AI Finds Hidden Health Clues in Routine CT Scans
An AI model extracted disease and future risk signals from routine CT scans, but clinical benefits still need testing.
Summary
Routine CT scans may reveal health risks beyond the reason they were ordered. University of Pennsylvania researchers developed Percival, an artificial intelligence model trained on more than 400,000 CT image volumes and accompanying reports from 52,000 participants. Its image patterns were associated with 35 laboratory measurements and 351 diagnosis categories. Models built from these patterns performed particularly well at identifying recorded diagnoses of several heart conditions and pancreatic cancer. Percival also outperformed comparison models in predicting later diagnoses, with its strongest results involving heart failure, cirrhosis, and acute kidney failure. However, most laboratory values were difficult to estimate directly. The findings suggest a possible way to extract more value from existing scans, not a reason to seek extra imaging. Independent validation and comparisons with radiologists remain important next steps.
Detailed Summary
Routine CT scans may contain clues about health far beyond the condition doctors originally wanted to investigate. Researchers at the University of Pennsylvania developed an artificial intelligence model called Percival to extract those clues, potentially helping clinicians recognize overlooked disease and estimate future risks without collecting additional images or exposing patients to extra radiation.
The model learned from more than 400,000 CT image volumes paired with radiology reports from 52,000 participants in the Penn Medicine BioBank. Its image patterns reflected age, sex, and body size. Researchers also found statistically significant associations with 35 laboratory measurements, including indicators of kidney and liver function, metabolism, blood counts, and inflammation. However, estimating most individual laboratory values directly remained difficult.
The model showed associations with 351 diagnosis categories across multiple organ systems. When researchers trained it to distinguish patients with and without recorded diagnoses, its strongest results included atrial fibrillation, coronary atherosclerosis, heart failure, and pancreatic cancer. It outperformed an image model trained without accompanying reports, suggesting that linking images with clinical descriptions added useful information.
For later diagnoses, Percival outperformed comparison models using images alone, organ measurements, or age and sex. Results were strongest for heart failure, cirrhosis, and acute kidney failure. If validated, this approach could help doctors use existing scans to flag patients who need further assessment. It does not establish that additional CT screening would benefit healthy adults.
These findings describe associations and prediction, not proof that the model detects disease earlier or improves outcomes. The article provides no detailed accuracy estimates, and direct comparison with radiologists remains a next step. Independent validation, reliable performance across different populations, and evidence that alerts lead to better care are needed before routine use. Patients should continue established screening and discuss scan findings with their clinician rather than seek imaging solely for prediction.
Key Findings
- Percival learned from more than 400,000 CT image volumes paired with radiology reports from 52,000 participants.
- Image patterns were associated with 35 laboratory measurements and 351 diagnosis categories, but most laboratory values remained difficult to estimate.
- Disease classification results were strongest for atrial fibrillation, coronary atherosclerosis, heart failure, and pancreatic cancer.
- Future diagnosis prediction outperformed comparison models, with strongest results for heart failure, cirrhosis, and acute kidney failure.
- The findings support further testing of existing scans, not additional CT screening for healthy adults.
Methodology
This Lifespan.io research news report summarizes a University of Pennsylvania study published in npj Digital Medicine. The evidence comes from training an image-and-text AI model on CT scans and radiology reports, then evaluating associations with laboratory results, recorded diagnoses, and later diagnoses. It is observational predictive research, not a trial demonstrating improved patient outcomes.
Study Limitations
The supplied article ends abruptly and does not provide detailed accuracy estimates, prediction time horizons, or sufficient information about independent validation. The primary paper should be checked for data separation, potential information leakage, population diversity, and whether later recorded diagnoses represent genuinely new disease. Direct comparison with radiologists and evidence of improved clinical outcomes remain outstanding.
Enjoyed this summary?
Get the latest longevity research delivered to your inbox every week.
Enter your email to subscribe: