A University of Hong Kong team built a deep-learning score from UK Biobank blood data to estimate risk for six cardiovascular diseases, which ScienceDaily calls a new AI blood test although the evidence described is a predictive model rather than a clinical test available to patients. [1]

CardiOmicScore combines measurements of 2,920 circulating proteins and 168 metabolites, and the release says it estimates risk for coronary artery disease, stroke, heart failure, atrial fibrillation, peripheral artery disease and venous thromboembolism, with signals among people classified as high risk appearing as far as 15 years before clinical onset. [1]

Those are model outputs, and although the release says performance exceeded conventional polygenic risk scores and improved when age and gender were added, it does not provide the cohort size, event counts, discrimination or calibration values needed to judge that comparison, while ScienceDaily is carrying edited university material rather than replacing the primary paper's methods and tables. [1]

The implementation gap is larger still, as external and prospective validation must show whether performance survives different populations, laboratories and disease rates, an assay would need reproducible measurements, an acceptable price and a plan for false positives, and a clinical-utility trial would then have to show that acting on the score improves outcomes beyond ordinary risk assessment.

Prediction can help prevention only after those stages, and until then CardiOmicScore is a retrospective biobank result with a long forecast horizon rather than evidence that an early warning changes a patient's health.

-- DAVID CHEN, Beijing