Mayo Clinic Deploys AI That Detects Heart Failure 48 Hours Before Onset

Published 2025-04-25 · Health AI

The Mayo Clinic has deployed a deep learning model that predicts acute decompensated heart failure up to 48 hours before clinical onset, giving physicians a critical window for early intervention. The system, which analyses routine 12-lead electrocardiogram data collected during standard hospital visits and inpatient monitoring, has been live across all three Mayo Clinic campuses — Rochester, Jacksonville, and Scottsdale — since February 2025.

The deployment represents one of the largest-scale applications of predictive AI in cardiology to date, processing approximately 4,200 ECG recordings per day and flagging roughly 60 patients weekly for heightened surveillance. Early results, presented at the American College of Cardiology's annual scientific session in March 2025, show that the system has reduced in-hospital heart failure mortality by 14% in flagged patients compared with historical controls.

The Model Behind the Prediction

The system, internally named ECG-HF-Net, is a multi-lead convolutional neural network trained on 2.4 million ECG recordings collected from 1.1 million unique patients across the Mayo Clinic health system between 2005 and 2023. The training dataset includes both 10-second standard ECGs and continuous telemetry strips from hospitalised patients, with outcomes linked through the Mayo Clinic's integrated electronic health record system.

What distinguishes ECG-HF-Net from earlier predictive models is its ability to detect subtle morphological changes in the ECG waveform that precede clinical heart failure decompensation but fall below the threshold of what even experienced cardiologists would flag as abnormal. The model identifies micro-alterations in QT dispersion, ST-segment vector orientation, and P-wave terminal force — features that individually carry minimal diagnostic weight but, when analysed in combination by a deep neural network, form a reliable early-warning signal.

Dr. Peter Noseworthy, a cardiac electrophysiologist at Mayo Clinic and co-principal investigator on the project, explained the clinical significance: "By the time a patient develops overt symptoms of heart failure decompensation — shortness of breath, fluid retention, elevated jugular venous pressure — the window for the most effective interventions has often narrowed considerably. If we can identify patients 48 hours earlier and initiate diuretic adjustment, fluid restriction, or closer monitoring, we can frequently prevent the admission altogether or at least reduce its severity."

Validation and Performance Metrics

The model was validated on a prospective cohort of 38,000 patients admitted to Mayo Clinic hospitals between January and December 2024 — data that was entirely held out from the training set. In this validation cohort, ECG-HF-Net achieved an area under the receiver operating characteristic curve (AUROC) of 0.89 for predicting heart failure decompensation within 48 hours, with a sensitivity of 82% at a specificity threshold calibrated to maintain a manageable false-positive rate.

At the operating point selected for clinical deployment — designed to produce approximately one false positive for every three true positives — the model's positive predictive value was 27%. While this may appear modest, it represents a substantial improvement over existing clinical prediction rules. The widely used Get With The Guidelines-Heart Failure risk score, which relies on manually entered clinical variables, achieves a positive predictive value of approximately 12% in similar populations.

The model performed consistently across major demographic subgroups. AUROC values were 0.88 for women and 0.89 for men, 0.87 for patients over 75, and 0.90 for patients under 65. Performance was slightly lower for Black patients (AUROC 0.85), a gap the research team attributes to the underrepresentation of Black patients in the Mayo Clinic health system's historical records — Black patients comprised 5.2% of the training dataset, compared with approximately 13% of the US population.

Integration into Clinical Workflow

The system is integrated directly into Mayo Clinic's ECG management platform, which is itself embedded within the Epic electronic health record system. When a patient's ECG triggers a positive prediction, an alert appears in the treating physician's inbox alongside the ECG waveform, the model's confidence score, and a brief explanation highlighting which ECG features contributed most strongly to the prediction.

The alert is advisory, not directive. Physicians retain full clinical autonomy and can dismiss alerts with a documented reason. During the first three months of deployment, approximately 58% of alerts resulted in a documented clinical action — typically ordering a BNP laboratory test, adjusting diuretic dosage, or requesting a cardiology consultation. The remaining 42% were dismissed, most commonly because the patient was already receiving appropriate heart failure management or because the physician judged the clinical context to warrant watchful waiting.

This override rate is intentional. Dr. Paul Friedman, chair of Mayo Clinic's Department of Cardiovascular Medicine and the project's senior investigator, has argued that a high override rate is preferable to alert fatigue: "If the system generates too many alerts and clinicians start ignoring them reflexively, we have solved nothing. We deliberately tuned the threshold to produce actionable alerts rather than exhaustive coverage. The 58% action rate tells us that clinicians find the information useful more often than not."

Regulatory and Ethical Considerations

ECG-HF-Net operates under Mayo Clinic's existing institutional review board protocols for clinical decision support tools. Because the system does not independently make clinical decisions — it provides information that physicians incorporate into their own judgement — it does not require FDA approval as a medical device. However, Mayo Clinic has voluntarily submitted the model's validation data to the FDA for review under the agency's Digital Health Center of Excellence, in anticipation of potential future guidance on clinical decision support software.

The deployment also raises questions about liability. If a physician overrides a correct alert and the patient subsequently deteriorates, who bears responsibility? Mayo Clinic's legal team has taken the position that the model is a decision-support tool, analogous to a laboratory test result, and that clinical responsibility remains with the treating physician. This position is consistent with current FDA guidance, but it has not been tested in litigation.

Scalability and External Validation

Mayo Clinic is planning a multi-site external validation study in partnership with the Cleveland Clinic, Massachusetts General Hospital, and the University of Tokyo, with enrolment expected to begin in late 2025. External validation is critical because models trained on data from a single health system often exhibit performance degradation when deployed in different patient populations with different disease prevalence patterns, different ECG acquisition protocols, and different coding practices.

The team is also exploring whether the model can be extended to predict other cardiovascular events, including atrial fibrillation recurrence after ablation, ventricular tachycardia episodes in patients with implantable cardioverter-defibrillators, and acute coronary syndrome in emergency department presentations. Early internal results for atrial fibrillation prediction are promising, with an AUROC of 0.84 in a preliminary dataset of 180,000 ECGs.

If external validation confirms the model's performance, Mayo Clinic intends to license the technology to other health systems through its Mayo Clinic Platform digital health initiative. The team has emphasised that any external deployment would require site-specific calibration to account for population differences and would need to comply with the EU AI Act's requirements for high-risk AI systems in European markets.

For more on AI applications in clinical diagnostics, visit our AI in Health research repository. Related coverage includes AlphaFold 3's drug-target interaction predictions and NVIDIA BioNeMo's protein design acceleration.

← Back to Publications