OpenRounds Editorial

Healthcare AI Daily Briefing

Monday, October 26, 2026

New to OpenRounds? Get the free daily email or see how we work.

Operations & Workflow

Mobile AI triage shows 75% sensitivity but low PPV for high-risk oral lesions in community screening.

Journal article (JMIR mHealth and uHealth)

In a prospective field study in eastern Taiwan, trained nonspecialists used a parameter‑locked mobile AI system to triage intra‑oral images from 602 participants. The AI flagged 68 red lesions; 12 (17.6%) were confirmed high‑risk and four more were found on expert review (16 total). Sensitivity for high‑risk lesion identification was 75% (12/16; 95% CI 50.5%‑89.8%) and PPV 17.6% (12/68; 95% CI 10.4%‑28.4%). Authors stress the system supports human‑supervised referral prioritisation, not autonomous diagnosis. [3]

Medical Imaging

Bilateral lung radiomics plus clinical data yields 0.803 hold‑out AUC for smoking status but remains investigational.

Journal article (Journal of imaging informatics in medicine)

In a retrospective study of 506 thoracic CT scans (253 smokers, 253 nonsmokers), concatenated right‑left lung radiomic features combined with age, sex, height, weight, and BMI, using ReliefF feature selection and a stacking classifier, reached a hold‑out ROC‑AUC of 0.803 (balanced accuracy 0.70, sensitivity 0.72, specificity 0.68). Clinical‑only random forest yielded a hold‑out ROC‑AUC of 0.610. The authors caution that performance is moderate, varies by sex subgroup, and lacks external validation, so the approach should be considered investigational. [2]

Medical Imaging

Proof‑of‑concept study finds ultra‑low‑field R1 mapping by field‑cycling imaging discriminates moderate‑to‑severe SVD.

Journal article (Magma (New York, N.Y.))

In a proof‑of‑concept study of 39 participants, field‑cycling imaging acquired a single image slice at 0.2, 2, 20 and 200 mT. Co‑registered 3 T FLAIR MRI identified 16 participants with moderate or severe white‑matter hyperintensities and 21 controls. The best histogram feature discriminated groups with sensitivity 0.94 and specificity 0.90; a deep‑learning classifier reached sensitivity 0.67 and specificity 1.00; two radiologists achieved sensitivities 0.81 and 0.88 with specificities 0.86 and 0.52, respectively. [1]

Sources

Ultra-low-field R(1) mapping at 0.2 mT obtained by field-cycling imaging correlates with cerebral small vessel disease burden. · Magma (New York, N.Y.) Original source
Radiomic Signatures of Smoking in Visually Normal Lung Parenchyma: Bilateral CT Analysis Using Machine Learning. · Journal of imaging informatics in medicine Original source
Mobile AI-Assisted Oral Lesion Triage for Community Oral Cancer Screening: Prospective Field Evaluation Study. · JMIR mHealth and uHealth Original source

Mobile AI triage shows 75% sensitivity but low PPV for high-risk oral lesions in community screening.

Bilateral lung radiomics plus clinical data yields 0.803 hold‑out AUC for smoking status but remains investigational.

Proof‑of‑concept study finds ultra‑low‑field R1 mapping by field‑cycling imaging discriminates moderate‑to‑severe SVD.

Sources

Get the free healthcare AI briefing