AI Is Learning to Read What Routine Scans Already Contain

The same trick, three organs

Three preprints published this week make the same move in three different organs: take data clinics already collect, and extract a signal that normally requires another test. A generative adversarial network reads left-ventricular mechanics out of a standard 12-lead ECG. A graph measure reads early Alzheimer's risk out of standard diffusion MRI. Two deep learning models read cancer molecular subtypes out of routine pathology slides. None of the three asks the clinic to buy a machine or order a new procedure, and all three are preprints, meaning their results have not been peer-reviewed. m:…1422 m:…9659 m:…9865

What each model found

The heart study derives synthetic left-ventricular mechanical parameters from 12-lead ECGs and finds they carry prognostic information beyond the usual tools. Synthetic e' velocity independently predicted all-cause mortality after acute myocardial infarction even after adjustment for the GRACE score, the standard risk score, and the synthetic measures added prognostic value beyond measured LVEF and AI-ECG-predicted EF. In a general-population cohort, MAPSE derived from ECGs was a significant predictor of myocardial-infarction-specific mortality, with a subdistribution hazard ratio of 0.49. The authors present this as a scalable, point-of-care risk tool that slots into existing ECG workflows. m:…1422

The brain study introduces the Route Concordance Index, a measure of white-matter routing from diffusion MRI, and reports that it detects conversion to mild cognitive impairment or Alzheimer's disease up to four years before clinical diagnosis, with an AUC of 0.831. It outperformed the established structural markers the authors compared it against, including hippocampal volume and entorhinal thickness, and the finding replicated in the ADNI cohort with an AUC of 0.797. m:…9659

The cancer study builds two models, HERA and HERMES, that classify molecular subtypes of vulvar squamous cell carcinoma directly from routine H&E-stained whole-slide images, bypassing the tissue-consuming p16 and p53 immunohistochemistry normally used. HERA predicted HPV/p16 status with an AUC of 0.84 in independent testing of 286 patients. HERMES-predicted subtypes reproduced the expected prognostic hierarchy, with five-year recurrence rates of 14.0 percent for HPV-associated disease, 31.1 percent for HPV-independent p53-wildtype, and 43.3 percent for HPV-independent p53-abnormal. But HERMES separated those last two groups only weakly, with AUCs of 0.68 and 0.63, and the authors flag that distinction as needing further work. m:…9865

Why routine data is the point

The shared implication is about where the bottleneck moves. When the input is a test every hospital already runs, the scarce resource is no longer data collection but validation: proving the model holds up across scanners, populations, and sites. That is a cheaper problem to have than building new imaging infrastructure, and it is also a harder one to hand-wave, because a model trained on one hospital's ECG machines or slide scanners has to survive contact with everyone else's. The three preprints all lean on the same promise, that the signal was sitting in the archive all along, and the same burden, that archive-derived signals need prospective proof before they change care. m:…1422 m:…9659 m:…9865

What is still unproven

All three results are preprint-stage, and the honest gaps differ by study. HERMES needs better separation of the HPV-independent subtypes and validation across scanners and populations. The heart and brain studies report strong retrospective discrimination but, as preprints, have not yet shown that their scores change treatment decisions or survive prospective trials. The pattern to watch is not whether more routine-data models appear, they will, but whether the next versions arrive with multi-site prospective validation attached, because that is the step that turns a hidden signal into a clinical one. m:…9865 m:…9659