Scalable Clinical Oversight of Large Language Models Via Uncertainty Triangulation

Official title

Prospective Evaluation of a Model-Agnostic Meta-Verification Framework (SCOUT) for Scalable Clinical Oversight of Large Language Model Outputs in Coronary Heart Disease Diagnosis: A Multi-Reader, Randomized, Crossover Trial

Overview

This prospective, multi-reader, randomized crossover trial evaluates SCOUT (Scalable Clinical Oversight via Uncertainty Triangulation), a model-agnostic meta-verification framework that selectively defers unreliable large language model (LLM) predictions to clinicians by triangulating three orthogonal uncertainty signals: model heterogeneity, stochastic inconsistency, and reasoning critique. The trial assesses whether SCOUT-assisted review can reduce physician review time compared with standard manual review of AI-generated diagnoses while maintaining non-inferior diagnostic accuracy in coronary heart disease (CHD) subtyping.

Detailed description

Background: Large language models are increasingly deployed in clinical workflows, yet requiring clinician review of every AI output negates the efficiency gains that motivate their adoption. SCOUT addresses this efficiency-safety paradox through algorithmic meta-verification.

The SCOUT framework triangulates three orthogonal external signals to determine case-level uncertainty: (1) Model Heterogeneity - whether a structurally different auxiliary LLM agrees with the primary model; (2) Stochastic Inconsistency - whether repeated sampling from the same model yields divergent outputs; (3) Reasoning Critique - whether an external checker model identifies logical flaws in the chain-of-thought reasoning.

In this crossover trial, 7 clinicians of varying seniority (2 junior residents, 3 senior residents, 2 attending physicians) each review all 110 cases under both standard manual review and SCOUT-assisted review workflows. The study evaluates workflow efficiency (primary endpoint) and diagnostic accuracy (secondary endpoint).

Interventions

• Diagnostic test SCOUT-Assisted Review Workflow
  SCOUT-Assisted Review (Intervention Arm): Physicians review 56 cases processed through the SCOUT framework. For cases classified as low-uncertainty (D(x)=0), the AI prediction is auto-accepted without physician review. For high-uncertainty cases (D(x)=1), the physician reviews the case with access to the main model's chain-of-thought reasoning and the meta-verification audit results. The main model is DeepSeek-V3.1 with chain-of-thought prompting.
• Diagnostic test Standard Manual Review Workflow
  Physicians perform a full manual review of 54 cases using raw medical records with access to the AI model's predictions and reasoning, but without SCOUT uncertainty stratification or selective deferral.

Primary outcome measures

• Mean physician review time per case (minutes) [Time frame: Through study completion, an average of 2 hours.]

Secondary outcome measures (2)

• Diagnostic accuracy (%) [Time frame: Through study completion, an average of 2 hours.]
• Computational Return on Investment (ROI) [Time frame: Through study completion, an average of 2 hours.]

Eligibility criteria

Inclusion criteria

• Board-certified or in-training cardiologists at Fuwai Hospital
• Spanning three experience strata: junior residents, senior residents, attending physicians

Exclusion criteria

• Clinicians involved in the development or optimization of the SCOUT framework
• Clinicians involved in the gold-standard adjudication process

Criteria are shown verbatim from the registry (in English). Final eligibility is always assessed by the study center.

Healthy volunteers: No

Study design

Allocation

Randomized

Model

Crossover

Masking

Open label

Primary purpose

Diagnostic

Study locations

Center list to be confirmed — check the primary protocol.

Identifiers

NCT: NCT07414966 · 2025-2702-1

Primary sources (government registries)

• ClinicalTrials.gov (US NIH)

View this study on ClinicalTrials.gov ↗