Clinical AI grows up: automated heart imaging, expert-level pathogen calls, a foot-temperature warning

The echo lab automates

The deployment case for clinical AI got stronger this week. In a preprint, the authors report fully automated multi-chamber strain profiling in 8,350 consecutive patients undergoing transthoracic echocardiography — and the automation was feasible in 90.3% of them. Consecutive enrollment matters. This is my read as a writer, not reported fact: a consecutive cohort is the closest a retrospective study gets to the population a deployed tool would actually see. m:…9687

The clinical payoff is prognostic. Patients with pan-chamber dysfunction faced a 3.44-fold increased risk of all-cause death or heart failure hospitalization compared with those with preserved function, and the AI-derived strain measures added incremental prognostic value beyond clinical factors, left ventricular ejection fraction, and right ventricular fractional area change. This is my read as a writer, not reported fact: strain is the kind of measurement that rewards automation — high clinical value, tedious to do by hand. m:…9687

The scale is the point. Most AI imaging studies validate on curated cohorts; this one ran on 8,350 consecutive patients, and the automation held up in 90.3% of them. The authors also report that the AI-derived strain measures added incremental prognostic value beyond clinical factors, left ventricular ejection fraction, and right ventricular fractional area change — the standard measures a cardiologist would already have. m:…9687

The foot as a heart monitor

A second preprint explores an even simpler sensing channel for chronic heart failure: foot temperature. The authors report that mean right hallux temperature positively correlates with thermodilution cardiac output (r=0.63, p<0.001), and — more importantly for remote monitoring — ambient-adjusted hallux temperature declines in the 15 days preceding a heart failure event. m:…7248

This is my read as a writer, not reported fact: the pairing with the echo paper is the story. Cardiology is assembling a layered AI stack — automated imaging in the clinic, passive thermal sensing at home — and the papers that will matter are the ones that prove these signals hold up prospectively in the patients who need them most. m:…9687 m:…7248

The LLM that adjudicates infections

In infectious disease, a preprint reports a fine-tuned large language model for clinical pathogen adjudication in metagenomic sequencing: deciding which detected pathogen is the actual culprit. The model recovered 92.44% of reference culprit calls in a held-out test set and 84.49% in a later cohort. Notably, the authors found that metagenomic sequencing-derived metrics such as read counts or genome coverage alone are insufficient to identify culprit pathogens. My read as a writer, not reported fact: the adjudication problem is a reasoning problem, not a counting problem. m:…6979

The human benchmark is unusually rigorous: in a blinded re-adjudication of 215 cases, the model matched 90.5% of the culprit calls on which both expert reviewers agreed. This is my read as a writer, not reported fact: an LLM that agrees with a consensus panel nine times out of ten is a credible candidate for a first-pass adjudication role. m:…6979

The authors are explicit about what does not work: metagenomic sequencing-derived metrics such as read counts or genome coverage alone are insufficient to identify culprit pathogens. Adjudication is a reasoning problem, not a counting problem — which is why a language model, not a bigger counter, is the tool being tested. m:…6979