A model trained on 34,000 paired chest CTs and echocardiograms finds reduced ejection fraction in scans ordered for something else entirely. Explore the data.
Source: European Heart Journal – Digital Health (2026) ·
Dr. Ashley Beecy (Chief AI Officer, Sutter Health)
Read today's newsletter →
A NewYork-Presbyterian/Weill Cornell team trained and validated a model on ~34,000 paired chest CTs and echocardiograms. The model detects abnormal ejection fraction from routine, non-gated, non-contrast chest CTs — images taken for lung, cancer, trauma, or anything else.
Each dot represents a simulated patient from the study cohort. The x-axis is the true ejection fraction (from the echo). The y-axis is the model's predicted EF. Dots below the 40% line have reduced EF — heart failure that wasn't the reason for the scan.
The diagonal line is perfect prediction. Dots near it means the model's EF estimate matches the echo. The red dots below 40% are the incidental findings — heart failure the scan wasn't looking for. The green dots are cases the model would correctly flag for follow-up.
This is a detection model, not a diagnostic. An incidental finding of reduced EF still needs a confirmative echo. The model says "look here" — it doesn't say "treat this." The value is in the routing, not the verdict.
34,000 paired scans is a specific dataset. The model was trained at one academic center on one patient population. Generalization to community hospitals with different CT protocols, different patient demographics, and different disease prevalence is not guaranteed. Prevalence matters: in a low-risk screening population, even a good model produces mostly false positives.
The business isn't the model — it's the "now what?" Who gets alerted when the incidental finding lands? Where does it go in the chart? What's the follow-up order? The model is the easy part. The incidental-finding routing layer is the standalone product.
Don't build another imaging model. The opportunity isn't a new classifier — it's the plumbing. The incidental-finding routing layer (who gets pinged, how it's tracked, who closes the loop) becomes a standalone product category before the imaging models themselves are commoditized.
The pathway, not the predictor: