What the research found

Researchers at the University of Pennsylvania developed an AI model called Percival that extracts hidden health information from CT scans beyond what radiologists typically report. The model was trained on over 400,000 CT scans paired with their radiology reports, learning to recognize patterns that correlate with patients' physiological state and disease risk.

When researchers analyzed what Percival's learned patterns represented, they found associations with dozens of blood biomarkers—including markers of kidney and liver function, glucose and lipid metabolism, blood cell counts, and inflammatory markers. The model also identified links between scan patterns and hundreds of diagnosed conditions across multiple organ systems, and crucially, it could predict which conditions patients were likely to develop later. Performance was strongest for cardiovascular outcomes (atrial fibrillation, coronary atherosclerosis, heart failure) and some cancers, as well as metabolic complications like acute kidney failure and cirrhosis.

However, the model's ability to directly estimate specific lab values from scans was weak for most measurements—only albumin, hemoglobin, and creatinine showed meaningful predictive power. This suggests the scans contain broad physiological "signal," but that signal doesn't translate neatly into precise numeric predictions.

Why it matters for you

If you're using MyKine to optimize health markers and catch problems early, this work points to untapped information already in your medical record. Many users have had CT scans for various reasons—screening, injury workup, or investigation of symptoms. Percival's findings suggest those images contain clues about your metabolic health, kidney and liver function, and cardiovascular risk that a standard radiology report might miss. Over time, if such models become clinically deployed, a routine CT could flag emerging disease risk before it shows up in your lab work.

That said, the predictive edge here is modest. The model outperformed simpler baselines (age, sex, basic organ measurements) but wasn't compared directly against a skilled radiologist reading the same scan. For now, this is best viewed as a research proof-of-concept that imaging data encodes more physiology than we routinely extract—not as a replacement for your tracked biomarkers or clinical interpretation. The strongest results came for cardiovascular conditions and liver disease, which may be of particular interest if those are areas of personal health focus.

One practical nuance: the study looked at diagnoses recorded within or shortly after imaging, so predictions partly reflect undiagnosed disease already present, not forecasting years into the future. Still, earlier recognition of existing disease—before it becomes symptomatic—is clinically valuable.

Caveats

• No direct comparison with radiologists. The study doesn't show Percival is better than a human radiologist at diagnosing the condition the CT was ordered to investigate—only that it detects broader disease associations.
• Temporal ambiguity. A "predicted" diagnosis could be disease present but unrecognized at scan time, not true long-term forecasting.
• Weak direct biomarker prediction. Most lab values could not be estimated accurately from scans; only three showed practical predictive power.
• Retroactive association study. Findings are based on observed diagnoses and records, not prospective validation in new patients.
• Early-stage clinical translation. No indication yet that this model is being used in clinical practice or validated against independent patient cohorts.