← Back to work
Clinical AI

CDS Agent: AI-Powered Clinical Decision Support

An intelligent clinical decision support tool that surfaces evidence-based recommendations at the point of care, reducing alert fatigue and improving guideline adherence in value-based care settings.

Role

Product Lead

Domain

Clinical AI / CDS

Stack

NLP, EHR Integration, FHIR

Setting

Value-Based Primary Care

Problem

Traditional clinical decision support tools generate excessive, low-value alerts that clinicians learn to ignore. In value-based care, missed guideline adherence directly impacts outcomes and cost performance. The challenge was designing a CDS system that clinicians would actually trust and use — one that delivered the right recommendation, to the right clinician, at the right time.

Approach

Led the product design and clinical validation of an AI-powered CDS agent that integrated directly into existing EHR workflows. The approach centered on three principles:

  • Context-aware triggering — using NLP to analyze clinical notes and surface recommendations only when clinically relevant, reducing noise
  • Evidence grounding — every recommendation linked to specific clinical guidelines (USPSTF, ADA, ACC/AHA) with confidence scoring
  • Workflow-native delivery — recommendations embedded in the clinical workflow rather than presented as interruptive pop-ups

Worked cross-functionally with clinical informaticists, data engineers, and care teams to define clinical rules, validate NLP accuracy, and iterate on the UX based on physician feedback.

Results

Pilot deployment across primary care clinics in a value-based care organization demonstrated measurable impact on both clinical quality and workflow efficiency.

40%
Reduction in alert fatigue
22%
Improvement in guideline adherence
3.2x
Clinician engagement vs. legacy CDS

Lessons

The biggest insight was that CDS adoption is fundamentally a design problem, not just an AI problem. The NLP pipeline was necessary but insufficient — what moved the needle was embedding recommendations within existing cognitive workflows so clinicians didn't have to context-switch. Physician co-design from day one was critical to getting the interaction model right.