Mayo Clinic's AI: ECG Heart Failure Detection & GenAI Search

The Challenge

As a leading academic medical center, Mayo Clinic manages millions of patient records annually, but early detection of heart failure remains elusive. Traditional echocardiography detects low left ventricular ejection fraction (LVEF <50%) only when symptomatic, missing asymptomatic cases that account for up to 50% of heart failure risks.[1] Clinicians struggle with vast unstructured data, slowing retrieval of patient-specific insights and delaying decisions in high-stakes cardiology.

Additionally, workforce shortages and rising costs exacerbate challenges, with cardiovascular diseases causing 17.9M deaths yearly globally. Manual ECG interpretation misses subtle patterns predictive of low EF, and sifting through electronic health records (EHRs) takes hours, hindering personalized medicine.[2] Mayo needed scalable AI to transform reactive care into proactive prediction.

Quantitative Results

• ECG AI AUC: 0.93 (internal), 0.92 (external validation)
• Low EF detection sensitivity: 82% at 90% specificity
• Asymptomatic low EF identified: 1.5% prevalence in screened population
• GenAI search speed: 40% reduction in query time for clinicians
• Model trained on: 1.1M ECGs from 44K patients
• Deployment reach: Integrated in Mayo cardiology workflows since 2021

Implementation Details

ECG Machine Learning Algorithm Development

Mayo Clinic's Department of Artificial Intelligence and Informatics spearheaded the ECG AI project starting in 2018. Researchers, led by Dr. Paul Friedman and Zachi Attia, collected a dataset of 1.1 million ECGs from 44,000 unique patients, labeling LVEF via paired echocardiograms. A convolutional neural network (CNN) was trained to predict binary low LVEF (<50%), achieving AUC 0.93 internally and 0.92 on external Mayo data, outperforming prior models.[1] FDA clearance was pursued, with deployment in clinical workflows by 2021 for risk stratification.

Challenges like data imbalance (low EF rare) were addressed via oversampling and augmentation. Integration with EHRs used Explainable AI (XAI) techniques for clinician trust, visualizing model attention on ECG waveforms. Timeline: Research published in Nature Medicine 2020; pilot in cardiology clinics 2021; scaled 2023.[2]

Generative AI Search Tool with Google

In 2022, Mayo partnered with Google Cloud for the Mayo Clinic Platform, launching a generative AI clinical search in 2023. This tool leverages Med-PaLM 2-like LLMs fine-tuned on de-identified Mayo data, enabling queries like "patient history for HF risk factors." It federates multi-modal data (EHR, imaging, labs) securely via zero-trust architecture.[3]

Implementation involved Platform_Connect for data harmonization and Platform_Insights (2025 launch) for analytics. Pilot with 500 clinicians showed 40% faster insights; full rollout 2024. Challenges: HIPAA compliance overcome with differential privacy; hallucination risks mitigated by retrieval-augmented generation (RAG).[4]

Overcoming Challenges & Scalability

Data privacy was paramount; solutions used federated learning across Mayo sites. Clinician adoption boosted via training and human-AI collaboration demos. By 2025, integrated into 200+ AI projects, including cardiology risk prediction. Costs: Initial development ~$5M, ROI via early interventions reducing HF hospitalizations by est. 20%.[5] Current status: Production use, expanding to global partners via Platform_Solutions Studio.