Targeted Artificial Intelligence Language mOdel Recommendations for Heart Failure (TAILOR-HF)

Overview

This study is an investigator-initiated, cluster-randomized implementation trial evaluating a large language model (LLM)-based clinical decision support (CDS) tool designed to improve guideline-directed medical therapy (GDMT) for adult patients with heart failure seen in outpatient cardiology clinics at Mass General Brigham. For eligible heart failure encounters, the CDS tool reviews existing electronic health record (EHR) data, including diagnoses, medications, vital signs, laboratory results, and recent notes, and generates brief, clinician-facing messages suggesting opportunities to initiate or optimize GDMT and highlighting relevant safety considerations. Messages are delivered to cardiology providers via Epic InBasket and/or institutional email prior to scheduled visits. The tool is advisory only and cannot place orders or change medications automatically; all treatment decisions remain at the discretion of the treating clinician and patient. Cardiology providers are assigned at the provider/clinic level to early implementation of the CDS tool versus usual care (no messages) during the initial phase. The primary outcome is GDMT optimization within 30 days of an index visit. Secondary outcomes include feasibility of CDS generation and delivery and a 30-day safety composite (e.g., heart failure hospitalization, acute kidney injury, hyperkalemia, hypotension or bradyarrhythmia plausibly related to GDMT).

Overview and Rationale Guideline-directed medical therapy (GDMT) for heart failure reduces hospitalizations and mortality, yet substantial underuse and suboptimal titration persist in routine practice, even in specialty cardiology clinics. Barriers include limited visit time, complex comorbidities, fragmented information across notes and structured data, and uncertainty about contraindications or prior intolerance. Electronic clinical decision support (CDS) tools that synthesize key patient information and highlight GDMT opportunities at the point of care may help close these gaps. Large language models (LLMs) can read both structured EHR data (e.g., diagnoses, medications, vital signs, laboratory results) and unstructured narrative notes to generate nuanced, patient-specific recommendations. We developed an LLM-based CDS tool that reviews an adult heart failure patient's EHR and produces a brief, free-text message to the treating cardiology provider summarizing heart failure status, suggesting potential GDMT changes consistent with contemporary guidelines, and flagging relevant safety issues (e.g., low blood pressure, bradycardia, renal dysfunction, hyperkalemia, prior documented intolerance). In retrospective testing, the tool's recommendations were generally concordant with expert clinician judgment. Study Design This is an interventional, cluster-randomized, provider-level trial conducted in adult outpatient cardiology clinics at Mass General Brigham. The intervention is a software-only, investigational clinical decision support device ("LLM-GDMT Clinical Decision Support Tool"). Eligible cardiology attendings and advanced practice providers are assigned at the provider/clinic level to one of two parallel arms during the initial phase: Early Implementation - LLM-GDMT CDS: Providers in this arm receive LLM-generated, clinician-facing messages for eligible heart failure encounters. For scheduled visits that meet predefined inclusion criteria, the tool reviews existing EHR data and generates a brief advisory message that is delivered via Epic InBasket and/or institutional email within the week prior to the visit. Usual Care (Delayed Implementation): Providers in this arm continue usual care and do not receive LLM-generated messages during the initial evaluation phase. EHR data are used to compute quality metrics for comparison. After the initial evaluation, the CDS tool may be expanded to these providers as part of routine care. Patients are not contacted for the study. All clinical decisions, including whether to start, stop, or adjust any medication, remain entirely at the discretion of the treating clinician in partnership with the patient. The CDS messages are advisory only and cannot place orders or directly change medications or monitoring plans. Population and Eligibility The study includes adult patients (age ≥18 years) with a documented heart failure diagnosis who are scheduled for outpatient visits with participating cardiology providers at Mass General Brigham clinics. Additional inclusion criteria require evidence supporting active or prior heart failure (e.g., diagnosis codes, loop diuretic use, echocardiographic findings, or documentation of heart failure signs or symptoms) and at least one prior cardiology visit. Exclusion criteria include hemodynamic instability (e.g., very low blood pressure or heart rate), advanced renal dysfunction below a specified estimated glomerular filtration rate threshold, selected advanced structural heart disease (e.g., cardiac amyloidosis, hypertrophic cardiomyopathy, heart transplant or left ventricular assist device recipients, severe valvular disease), and encounters in adult congenital heart disease clinics. Intervention and Workflow For eligible encounters in the early-implementation arm, the CDS tool operates within Mass General Brigham's secure technical environment (Epic, enterprise data warehouse, and Azure OpenAI within the MGB tenant). The tool retrieves relevant structured and unstructured EHR data, uses a large language model to synthesize this information, and generates a brief, human-readable message. The message typically includes: * Confirmation of current heart failure status and key comorbidities. * A summary of current GDMT and potential opportunities for initiation or titration of evidence-based therapies, taking into account recent vital signs, laboratory values, and documented contraindications or prior intolerance. * Safety considerations (e.g., blood pressure, heart rate, renal function, potassium) relevant to GDMT changes. Messages are delivered to the treating provider via Epic InBasket and/or institutional email in advance of the visit. The tool does not write orders, modify medication lists, or send any direct communication to patients. Providers may choose to use, modify, or ignore the suggestions based on their clinical judgment and patient preferences. Outcomes and Analysis The primary outcome is GDMT optimization within 30 days of the index visit, defined as initiation of at least one new GDMT class not previously prescribed and/or uptitration of at least one existing GDMT medication in eligible patients. Secondary outcomes include: (1) feasibility and fidelity of CDS implementation (e.g., proportion of eligible encounters for which messages are successfully generated and delivered), (2) a 30-day safety composite that includes heart failure hospitalizations, emergency department visits related to decompensated heart failure, acute kidney injury, hyperkalemia above predefined thresholds, and clinically significant hypotension or bradyarrhythmia plausibly related to GDMT, and (3) provider-reported acceptability and perceived usefulness, measured via an optional anonymous survey. The study anticipates including up to 2,500 unique adult heart failure patients across participating clinics during the implementation period. Analyses will account for clustering at the provider level and will compare GDMT optimization and safety outcomes between early-implementation and usual-care arms during the initial evaluation phase. The overall goal is to determine whether an advisory, LLM-based CDS tool can be implemented safely and feasibly in routine outpatient cardiology practice and whether it improves uptake of guideline-directed heart failure therapies.