Real-World Effectiveness of AI-Enabled Clinical Decision Support in Primary Care: Systematic Review.
Yash Jain, Dishan Wu, Zhong Wang
✦ AI-curated · Sources linked
In 30 seconds
This systematic review synthesized real-world evaluations of clinician-facing AI-enabled clinical decision support in primary and ambulatory care, covering studies published from 2011 through 2025. Of 24 eligible studies, 10 evaluated AI-enabled tools and formed the appraised evidence base; 14 rule-based studies were included only for context. Findings varied across detection, diagnostic performance, care processes, and efficiency. AI electrocardiography was associated with more new diagnoses of low ejection fraction (OR 1.32, 95% CI 1.01-1.61), but neither trial assessing patient-important outcomes demonstrated benefit from AI-CDSS.
Key findings
- AI electrocardiography increased new low ejection fraction diagnoses (OR 1.32, 95% CI 1.01-1.61); an EHR dementia marker used alone did not.
- Diagnostic performance varied: melanoma discrimination was high (AUROC 0.960), while glaucoma discrimination was moderate (AUROC 0.80; sensitivity 65%).
- Nonrandomized studies reported urinary tract infection treatment success rising from 75% to 80% and changes to diabetic retinopathy screening criteria in 3 of 4 general practitioners.
- Neither trial assessing patient-important outcomes demonstrated benefit; one musculoskeletal functional outcome favored usual care.
Why it matters
The review distinguishes promising signals in detection and care processes from evidence of improved patient health. For clinicians and health systems, this cautions against treating diagnostic accuracy or workflow changes as proof of clinical benefit. The mixed findings and limited outcome measurement underscore the need to evaluate how AI support affects decisions and patient-important outcomes together.
What to keep in mind
The evidence was heterogeneous, and limited outcome measurement prevented causal explanations for differences. Only 10 AI-enabled studies formed the appraised review population, and patient-important outcomes were assessed in just two trials.