We Got a Cardiac MRI Analysis AI SaMD Cleared in 5.5 Months, PCCP Included
Heartvue.ai - Dr. Jeffrey Dendy
FeaturedAI/ML Submission in 3 MonthsFDA Strategy & Pre-Sub in 4 WeeksAI/ML510kCybersecurityWeb-AppDICOM
Testimonial
This was our first submission at Heartvue.AI, and it was not a simple one. Between the AI/ML performance questions, the PCCP, and several rounds of deficiency responses on tight timelines, there were plenty of moments where we needed more than process management. We needed people who deeply understood both the science and how FDA thinks. Mary and Matt were that, every single time. They caught issues before FDA did, they explained their reasoning clearly, and they were calm and steady. We will be recommending Innolitics to anyone who asks (and probably to some who don't).
Jeffrey M. Dendy, M.D.
CEO of [Heartvue.ai](http://heartvue.ai/), Assistant Professor of Cardiovascular Medicine, Vanderbilt University Medical Center
Summary
Heartvue.ai is a cardiologist-founded startup that built Heartvue.Proton, an AI/ML-powered platform that automates cardiac MRI analysis: 17 linear measurements, ventricular volumes and ejection fraction, and blood flow quantification, all under physician review. They came to Innolitics in May 2025 with roughly 20 trained ML models and zero regulatory experience. Fifteen months later, Heartvue.Proton was 510(k) cleared (K260811) on their first submission, with an FDA-authorized PCCP that lets them update their models post-market without new submissions. PCCP, or Predetermined Change Control Plan, is a regulatory tool that allows companies to "pre-clear" certain changes without having to go back to the agency with another 510(k).
The Problem
Cardiac MRI is the gold standard for assessing heart structure and function, but manual analysis takes 45 to 90 minutes per study. Radiologists must trace cardiac chambers across dozens of slices, compute volumes and flows, and compare results against normative data, all under high case volumes. The result: quantitative measurements are often skipped in favor of "eyeball" assessments, and the modality's superior accuracy goes underused.
Dr. Jeffrey Dendy, a practicing cardiologist at Vanderbilt, built Heartvue.Proton to fix the problem he lived every day. The models worked. The path to market did not exist yet.
The Challenge
Heartvue had never submitted anything to FDA. And this was not an easy first device:
A multi-model AI/ML system. Roughly 20 ML models across three algorithm families (2D landmark measurements, volumetric segmentation, phase-contrast flow), each with different FDA performance evaluation expectations. An unfocused approach would have ballooned into 21 separate validation protocols.
Evolving AI/ML expectations. Training data documentation, generalizability across scanners and demographics, standalone performance testing, and a PCCP for post-market model updates.
Cloud architecture under cybersecurity scrutiny. A PACS-connected, cloud-hosted system with DICOM and HL7 interfaces, reviewed under FDA's 524B cybersecurity requirements.
Startup constraints. Limited capital, limited bandwidth, and a timeline where every month mattered.
What We Did
We ran the full regulatory lifecycle, from first FDA contact to clearance letter:
De-risked the strategy with a Pre-Submission first. We put every consequential question in front of FDA before Heartvue spent money on validation studies: predicate selection, ground truth methodology, per-model metrics and sample sizes, the PCCP concept, and the non-device CDS boundary for normative data references. FDA praised the package and validated the strategy.
Right-sized the validation program. We consolidated the evaluation of ~20 models into a master performance plan organized by model category, with clinically meaningful endpoints for each family (point-to-point distance for landmark models, where Dice would have been the wrong tool).
Authored the 510(k) end to end. Device description, substantial equivalence, performance reports, PCCP, and cybersecurity documentation, compiled into eSTAR and submitted March 2026.
Used every FDA touchpoint. We requested an early orientation meeting so the review team saw a live demo before opening the documentation, which set up a communicative, collaborative review.
Turned deficiency responses around in days. When FDA raised a demanding cybersecurity question late in interactive review with a same-week deadline, our team delivered a complete, internally consistent response strategy within days. We caught inconsistencies before FDA did.
The Result
FDA 510(k) cleared August 26, 2026 (K260811), 5.5 months after submission, on Heartvue's first submission ever.
The clearance includes an FDA-authorized PCCP, so Heartvue can retrain and improve its ML models within pre-defined bounds without filing a new 510(k) each time. From engagement kickoff to clearance letter: 15 months.
Heartvue is now moving to commercial rollout, and Innolitics continues to support them on registration and listing, UDI, and post-market readiness.
Heartvue.ai is a medical imaging AI company based in Brentwood, Tennessee, founded by Dr. Jeffrey Dendy, a practicing cardiologist and Assistant Professor of Cardiovascular Medicine at Vanderbilt University Medical Center.
Heartvue.Proton ingests cardiac MR studies directly from hospital PACS, runs its ML models automatically, and presents semi-automated results for physician review in a web viewer, with finalized reports delivered to the EHR over HL7.
The platform automates 17 linear 2D measurements, ventricular volumes, ejection fraction and mass, and aortic and pulmonary flow quantification, with every measurement reviewable and editable by the physician.
