the Epic Sepsis Model at one academic hospital system (27,697 patients, 38,455 hospital stays) reported an area under the curve of 0.63. It missed 67% of the patients who developed sepsis while alerting on 18% of all hospitalized patients. [6] This was one health system, retrospective, and the paper has a published erratum for Table 2 and the results text. [6] Lesson: wide use of a tool is not evidence that it works. See Clinical Validation and Real-World Performance of Health Software.

Over-trust. See Hallucinated Health Information and Automation Bias for how people defer to confident-looking outputs.

What this means for patients

• If a clinician mentions a decision tool, you can ask what it is based on, and what the clinician decides when it disagrees with their judgment.
• A tool that gives you, as a patient, a specific recommendation is on the device side of the FDA's line. [1] See Patient-Facing Health Chatbots.
• Decision tools support care decisions; they do not diagnose you. Image-reading tools are covered in AI-Assisted Diagnosis.