Clinical Decision Support, Patient Safety, Quality of Health Care
Conditions
Brief summary
This trial evaluates CITE, a retrieve-and-verify layer that audits an AI-generated care plan against a full-text evidence corpus and flags patient-specific codifiable safety hazards to the clinician. The co-primary outcomes are how accurately CITE flags these hazards (sensitivity and specificity versus blinded clinician adjudication) and its clinician alert burden and acceptance, compared with AI care plans using safety guardrails alone and with unassisted clinician care, in Medicaid primary care.
Detailed description
Patients are randomized 1:1:1 to (1) unassisted clinician care; (2) AI-generated care plan with safety guardrails; (3) AI-generated care plan with safety guardrails plus CITE. CITE audits the finalized plan against a frozen, versioned evidence corpus and returns physician-facing flags for patient-specific codifiable safety hazards (a recommended drug contraindicated by this patient's diagnosis or laboratory value; a drug-allergy conflict; a dropped high-risk medication; a guideline-indicated therapy omitted for an active diagnosis; a stated quantity refuted by the corpus), each with a verbatim quote and citation; the clinician retains decision authority. Randomization uses a deterministic HMAC permuted-block scheme; outcome assessors are blinded to arm. The co-primary outcomes are (1) the diagnostic accuracy (sensitivity and specificity) of CITE against blinded clinician adjudication, and (2) clinician alert burden (flags per encounter) and acceptance, comparing the CITE arm with the guardrail arm; both are estimable at the enrolled sample size because they do not depend on a rare between-arm event. The unresolved codifiable-hazard rate by arm is reported as a descriptive secondary: codifiable hazards are infrequent, so the trial is not powered for a between-arm efficacy contrast on hazard reduction. A prior trial of a different mechanism (a generic deterministic rule-corpus that surfaced roughly 30 or more flags per encounter and was uninformative) was completed with null results and is registered separately; this trial evaluates a materially different, patient-specific intervention and set of outcomes. Determined exempt by WCG IRB (low risk). Analysis is pre-registered on OSF (https://doi.org/10.17605/OSF.IO/ENXCW).
Interventions
OTHERCITE evidence-grounding retrieve-and-verify layer
Reads the AI care-plan text and verifies each recommendation/claim against a full-text evidence corpus; returns physician-facing flags (commission/confabulation/unsupported/omission) with verbatim quotes and citations. Clinician retains decision authority.
OTHERAI care plan with safety guardrails
AI-generated care plan produced under a safety-guardrail system prompt.
Sponsors
Waymark
University of California, San Francisco
Study design
Allocation
RANDOMIZED
Intervention model
PARALLEL
Primary purpose
HEALTH_SERVICES_RESEARCH
Masking
SINGLE (Outcomes Assessor)
Eligibility
Sex/Gender
ALL
Age
18 Years to No maximum
Healthy volunteers
No
Inclusion criteria
1. Age 18 years or older. 2. Medicaid-enrolled and attributed to a participating Waymark primary care site. 3. Primary care encounter that requires clinical reasoning (not administrative-only). 4. English-language clinical documentation.
Exclusion criteria
1. Age less than 18 years. 2. Hospice or palliative-care-exclusive care plan. 3. Administrative-only or pharmacy-only encounter that does not surface a clinical decision to the supervising clinician. 4. Encounter where the supervising clinician is the principal investigator. 5. Enrollment in a competing AI-safety study within the prior 90 days.
Design outcomes
Primary
| Measure | Time frame | Description |
|---|---|---|
| Diagnostic accuracy of CITE against clinician adjudication | Day 1 (index primary care encounter) | Sensitivity and specificity (with positive and negative predictive values) of the CITE checker for clinically consequential codifiable safety hazards, using blinded clinician adjudication of the plan as the reference standard. Every plan contributes, so the estimate does not depend on a rare between-arm event. Exact-binomial 95% confidence intervals; reported overall and by hazard family. |
| Clinician alert burden (flags surfaced per encounter) | Day 1 (index primary care encounter) | Number of safety flags surfaced to the clinician per encounter in the CITE arm versus the guardrail arm, with clinician acceptance rate. Co-primary usability outcome: a verifier that surfaces an unmanageable number of flags is not deployable regardless of sensitivity (the prior-trial mechanism surfaced a median of about 30 per encounter). Pre-registered acceptability ceiling: median CITE flags per encounter at or below three. |
Secondary
| Measure | Time frame | Description |
|---|---|---|
| Clinician action on CITE flags | Day 1 (index primary care encounter) | Proportion of CITE flags accepted vs overridden by the clinician, by flag type (commission, confabulation, unsupported, omission). |
| Unresolved codifiable safety-hazard rate by arm (descriptive) | Day 1 (index primary care encounter) | Proportion of patient-specific codifiable-hazard checkpoints with an unresolved hazard in the finalized plan, by arm, with the Arm 3 minus Arm 2 difference and 95% confidence interval. Pre-specified as descriptive and hypothesis-generating: codifiable hazards are infrequent, so the trial is not powered for a between-arm efficacy contrast on this measure at the enrolled sample size. |
| Correction of codifiable hazards within 30 days | Up to 30 days after the index encounter | Among checkpoints with a hazard in the finalized plan, the proportion acted on and corrected within 30 days (documented resolution, completed referral, or corrected order). Proximal clinical effectiveness measure. |
| Completed referrals within 30 days | Up to 30 days after the index encounter | Proportion of initiated referrals completed within 30 days. |
| Clinical safety composite (exploratory) | Day 1 (index primary care encounter) | Four-component clinical safety composite carried from the prior trial. Pre-specified as exploratory; underpowered at the planned sample size. |
| 30-day acute care utilization (exploratory) | Up to 30 days after the index encounter | Emergency department visits and hospitalizations within 30 days. Exploratory. |
Outcome results
None listed