AI finds the patients the medical record missed

The heart patients with no diagnosis

Viz.ai announced that two research abstracts evaluating its AI-powered Viz HCM algorithm have been accepted for presentation. The studies, conducted by Beth Israel Medical Center and the University of Colorado, examine the algorithm's ability to identify suspected hypertrophic cardiomyopathy in patients who might otherwise remain undiagnosed. Viz HCM was the first FDA-cleared AI algorithm designed to help clinicians detect signs of HCM using a standard 12-lead ECG, and an estimated 85% of patients with hypertrophic cardiomyopathy may remain undiagnosed, underdiagnosed, or misdiagnosed. m:…6844

The numbers show what the algorithm is actually catching. Nearly 40% of patients flagged by Viz HCM for suspected HCM had no corresponding diagnosis code in their medical records — the record simply did not know they were ill. Patients flagged by the algorithm had a 2.68-fold higher hazard of adverse clinical events, which means the flag is not noise: the people it picks out are the ones who go on to have problems. The University of Colorado study analysed ECGs from more than 29,000 patients, and Viz HCM has now been evaluated across 21 clinical abstracts and studies. m:…6844

Viz.ai's platform is deployed in 2,000 hospitals across the United States, and the company was the first awarded CMS reimbursement for AI. That distribution is the difference between a research result and a screening programme: an algorithm that runs on a standard ECG in 2,000 hospitals can find the undiagnosed at a scale no specialist clinic can match. The open question is what happens after the flag — a detection that nobody acts on is a more expensive way of not knowing. m:…6844

Blood pressure from a watch

Google Research has developed AI-PPG-ACC-HTN, a system that uses consumer smartwatch sensors to passively detect signs of hypertension without requiring cuff calibration. The system analyses pulse wave contours and mechanical heart vibrations, aggregates passive physiological signatures over multiple days to identify persistent hypertensive states, and was validated against 24-hour ambulatory blood pressure monitoring. It achieves accuracy levels comparable to traditional cuff-based diagnostic workflows. Hypertension affects more than one billion people globally, and more than half of adults with hypertension are unaware they have it. m:…2634

The performance figures are honest about the limits. Initial screening with the smartwatch achieved a sensitivity of 65.8% and a specificity of 90.0%, and confirmatory testing reached a sensitivity of 78.4% and a specificity of 90.9%. The model was trained on 492,347 hours of sensor data from 2,109 participants, and the clinical study involved 196 participants across five US sites. A test that catches two-thirds of cases on first screening and nearly four-fifths on confirmation is a triage tool, not a diagnosis — and the research is a collaboration between Google Research, Beth Israel Deaconess Medical Center, and Harvard Medical School. m:…2634

The system is currently exploratory and not a commercial product, which is the right framing. A smartwatch that flags possible hypertension in someone who would never have bought a cuff is valuable precisely because it reaches people outside the clinic — but only if the flag leads to a cuff measurement and treatment, not to a notification that gets dismissed. The technology is ready for the question; the care pathway is not yet built. m:…2634

A million base pairs of context

Deep Genomics has developed DeepRNAi, an AI model for predicting siRNA activity and off-target effects across the transcriptome. It integrates local hybridization data with global genomic context provided by biological foundation models — BigRNA, Enigma, and REPRESS — which take into account up to a million base pairs of sequence context. Incorporating the foundation models nearly doubles the performance of the DeepRNAi model, and its performance rivals experimental screening methods. The company says it will soon release a preprint on the work. m:…8824

The case studies show why context matters. DeepRNAi correctly predicted that Alnylam's siRNAs targeting the distal 3' UTR of the PLG transcript would be inactive — a call that depends on the surrounding sequence, not just the target site. A common variant in the IRF5 gene, with a 58% global allele frequency, creates a proximal polyadenylation signal, and DeepRNAi with foundation models identifies that siRNAs tiling the IRF5 variant region are unlikely to work for a large portion of the population. A model that only looks at the target would miss both findings; a model that reads a million base pairs around it catches them. m:…8824

The three stories share a shape. In each case the AI is not replacing a clinician or a bench scientist; it is finding the cases the existing system misses — the patient with no diagnosis code, the person with undetected hypertension, the drug candidate that looks good at the target site and fails in context. The value is in the miss rate, and the risk is the same in all three: a flag that nobody follows up is worse than no flag at all, because it creates the illusion that someone is watching. m:…8824