Approach

From Signal to Insight

Every echocardiogram, every CT, every ECG begins as raw digital data — a mixture of signal, noise, and variation. Our work is to reconstruct these into faithful representations from which two kinds of value emerge: automation of what the skilled clinician already does, and augmentation through signal that no clinician could see at all.

Automation + Augmentation

Every echocardiogram, every CT, every ECG arrives first as digital data — a dense mixture of signal, noise, and redundancy. Our work is to take that raw input and reconstruct it into faithful representations that discard what is meaningless and preserve what matters. From those representations, two kinds of value emerge along a single axis.

Automation is the reliable, reproducible, and precise reading of what the skilled clinician already sees — LVEF, wall motion, chamber size, rhythm, valve anatomy. Augmentation is the extraction of hidden labels — phenotypes, pre-clinical disease, and risk signatures that no expert reader can reliably identify unaided. One gives us the floor of human performance, replicated everywhere. The other gives us the ceiling we could never reach alone.

① INPUT ② DIGITAL DATA ③ REPRESENTATION ④ VALUE ECG ECHO CT Hidden labels Human-visible labels Clinically insignificant Low-yield data Noise · artifact High- yield Low- yield Layered signal & noise Representation learning discard noise · preserve signal AUTOMATION Reproducible · Precise · What the clinician sees PanEcho → LVEF · valve disease · wall motion AI-ECG → rate · rhythm · intervals · axis Scales expert performance · removes variance AUGMENTATION Inference · Discovery · What the eye cannot see CTRCD risk from ECG Aortic stenosis phenotype from TTE Perivascular fat attenuation (FAI) on coronary CT Epicardial fat radiomic profile on CCTA

From raw input to resolved insight: clinical imaging is decomposed into layered digital data, compressed through representation learning into faithful signal, and returned to the clinician along two value axes — automation of the human-visible, and augmentation through the hidden.

Four Principles

Four design principles that recur across a decade of building, validating, and deploying cardiovascular AI biomarkers.

01

Augment, don't replace.

AI earns its place by giving clinicians time and reach — not by adding cognitive load.

Eur Heart J 2024 [1] → 02

Hidden labels matter most.

High-value models don't re-read what an expert can read; they surface what no expert can see.

JACC 2024 [2] → 03

Target, don't blanket.

Deployment is the biomarker. An EHR-guided gate turns a high-AUC model into a high-yield tool.

NEJM AI 2025 [3] → 04

Value > performance.

Impact, equity, and workflow fit — not AUC alone — separate good papers from good tools.

JACC Imaging 2026 · PRIME 2.0 [4] →

Targeting the Augmentation Biomarker

A high-performance augmentation biomarker is only half the story. Who receives the inference, and when, is what determines whether it translates into benefit. Blanket deployment multiplies false positives, alert fatigue, and inequitable yield. Targeted deployment — gating the AI on the patient's longitudinal EHR phenotype at that moment — turns the same algorithm into a precise, high-value clinical tool.

UNTARGETED TARGETED ECG unselected population AI model Alert fatigue · low precision · inequitable yield ECG same population EHR GATE AI model Higher precision · fewer false positives · equitable benefit true positive false positive

We can build powerful tools — but who benefits depends on where we deploy them. Moving from untargeted to EHR-gated, phenotype-aware deployment is what converts a high-performance augmentation biomarker into a high-value clinical one. After TARGET-AI, NEJM AI 2025 [3].

The Clinician as Bayesian

The pace of AI innovation now exceeds any individual's ability to keep up. The answer is not to retreat — it is to reimagine our role. The clinician of the next decade will operate as a Bayesian integrator: bringing the prior (their training, their knowledge of the patient, the evidence base) into conversation with the AI output (noisy, powerful, but not infallible), updating beliefs, and using each as a reference mark to check the other.

This is the practice we are building toward — one in which AI makes medicine more human, not less.