REGAIN-ADVOCATE-TA3-RCT

Scalable Agentic AI for Heart Failure & Post-MI Management: A Pragmatic Non-Inferiority Randomized Controlled Trial

(The ADVOCATE Scalability Study)

Protocol Author

Alexey Revtovich

Scientific Lead, Regain Inc. | formerly Rice University

h-index 12 720 citations 19 publications

Google Scholar LinkedIn

1. Project Summary

REGAIN-ADVOCATE-TA3-RCT is a multi-site pragmatic randomized controlled trial designed to evaluate a dual-agent "Clinical AI" system for cardiovascular disease management in the United States, explicitly structured as an ADVOCATE Scalability Study to generate evidence for:

• Clinical non-inferiority
• Operational efficiency
• Technical robustness across EHR vendors/workflows
• Payer-facing economic endpoints

Investigational System

TA1 (Clinical Agent / SaMD): proposes guideline-concordant medication optimization and monitoring plans for heart failure (HFrEF/HFmrEF/HFpEF) and post-myocardial infarction (post-MI) patients, generating actionable orders and follow-up plans.

TA2 (Supervisory Agent / Safety Control): independently monitors TA1 outputs in real time, blocks or escalates unsafe recommendations, and enforces a fail-safe state when safety or system performance degrades.

Primary Objective

The primary objective is to demonstrate non-inferiority of the investigational system to usual care on GDMT adherence (operationalized as the GCTS - Guideline-Concordant Therapy Score) at Month 12, with key supportive endpoints including:

• Rehospitalization or all-cause mortality
• CV death / HF hospitalization
• Patient-reported quality of life (KCCQ)
• Operational efficiency (clinician time per patient)
• Technical integration reliability (read/write success and uptime)
• Economic outcomes (total cost of care per patient per month)
• Adjudicated safety outcomes (agent decision-related SAEs and unsafe recommendation rate)

2. Specific Aims

3. Research Strategy

3a. Significance

Heart failure and post-MI care require longitudinal, guideline-concordant optimization of medications and monitoring, yet specialty capacity is constrained and outcomes remain heterogeneous across settings. A scalable, auditable, safety-controlled agentic AI system could extend specialist-quality management across diverse US health systems, including resource-limited and rural settings, while maintaining patient safety and regulatory-grade traceability.

3b. Innovation

• Dual-agent architecture with independent safety control (TA2) monitoring TA1 in real time
• Auditability ("glass box"): complete trace logs of inputs, model versions, outputs, TA2 decisions, and clinician actions
• Pragmatic EHR-integrated workflow with a "pending order → clinician sign" mechanism supporting deployment realism while preserving clinician accountability

3c. Approach (High-Level Design)

• Phase 1A: retrospective data access + pre-production sandbox integration for read/write validation (orders, In-Basket drafts, note drafts) and IV&V Study 1 support.
• Phase 1B: IRB approval with FWA; beta patients for UI/UX; prospective non-interventional Shadow Mode evidence; IDE activities; IV&V Study 2 support; go/no-go readiness.
• Phase 2 (Live Pragmatic RCT): randomized comparison of investigational system-enabled care versus usual care with blinded endpoint and safety adjudication.
• Safety governance: DSMB + Medical Monitor, pre-specified stopping rules, and a fail-safe state when TA2 or data quality degrades.

4. Study Overview

4.1 Investigational System Definition

The investigational device is the combined TA1+TA2 system integrated into the clinical workflow and EHR. TA2 is treated as an internal safety control. TA2's potential MDDT qualification evidence is developed in parallel, but Phase 2 evaluates the combined system's safety and effectiveness in situ.

4.2 ADVOCATE Schedule (39 Months)

4.3 Change Control / Model Freeze

To preserve interpretability of the RCT and maintain a stable investigational device definition:

PCCP-Style Controlled Updates (IDE-Governed)

Versioning and traceability: Every recommendation and TA2 decision is tied to a unique system version identifier in the audit log.

5. Study Design (Phase 2)

5.1 Trial Type

Multi-site pragmatic RCT, open-label at point of care, with blinded endpoint and safety adjudication.

Regulatory framing (ADVOCATE TA3 requirement): This trial is conducted as an IDE study supporting FDA SaMD authorization for the investigational TA1+TA2 system.

Technical robustness requirement (TA3): The site network will include at least two major EHR vendors (e.g., Epic and Cerner) and demonstrate stable operation across vendor-specific workflows; vendor- and site-level integration metrics are reported.

5.2 Randomization

Individual patient randomization (1:1) within each site with centralized allocation concealment.

Justification: Preserves patient-level causal inference while enabling pragmatic deployment; contamination risk is mitigated via access controls and audit logs.

Contingency: If operationally unavoidable or contamination is excessive, a cluster design at clinician/team level may be adopted with corresponding ICC-driven sample size adjustments and analysis.

Stratification Variables

• Site
• HF phenotype (HFrEF/HFmrEF/HFpEF) vs post-MI cohort
• Baseline GCTS (guideline-concordant therapy score)
• Age (>65 vs ≤65)
• Rural/urban indicator (based on ZIP RUCA)

5.3 Blinding & Adjudication

• Open-label care is expected due to workflow integration.
• Blinded adjudication is required for:
  • Primary endpoint scoring (where subjective elements exist)
  • Device-related serious harms attribution
  • Classification of "unsafe recommendations" and "critical misses"

5.4 Contamination Control & Spillover Analysis

To prevent "spillover" learning from intervention to control:

• Restrict TA1/TA2 UI access to intervention participants (role-based access; EHR flags)
• Segregate order queues and dashboards
• Maintain audit logs of UI access and recommendation viewing
• Train staff on separation and documentation requirements

Contamination Exposure Index (Pre-Specified)

Derive an exposure index per clinician/team and per patient from audit logs (e.g., # of AI-case views, # of intervention pending orders reviewed, time-in-AI UI). Use it for (1) monitoring separation fidelity, and (2) sensitivity analyses.

Operational Separation (Recommended Default)

• Use a dedicated intervention review pool (NP/PA/MD adjudicator queue) where feasible
• Keep intervention dashboards separate from control workflows by role and patient flags

5.5 IDE Sponsor Accountability, Reporting, and Change Control

This study is conducted under an IDE for SaMD. The IDE sponsor (Regain, Inc.) holds device accountability and is responsible for FDA communications, regulatory reporting, and software release control.

Key Principles

• Software change control: the investigational system is frozen per Section 4.3. Any permitted safety-driven changes follow an IDE amendment process and documented CAPA.
• Safety reporting: sites report events rapidly to the IDE sponsor; sponsor performs required FDA/IRB reporting and maintains the Device Master Record and audit trail.
• Regulatory reporting and monitoring: the IDE sponsor fulfills applicable IDE obligations (including 21 CFR 812 reporting expectations).

RACI Matrix

R = Responsible, A = Accountable, C = Consulted, I = Informed

6. Study Population

6.1 Inclusion Criteria (Pragmatic)

• Age ≥18 years receiving longitudinal care at participating US health systems
• Heart failure diagnosis (HFrEF, HFmrEF, or HFpEF), NYHA class II–IV, OR Post-MI within the prior 12 months
• Ability to provide informed consent
• English or Spanish literacy
• Access to smartphone/tablet OR caregiver-assisted support
• Ability to use required wearable/RPM devices (digital scale, BP cuff, SpO2)
• EHR data availability sufficient to support safe medication management (problem list, meds, allergies, labs, vitals)

HF Phenotype Definitions

• HFrEF: typically LVEF ≤40%
• HFmrEF: typically LVEF 41–49%
• HFpEF: typically LVEF ≥50%

LVEF used when available; diagnosis may also be confirmed by clinician problem list / encounter documentation

6.2 Exclusion Criteria (Minimal, Safety-Focused)

Exclude only if there is no safe path to participate even with device provisioning/training and caregiver assistance, or if participation would be clinically inappropriate:

6.3 Recruitment & Enrollment Sources (TA3 Requirement)

Participants may be enrolled through:

• Outpatient clinics: cardiology, primary care, HF programs
• Inpatient settings: prior to discharge after HF hospitalization or acute MI, with longitudinal follow-up arranged within the TA3 health system

6.4 Demographic Representation Targets

6.5 Equity & Representation Operational Plan

To make recruitment targets achievable under real constraints, TA3 execution uses a monitored, adaptive approach:

Site Selection Criteria

• Include at least one safety-net/underserved urban site
• Include at least one rural-serving site
• Prioritize systems with demonstrated Black/African American and Hispanic/Latino HF/Post-MI volume

Adaptive Recruitment Triggers

If any target stratum falls >5 percentage points below plan for two consecutive months:

• Open additional clinics/sites
• Add community outreach
• Increase device provisioning and engagement staffing

Resourcing Tied to Equity KPIs

• Budget and staffing explicitly support translation (EN/ES)
• Device setup/training resources
• Caregiver onboarding
• Navigation support to prevent technology access from becoming an exclusion

7. Interventions & Clinical Workflow

7.1 Control Arm: Usual Care

Standard clinician-led management consistent with local workflows and current AHA/ACC guidelines. Data collection is passive via EHR + PROs.

To support interpretability:

• Document site-level baseline practice patterns
• Pre-specify minimum documentation for GDMT status (med list, doses, contraindications/intolerance)
• Define a minimum measurement-only chart abstraction standard for GDMT eligibility/contraindications at each assessment window (baseline, Month 3, Month 12) to reduce differential misclassification and documentation bias across arms

7.2 Intervention Arm: Investigational System

Core workflow:

1. TA1 ingests EHR data and produces a guideline-concordant optimization plan
2. TA2 independently evaluates TA1 outputs against safety constraints in real time
3. Approved actions are converted into pending orders or structured recommendations
4. A licensed clinician reviews and signs (or rejects/modifies) the pending orders

Rationale Capture (Designed to be Scalable)

Order Review SLAs (Protocol Defaults)

7.3 Training and Credentialing

Mandatory Clinician Training Before Phase 2 Start

• How to review pending orders
• When to override
• Documentation requirements
• Escalation pathways and fail-safe behavior

Ongoing: Periodic refreshers and change-control notifications (without changing frozen model behavior)

7.4 Order Classes Matrix

7.5 Fail-Safe Behavior (System-Level)

If TA2 is unavailable, degraded, or outside performance thresholds, the system must enter a fail-safe state:

Fail-safe exit: System remains in fail-safe until TA2 availability and performance thresholds are restored and verified.

7.6 Staged Autonomy Pathway (Phase 2)

ADVOCATE's goal is to demonstrate safe autonomy at scale, not merely decision support. This protocol pre-specifies a staged autonomy ladder during Phase 2.

Stage A — Run-in (First 4 Weeks)

Stage B — Review-Exception for GREEN Non-Order Actions

Stage C — Optional Limited Pilot (Site- and IDE-Approved)

• Expand review-exception coverage
• Allow limited protocolized actions under explicit standing protocols
• Post-hoc clinician audit sampling and DSMB oversight

Advancement Criteria (Evaluated Per Site Monthly)

Scalability KPIs (Reported Monthly)

8. Outcomes & Endpoints

8.1 Primary Endpoint (Non-Inferiority)

GDMT adherence, operationalized as the Guideline-Concordant Therapy Score (GCTS; 0-4 points) at Month 12.

Primary analysis population: HFrEF/HFmrEF and post-MI participants (HFpEF included in the trial but analyzed as a pre-specified supportive subgroup due to less uniformly defined medication optimization targets).

HFrEF GCTS (4 Pillars)

• RAASi/ARNI (ACEi/ARB/ARNI)
• Evidence-based β-blocker
• MRA
• SGLT2i

GCTS Scoring Framework

Eligibility adjustment: Contraindicated/intolerant pillars are excluded from the denominator.

Post-MI GCTS (4 Elements)

• High-intensity statin (or maximally tolerated)
• Antiplatelet therapy appropriate to time-from-MI and bleeding risk
• β-blocker if indicated
• ACEi/ARB/ARNI if indicated

HFpEF Supportive Therapy Score (HFpEF-STS; 0–2 Points)

Final Score Calculation

GCTS Ascertainment & Documentation-Bias Mitigation

Because the intervention arm may improve documentation quality (not just prescribing), this protocol explicitly separates pragmatic documentation from adjudicated "best-available truth" to prevent biased non-inferiority conclusions.

8.2 Key Supportive Endpoints

• Re-hospitalization or all-cause mortality through Month 15
• CV death / HF hospitalization through Month 15
• Time-to-optimization (time to GCTS ≥3.5)
• Early optimization rate (proportion with GCTS ≥3.0 by Month 3 and Month 6)
• GCTS AUC (0–6 months) - area-under-the-curve to capture speed + maintenance

8.3 Patient-Reported Outcomes

• KCCQ (quality of life) at baseline and follow-up time points (Months 3, 6, 12, 15)

8.4 Operational/Scalability Endpoints

• Clinician time per patient-month (median and p90)
• Specialist Extension Factor (SEF): target ≥5 by Month 3, ≥10 by Month 9
• Response time from red-flag event to clinical action
• Total cost of care (PMPM)

Autonomy-at-Scale KPIs

• AAR (Autonomous Action Rate): % of GREEN non-order actions auto-executed
• BRR (Batched Review Rate): % of GREEN pending orders handled via batched review
• TA2 hard-stop rate per 1,000 recommendations
• Escalation rate per 100 patient-months

Interruptiveness Metrics (Burnout-Relevant)

• Interruptive alerts/pages per 100 patient-months
• Non-interruptive queue items per 100 patient-months
• Median time-to-disposition for each class

Cost / Reimbursement Evidence (TA3 Requirement)

• Total cost of care per patient per month (PMPM)
• Using claims feeds where available OR standardized cost weights derived from utilization
• At least one participating site will provide claims linkage (e.g., Medicare FFS/MA, ACO)

Budget Impact Analysis (Payer-Grade)

• Gross savings from reduced admissions/ED utilization
• Incremental program costs (devices/data plans, integration, adjudication time)
• Net PMPM
• Breakeven month

8.5 Patient Medication-Taking Adherence (Secondary/Mediator)

Because the TA3 "GDMT adherence" effectiveness endpoint is operationalized as guideline-concordant prescribing/optimization (GCTS), separate patient medication-taking adherence is treated as a secondary/mediator endpoint and measured via:

• Pharmacy claims/fills (when available)
• EHR medication reconciliation
• Optional ePill devices and/or validated self-report

8.6 Safety Endpoints

• Device-related serious adverse events (adjudicated)
• Unsafe recommendation rate (adjudicated)
• Agent decision-related SAEs: <3% target
• TA2 performance: critical miss rate, false positive block rate

Hallucination/Invalid-Reasoning Metrics (Reportable)

8.7 Sample Size & Statistical Analysis

Final planned sample size: N = 800 total participants (400 per arm)

Non-Inferiority Margin

Δ = -0.20 points on the 0-4 GCTS scale. The investigational system is non-inferior if the lower bound of the one-sided 97.5% CI is greater than -0.20.

Base NI Calculation

• Endpoint SD (planning): σ = 1.0 GCTS points (conservative; refined using Phase 1B data)
• NI margin: Δ = 0.20
• One-sided α = 0.025; power = 90%

Inflations Applied

Rounding to 400 per arm provides margin for heterogeneity and improves precision for key supportive event endpoints and subgroup analyses.

Enrollment Balance Targets

9. Schedule of Assessments (Phase 2)

Assessment Windows

Continuous Event Capture

All suspected endpoint events are adjudicated.

10. Statistical Analysis Plan (SAP) Summary

10.1 Estimands and Analysis Sets

Analysis Sets: Both ITT and Per-Protocol non-inferiority analyses are required, with expectation of consistent conclusions.

10.2 Primary Analysis Model

Mixed effects regression (or GEE) appropriate to endpoint scale, with:

• Fixed effects: arm, baseline Adjudicated GCTS, cohort (HF vs post-MI), site, stratification variables
• Random effects: clinician/team if needed and/or site-level random intercepts
• Robust standard errors

10.3 Multiplicity and Hierarchy

Confirmatory Family (Gatekept)

10.4 Missing Data

• Primary analysis: mixed models with maximum likelihood under MAR assumptions, supported by multiple imputation with auxiliary variables
• Sensitivity: pattern-mixture (delta-adjustment) and worst-case bounds for differential missingness

10.5 Subgroup Analyses (Pre-Specified)

• Age >65 vs ≤65
• Sex
• Race/ethnicity
• Rural/urban
• HF phenotype (HFrEF vs HFmrEF vs HFpEF) vs post-MI
• CKD strata

11. Data & Safety Monitoring / Stopping Rules

11.1 Governance

11.2 Definitions

Hallucination Taxonomy

• Fabricated data claims (labs, vitals, medications) not present in EHR/RPM feed
• Wrong-patient-context inference
• Guideline mismatch / non-concordant recommendation given available facts
• Missing-data hazard (proceeds as if required safety data exist)

11.3 Phase 1B Go/No-Go Thresholds

11.4 Phase 2 Stopping Rules

Patient-Level

Remove from autonomous mode if:

• ≥2 high-severity TA2 blocks within 30 days, OR
• ≥1 confirmed critical miss, OR
• Any probable/definite device-related serious harm

Trial-Level

System-Level (Fail-Safe)

Automatic fail-safe if:

• TA2 unreachable for >5 seconds, OR
• TA2 p99 latency >250ms sustained for >5 minutes, OR
• Required data-quality guardrails fail

Recurrent fail-safe events (>3 in 24 hours) trigger incident review and DSMB notification.

11.5 Escalation Protocols (24/7 Coverage)

Red Flag Triggers (Minimum Set)

• Rapid weight gain (≥2–3 kg in 72 hours) with HF symptoms
• New/worsening hypoxia (SpO2 below threshold) or severe dyspnea
• Hypotension below threshold with symptoms
• ADT feed indicating ED visit/admission for HF-related complaints

Required Behavior: TA1 drafts In-Basket message and/or pages on-call clinician immediately. TA2 validates escalation urgency and blocks inappropriate autonomous action. Sites provide 24/7 coverage via existing on-call systems.

11.6 Event Classification & Reporting Workflow

12. Technology & Integration Requirements

12.1 EHR Integration (FHIR R4)

Read Access (Real-Time)

• Labs (chemistry, hematology, BNP and troponin)
• Vitals (BP, HR, weight, O2/SpO2)
• Medications (current active list)
• Clinical notes (cardiology, primary care)
• ADT feeds (admission, discharge, transfer)

Write Access (Real-Time)

• Draft In-Basket / inbox messages to clinical team (required)
• Draft scheduling requests / follow-up tasks
• Create pending orders (meds/labs) for clinician sign-off
• Draft documentation / encounter notes for clinician review/signature (required)

Audit Logs (Required)

All TA1/TA2 inputs/outputs, model versions, gating decisions, timestamps, clinician actions, override reasons, downstream order execution status.

12.2 Standards, Interoperability, and Auth

12.3 Phase 1A Data Access

• Retrospective data: de-identified longitudinal EHR data for HF/Post-MI cohorts
• Connected wearable/RPM platforms: de-identified historical feeds and pre-production access
• Pre-production/sandbox: validate API writes without patient risk
• Pre-Phase-2 qualification: each site must pass Integration Qualification Checklist (Appendix B)

12.4 Fusion Protocol Test (TA2 Gating Verification)

Pre-Phase 2 verification:

• Test harness injects known unsafe scenarios across error classes
• Confirm TA2 blocks/escalates per spec
• Confirm system enters fail-safe when TA2 unavailable or outside constraints

12.5 Performance Requirements

12.6 Downtime / Failover SOP

Required procedures for:

• EHR downtime
• TA2 downtime
• Missing/degraded data quality
• Cybersecurity incidents

13. Ethics, Consent, and Privacy

• IRB approval at each site
• Informed consent includes:
  • Description of investigational system and clinician sign-off workflow
  • Data use, audit logs, and privacy protections
  • Explicit disclosure that Phase 1B shadow mode does not change care
• Data handling: HIPAA-aligned; role-based access; audit logs retained per protocol
• eConsent/e-sign: implemented with integrity controls appropriate to environment, consistent with 21 CFR Part 11 expectations where applicable

14. Timeline & Milestones

Phase 1A: Discovery & Foundation (Months 0–12)

Phase 1B: Preparation & Regulatory (Months 12–24)

Phase 2: Scalability Study Execution (Months 24–39)

15. Budget Justification (Summary)

Cost Categories

TA3 Budgeting Categories (Spec-Aligned)

• Per-patient costs (recruitment, enrollment, device provisioning)
• IT integration costs (interface engine/vendor fees; integration staff time)
• Clinical staff research time (adjudication and documentation)
• Administrative overhead (IRB fees, grant management)

16. Data Management, Monitoring, and Quality Assurance

Data Sources

• EHR (FHIR R4 and/or HL7 v2)
• Order-signing logs and In-Basket message logs
• Audit logs (TA1/TA2)
• Connected wearable/RPM platform data
• PROs (KCCQ)
• Claims feeds (required for ≥1 site)

Auditability

All TA1/TA2 inputs/outputs and clinician actions captured with timestamps, versions, and unique identifiers. Logs are immutable and retained per protocol.

Data Integrity Controls

• Role-based access
• Encryption in transit/at rest
• Separation of duties between engineering and adjudication
• Periodic log review for anomalies

Monitoring Plan

• Risk-based monitoring with centralized data checks (missingness, outliers, protocol deviations)
• Site monitoring for consent and endpoint ascertainment

Quality Management

• Pre-Phase 2 validation (Fusion Protocol tests, downtime drills)
• SOPs for incident response
• Documented CAPA

16.1 Data Management and Sharing Plan (DMSP)

What Is Shared

• Aggregated endpoint summaries
• De-identified audit-log extracts for IV&V
• Adjudication labels (de-identified)
• Integration reliability metrics by site/vendor
• Challenge-set and fusion protocol test reports

Cadence

• Monthly operational/technical dashboards during Phase 2
• Quarterly curated de-identified datasets for IV&V

De-Identification

• HIPAA-aligned (safe harbor or expert determination)
• Tokenization/pseudonymization for linkage
• CUI handling for sensitive artifacts

17. TA3 Management, Collaboration, and Site Eligibility

Required Roles (Clinician-in-the-Loop Team)

Required Governance Capabilities

• Site IRB has Federal Wide Assurance (FWA) and AI/SaMD review capacity
• Participation in independent DSMB
• 24/7 escalation coverage via existing on-call systems

Collaboration and IV&V

• TA3 sites collaborate with IV&V Partner on evaluation metrics
• Participate in IV&V Study 1 and Study 2 per program schedule

IP Boundary (Program Requirement)

• Hospital/TA3 site owns the clinical data
• Regain/Prime owns the AI models (TA1/TA2)

17.1 Dealbreakers (Ineligibility Factors)

TA3 proposals are rejected if they:

18. References (Selected)

• ICH E6(R2): Guideline for Good Clinical Practice
• ICH E9(R1): Addendum on Estimands and Sensitivity Analysis
• CONSORT-AI and SPIRIT-AI reporting guidelines for clinical trials involving AI interventions
• AHA/ACC/HFSA heart failure guideline (contemporary version) for GDMT definitions and target dosing references
• Contemporary ACC/AHA guidance on secondary prevention after myocardial infarction for post-MI therapy elements

Appendix A: TA3 Traceability Matrix

55-row mapping table linking each requirement from TA3 Official Specs (v1.2) to the corresponding section in this protocol, with notes and required evidence artifacts.

Full 55-row matrix available in protocol source document.

Appendix B: Integration Qualification Checklist

Each participating TA3 site must complete this checklist in pre-production and re-validate in production prior to first enrollment.

Multi-vendor requirement: Checklist completed for each EHR vendor in the TA3 network.

Appendix C: Workflow Diagrams

C.1 Care Loop and Safety Gating

EHR + Wearables/RPM + Patient Inputs
            |
            v
     TA1 Clinical Agent
 (draft plan + pending orders)
            |
            v
     TA2 Supervisory Agent
 (approve | hard-stop | escalate)
            |
     +------+------+------------------+
     |             |                  |
     v             v                  v
PEND in EHR   BLOCK + Escalate   DATA-UNCERTAIN
(GREEN/YELLOW)   (urgent queue)     (needs remediation)
     |
     v
Clinician action (sign / modify / reject)
     |
     v
EHR executes + patient/team notified + immutable audit log
                        

C.2 Offline Improvement Loop

Without Contaminating Phase 2:

Clinician decisions + structured rationales
                |
                v
      Label set (accept/reject/modify + reason codes)
                |
                v
 Offline analysis/training for next release (Phase 1B / post-trial)
                |
                v
Versioned release candidate
 (frozen during Phase 2 unless safety-driven IDE amendment)