Echocardiography-Based, Artificial Intelligence-Enabled Electrocardiography (AI-ECG) for Diastolic Hemodynamics Phenotyping in Acute Heart Failure (AHF)

This study demonstrates that in a large cohort of acute heart failure patients, artificial intelligence-enabled electrocardiography (AI-ECG) provides universally feasible, scalable diastolic function grading that correlates with hemodynamic severity and independently predicts mortality and rehospitalization, offering a pragmatic solution when echocardiographic assessment is indeterminate or unavailable.

The Gist

The Big Idea: Turning a Simple Heartbeat into a Crystal Ball

Imagine you have a car. Usually, to know if the engine is overheating or if the transmission is slipping, you need a mechanic to pop the hood, hook up a computer, and run a deep diagnostic. That's like an Echocardiogram (an ultrasound of the heart). It's the gold standard, but it's expensive, takes time, and sometimes the "windows" to see inside are too blurry to get a clear picture.

Now, imagine if you could just listen to the car's engine sound for a few seconds and instantly know exactly how much stress the engine is under, even if the mechanic couldn't see inside. That is what this study is about.

The researchers developed an Artificial Intelligence (AI) system that looks at a standard 12-lead Electrocardiogram (ECG)—the little heart rhythm strip you get in the ER or at the doctor's office—and uses it to "see" the hidden pressure inside the heart.

The Problem: The "Black Box" of Heart Failure

When people are hospitalized with Acute Heart Failure (AHF), their hearts are struggling to pump or relax properly. A major cause of this is diastolic dysfunction—basically, the heart is too stiff to fill up with blood, causing a backup of pressure (like a clogged drain).

• The Old Way: Doctors try to measure this pressure using an ultrasound (Echo). But in about 44% of patients, the ultrasound is too blurry or the heart is beating too fast to get a clear answer. It's like trying to take a photo of a race car through a foggy window; you just can't tell how fast it's going.
• The New Way: The AI looks at the electrical signal of the heart (the ECG). Even if the ultrasound is blurry, the electrical signal is always clear. The AI acts like a super-smart detective who can read the "clues" in the electrical waves that human doctors miss.

What They Did

The team tested this AI on 11,513 patients who were hospitalized for heart failure across many different hospitals in the US.

1. They fed the AI the patients' ECGs.
2. The AI gave each patient a "Grade" of how stiff their heart was (Normal, Grade 1, Grade 2, or Grade 3).
3. They compared these AI grades to what actually happened to the patients over the next few years.

The Big Discoveries

1. The AI Never Gives Up
While the ultrasound failed to give a clear answer in nearly half the patients, the AI gave a grade for 100% of them. It works even when the "foggy window" problem exists.

2. The Grades Match the Danger
The AI's grades were spot-on.

• Grade 1 (Normal/Good): These patients had lower stress markers and lived longer.
• Grade 3 (Severe/Stiff): These patients had much higher pressure in their hearts. They were sicker, had more comorbidities, and were much more likely to die or be readmitted to the hospital.
• Analogy: Think of the AI grades like a weather forecast. A "Grade 3" prediction is like a "Category 5 Hurricane Warning." The patients with this grade were indeed in a storm, while "Grade 1" was just a sunny day.

3. It Spots the "Hidden" Danger
Here is the most surprising part: The length of time patients stayed in the hospital was the same for everyone, regardless of their AI grade.

• The Metaphor: Imagine two people leave a hospital. One has a broken leg (obvious danger), and the other has a ticking time bomb in their chest (hidden danger). Both leave after 5 days. The doctor sees the broken leg and knows to be careful. But the AI saw the ticking time bomb in the second patient's heart rhythm.
• The study found that patients with "hidden" severe stiffness (Grade 3) were just as likely to die as those with obvious heart failure, even though they were discharged at the same time. The AI helps doctors spot these "ticking time bombs" before they go home.

4. The "Second Look" (Follow-up)
The researchers also looked at patients who got a second ECG after they went home.

• If the AI showed the heart pressure was improving (going down), the patient had a much better chance of survival.
• If the pressure stayed high or got worse, the risk remained high.
• Analogy: It's like checking the oil light in your car after a repair. If the light turns off, you're good to go. If it stays on, you need to go back to the shop. The AI allows doctors to check this "oil light" using a simple ECG at a follow-up visit.

Why This Matters

This isn't just about fancy technology; it's about saving lives by seeing what was previously invisible.

• It's everywhere: Every hospital has ECG machines. They are cheap, fast, and done on almost everyone.
• It's a safety net: It catches the patients who are sicker than they look, ensuring they get the right care and follow-up.
• It's a new tool: It doesn't replace the ultrasound; it complements it. When the ultrasound is blurry, the AI is the flashlight that helps the doctor see the truth.

The Bottom Line

This study shows that by using AI to listen to the heart's electrical song, we can predict how stiff the heart is and how likely a patient is to survive, even when traditional tests fail. It turns a routine, 10-second heartbeat check into a powerful tool for spotting hidden heart failure risks, helping doctors treat patients more effectively before they leave the hospital.

Technical Summary

1. Problem Statement

Acute Heart Failure (AHF) is characterized by significant heterogeneity in diastolic hemodynamics, particularly regarding Left Ventricular (LV) filling pressures (FP) and diastolic function (DF).

• Clinical Gap: While echocardiography is the reference standard for assessing DF and FP, it is often operationally infeasible to perform universally in acute settings. In large cohorts, guideline-based echocardiographic DF grading is indeterminate in at least 20–44% of patients due to factors like tachycardia, atrial fibrillation, or poor acoustic windows.
• Symptom Limitations: Symptom-based triage (reliance on dyspnea) is unreliable, as many AHF patients present with fatigue or edema rather than dyspnea, leading to under- or over-diagnosis.
• Need: There is a critical need for an objective, scalable, non-invasive, and universally available biomarker to assess diastolic hemodynamics and stratify risk in AHF patients, especially when echocardiography is unavailable or inconclusive.

2. Methodology

Study Design & Population:

• Type: Retrospective, multisite cohort study.
• Cohort: 11,513 adults hospitalized for AHF across Mayo Clinic sites (Minnesota, Arizona, Florida) between 2013 and 2023.
• Inclusion Criteria: Hospitalization $\ge$24 hours, clinically documented AHF (ICD coding), receipt of $\ge$1 dose of IV loop diuretic, and paired 12-lead ECG and Transthoracic Echocardiogram (TTE).
• Exclusion: Patients included in the original AI-ECG model derivation/validation cohorts.

AI-ECG Model:

• Architecture: A modified ResNet-18 Convolutional Neural Network (CNN) adapted for ECG waveforms.
• Input: 10-second, 12-lead ECGs sampled at 500 Hz.
• Training: Originally trained on >200,000 paired ECG-TTE studies using the Mayo Clinic 2020 algorithm (aligned with 2016 ASE/EACVI guidelines).
• Output:
  1. Discrete DF Grades: Normal, Grade 1, Grade 2, or Grade 3 (assigned based on the highest probability class).
  2. Continuous Filling Pressure (FP) Probability: Calculated as $P(FP) = P(\text{Grade 2}) + P(\text{Grade 3})$, ranging from 0 to 1.

Data & Analysis:

• Reference Standard: Echocardiographic DF grading (where determinate) and invasive Pulmonary Capillary Wedge Pressure (PCWP) from right-heart catheterization.
• Outcomes: Primary outcome was all-cause mortality; secondary outcome was HF rehospitalization.
• Statistical Approach:
  • ROC analysis for model performance against TTE.
  • Multivariable Cox proportional hazards models adjusted for demographics, comorbidities, LVEF, and medications.
  • Interaction analyses for subgroups (LVEF, rhythm, echo determinacy).
  • Longitudinal analysis of serial ECGs (7–180 days post-discharge) to assess changes in FP probability and their impact on mortality.

3. Key Contributions

1. Universal Feasibility: Demonstrated that AI-ECG can provide a diastolic function grade in 100% of patients, whereas guideline-based echocardiography was indeterminate in 44% of clinically eligible patients.
2. Hemodynamic Phenotyping: Validated that AI-ECG grades correlate strongly with established severity markers (NYHA class, NT-proBNP, MAGGIC scores) and invasive PCWP, effectively stratifying patients into distinct hemodynamic phenogroups.
3. Independent Prognostic Value: Showed that AI-ECG DF grades provide independent prognostic information for mortality and rehospitalization beyond traditional risk factors (including MAGGIC score), even in patients with indeterminate echocardiograms.
4. Longitudinal Monitoring: First demonstration that serial AI-ECG measurements can track post-discharge hemodynamic recovery. Improvement in FP probability was independently associated with a 15% reduction in mortality.
5. Workflow Integration: Proposed a pragmatic solution using the ubiquitous 12-lead ECG to identify "occult" high-risk patients who might otherwise be discharged prematurely despite elevated filling pressures.

4. Key Results

• Model Performance: In the subset with determinate TTE ($n=2,582$), the AI-ECG continuous FP probability discriminated TTE Grade 2–3 dysfunction with an AUC of 0.85 (95% CI 0.83–0.86).
• Correlation with Severity:
  • Higher AI-ECG grades correlated with older age, higher comorbidity burden, elevated NT-proBNP, and worse NYHA class.
  • Invasive Validation: Mean PCWP increased progressively across grades (Normal/Grade 1: 18 mmHg; Grade 2: 19.8 mmHg; Grade 3: 22.9 mmHg; $P<0.001$).
• Prognostic Outcomes:
  • Mortality: After multivariable adjustment, Grade 2 and Grade 3 AI-ECG DF were independently associated with increased mortality compared to Normal/Grade 1 (HR 1.25 and 1.44, respectively).
  • Risk Stratification: Combining AI-ECG DF with MAGGIC scores created ordered risk gradients. Patients with high MAGGIC scores and Grade 2–3 DF had the highest mortality (HR 3.4 vs. reference).
  • Length of Stay: Despite vastly different hemodynamic burdens and outcomes, hospital length of stay did not differ significantly across AI-ECG grades, suggesting advanced diastolic dysfunction is often unrecognized in routine care.
• Longitudinal Findings:
  • Patients with Grade 2/3 DF showed a decrease in estimated FP probability post-discharge (treatment response).
  • Patients with Normal/Grade 1 DF showed a paradoxical increase in FP probability, suggesting unresolved interstitial congestion.
  • Improvement in FP probability (post-discharge) was associated with a 15% lower mortality risk (HR 0.85), while worsening did not show a consistent adverse gradient beyond baseline risk.

5. Significance and Clinical Implications

• Bridging the Diagnostic Gap: AI-ECG transforms the standard 12-lead ECG into an actionable hemodynamic biomarker, addressing the critical gap when echocardiography is indeterminate or unavailable.
• Risk Stratification: It identifies a high-risk "occult" population (Grade 3 DF) with median survival of only 2.8 years—comparable to high-risk chronic HF trial populations—who may benefit from intensified follow-up or specialist referral.
• Post-Discharge Management: The ability to track FP probability trajectories via serial ECGs offers a novel method for monitoring hemodynamic recovery, potentially identifying patients who need intensified diuresis or therapy adjustment before readmission.
• Scalability: As a software-based solution requiring no new hardware, it offers a scalable, low-cost method to integrate advanced hemodynamic phenotyping into routine acute care workflows globally.

Limitations: The study is retrospective and single-system (though geographically diverse). Echocardiographic data was required for inclusion, potentially limiting generalizability to patients who never receive an echo. Longitudinal findings are exploratory and require prospective validation.