A vision-language model trained on more than 400,000 CT volumes paired with radiology reports extracted physiological signals far beyond each scan's original clinical question, showing associations with 35 laboratory measurements and 351 diagnosis categories. It outperformed image-only models and organ-measurement baselines at predicting later diagnoses, with the strongest performance for heart failure, cirrhosis, and acute kidney failure. Direct comparison against human radiologists has not yet been performed.
Key Points
- Model trained on 400,000 CT volumes paired with radiology reports
- Embeddings linked to 351 diagnoses and 35 laboratory markers
- Best prediction: heart failure, cirrhosis, acute kidney failure
Longevity Analysis
Every routine scan already contains far more physiological information than any report captures, and much of it concerns the slow accumulation of cardiac, hepatic, and renal burden that defines late-life decline. Reading those latent patterns converts imaging performed for an unrelated complaint into an early signal of where circulation, filtration, and detoxification capacity are drifting — years before symptoms force the question. Earlier recognition only matters when it is acted on consistently, but it widens the window in which intervention still changes the trajectory.
Original published by LifeSpan.io, by Arkadi Mazin.