Scientific and Strategic Leadership in Clinical AI, Mayo Clinic

Zachi I. Attia, PhD MBA

For more than a decade my work has focused on a simple but fundamental question: Can artificial intelligence discover clinically meaningful signals that humans cannot see-and can these discoveries reliably change patient care?

Our team has taken this idea from early AI-ECG models, through large prospective studies, FDA clearance, CMS reimbursement, and now into multimodal foundation models for the enterprise. We build systems that move beyond model accuracy and toward clinical trust, safety, and impact.

Explore Our Work Join Our Team

Lead, Sovereign AI, Mayo Clinic Enterprise Generative AI Initiative

Executive MBA, MIT Sloan (2025)

PhD, Computational Biology, University of Minnesota

MSc,BSc Electrical and Computer Engineering, Ben-Gurion University

20+ years applied R&D, 11+ years AI-healthcare

Inventor on 20 patents, 3 FDA-cleared AI diagnostics

Vision

Clinical AI is shifting from narrow predictors to system-level tools that will support real clinical decisions. My work focuses on three principles:

Evidence First

AI in medicine must be validated like a diagnostic, not like a tech demo. Our group designs prospective studies, randomized trials, real-world evaluations, and rigorous comparisons to standard care.

Clinical Use > Technical Novelty

We treat models as candidates for deployment. If it cannot be implemented, validated, or used by clinicians, we don't count it as progress.

AI Systems Must Be Safe

As part of Mayo Clinic's Generative AI initiative, we build enterprise guardrails such as CURE, a system for hallucination detection and reduction in LLMs. Safety and reliability are first-class requirements, not afterthoughts.

Research Journey

Our work represents a systematic progression from proving AI can enhance cardiac diagnostics to deploying these tools in real clinical settings and building next-generation foundation models.

Phase 1

Seeing the Invisible (AI-ECG)

We demonstrated that deep learning can detect reduced ejection fraction, atrial fibrillation, cardiomyopathy, and other conditions using standard ECGs-even when findings are invisible to expert readers.

Impact: Established the concept that routine clinical signals contain far more information than humans can perceive.

Phase 2

Beyond ECG (Echocardiography & Imaging)

We adapted these ideas to echocardiograms, developing models that infer ejection fraction and structural features-even from a single frame.

Impact: Showed that the principle generalizes across modalities and that minimal imaging can still contain actionable information.

Phase 3

Trials, FDA Clearance, CMS Reimbursement

We led or co-led the first randomized clinical trials of AI-ECG screening, obtained multiple FDA 510(k) clearances, and achieved CMS reimbursement. These AI tools have since been used in more than 800,000 patient encounters.

Impact: AI moved from algorithm to regulated clinical tool-closing the translation gap most AI research never crosses.

Phase 4

Multimodal Early Fusion Foundation Models

Today our team is developing multimodal foundation systems integrating ECG, echocardiography, imaging, waveforms, and clinical text in early fusion architectures. These models will serve as core backbones for diverse clinical tasks across Mayo Clinic.

Impact: Toward AI systems that understand complex clinical states, not isolated predictions.

Scientific Approach

We build AI systems that expand what clinicians are capable of, revealing patterns no human can see and enabling decisions that were previously impossible. The future of care belongs to teams where human expertise is amplified by intelligent systems, unlocking a level of precision and foresight that fundamentally changes how medicine is practiced.

Rigorous Validation

Prospective studies
Pragmatic multicenter trials
External validation across populations
Comparison to gold-standard measurements

Inside the Model’s Reasoning

Testing where model performance comes from
Falsifiable explainability methods
Perturbation studies and simulated experiments
Physiologically plausible features

Human-AI Complementarity

How clinicians interact with AI recommendations
Designing outputs that improve decision-making
Trust and transparency in clinical context

Deployment & Integration

Collaboration with cardiology, IT, operations
Regulatory science and FDA pathways
Workflow integration with measurable impact

Explainability in AI Systems

We go beyond traditional heat maps to understand what models are actually using to make predictions. This includes perturbation studies, simulated signal experiments, counterfactual testing, and evaluating whether the model relies on physiologic features instead of artifacts or bias.

We apply the same approach to large language models, studying how they reason through multi-step clinical tasks, how errors arise, and how to prevent unsafe behaviors before they reach patient care.

Goal: Build AI systems that behave safely, consistently, and transparently so clinicians can trust the outputs.

Join Our Team

Our group includes researchers, clinicians, engineers, analysts, and regulatory experts from more than eight countries, working together on AI systems that matter.

We Welcome

Ideal Experience

Current Focus Areas

We value people who can take an idea from concept to working prototype to clinical study.

Connect on LinkedIn