Healthcare AI Daily Briefing

Research & Evidence

NLP extraction found social risk factors across all 17 domains in 4,844 CVD patients.

Preprint

Using a custom NLP algorithm to extract 17 social and behavioral domains from EHR notes identified patients in every domain and higher prevalence for each, while ICD-10 codes and questionnaires captured only a subset; among 4,844 CVD patients, 526 (10.9%) were frequent ED users. [1]

Diagnostics & Pathology

Deep learning predicts breast cancer subtypes from H&E slides with 0.81 accuracy.

Journal article

A modified EfficientNetB0 model trained on 1,000 H&E‑stained images achieved overall classification accuracy of 0.81 on internal test set. Luminal A and Basal‑like subtypes performed best (F1 0.93 and 0.86). External validation on 50 patients from another hospital yielded accuracy 0.78. [2]

Clinical Practice

Many health systems cite document delays hurting patient care.

According to a Documo report, 48% of healthcare organizations say document delays often or always negatively affect patient care. Additionally, 60% of payers and 64% of providers report employees using shadow AI tools, highlighting workflow and security challenges. [3]

Operations & Workflow

Speaker notes self‑scheduling needs specialist routing to avoid missteps.

Podcast from Healthcare AI Pioneers

A speaker on Healthcare AI Pioneers explained that while self‑scheduling tools empower patients, they often fail to direct them to the correct specialty or location, leading to mis‑scheduling; precise agentic navigation aims to connect patients to the right care. [4]

Diagnostics & Pathology

Speaker urges trials showing AI tools beat death as comparator.

Podcast from AI and Healthcare

A speaker argued AI tools should be measured against the real standard of care, noting that in many places the comparator is death, and called for trials to prove benefit. [5]