Healthcare AI Daily Briefing

Research & Evidence

Multimodal foundation model achieves 0.895 diagnostic accuracy across six specialties.

Journal article (Nature Medicine)

The study reports EVLF-FM, trained on 1.3 million samples from 18 datasets spanning dermatology, hepatology, ophthalmology, pathology, pulmonology, and radiology, achieved the highest average diagnostic accuracy of 0.895 in internal validation, outperforming leading generalist and specialist models. In visual grounding across nine modalities, it reached an average Acc@0.5 of 0.842, with human-evaluated reasoning scoring 4.25/5 on the S.C.O.R.E. framework, though prospective clinical studies are required before deployment. [1]

Clinical Practice

SPARC transformer model predicts cancer adverse events across external cohorts.

Preprint, not peer reviewed

Researchers developed SPARC, a transformer-based foundation model trained on 101,920,379 measurements from 224,047 cancer patients, to predict cancer-related adverse events. The model achieved state-of-the-art performance on held-out development data, a prospective validation cohort of 23,126 patients from the same center, public external datasets of 26,012 patients, and multi-hospital cohorts of 562,620 and 206,840 patients; pooled multi-institution fine-tuning improved generalizability over single-institution fine-tuning and identified patients most likely to benefit from G-CSF. [2]

Research & Evidence

TRACER foundation model shows screening promise but detection burden varies across health systems.

Preprint, not peer reviewed

A decoder-based foundation model pretrained on 4.5 million patients across four health systems retained discrimination when tested on 16 incident diseases in over two million patients from two excluded health systems. At 95% specificity, flagged patients were up to eight times more likely to receive a diagnosis within a year than unselected patients. However, evaluating at this screening-relevant operating point revealed that detection and false-alert burden varied across diseases and health systems even when global discrimination metrics appeared stable, underscoring the need for external evaluation at clinically appropriate false-positive rates. [3]

Operations & Workflow

SHI Summit highlights cost discipline as key to healthcare AI value.

At the SHI Summit, Healthcare IT Today founder John Lynn reported that healthcare organizations must prioritize cost discipline and value economics to prevent AI costs from derailing future projects. Lynn noted that while healthcare struggles to generate new revenue compared to other industries, every AI implementation still requires a clear value equation linking spending to measurable outcomes. [4]

Industry & Products

AMA survey finds 81% of physicians now use AI professionally, 88% fear skill loss.

According to the 2026 AMA survey cited in the podcast launch, physician AI adoption more than doubled since 2023, while 88% reported at least some concern about losing clinical skills to AI and 70% worried about skill erosion among current medical students and residents. The discussion highlighted deliberate practice rotations and careful history-taking as safeguards against uncritical deference to algorithmic recommendations. [5]

Clinical Practice

Google announces AI certificate for healthcare professionals.

Google announces the Google AI for Healthcare Professionals Certificate, a self-paced program under five hours built by Google clinicians and reviewed by academic organizations, offering up to 5.0 AMA PRA Category 1 Credits. The curriculum covers evaluating AI outputs, patient communication, administrative task automation, and research synthesis, with NYU Grossman, UAB, Touro, and Purdue Nursing already offering it to faculty and students. [6]

Industry & Products

AI clinical trial screening nearly doubles enrollment in JAMA study.

Podcast from The Signal Room

The interview reports an AI screening tool with human-in-the-loop confirmation achieved an 84% enrollment increase (p=0.04) in a prospective JAMA-published trial, with AI-flagged patients showing reduced no-show rates and faster time to diagnostic procedures. Dr. Blood describes deploying this workflow at Mass General Brigham and through AIwithCare to other health systems, emphasizing clinician decision authority over automated selection. [7]