Healthcare AI Daily Briefing

Diagnostics & Pathology

2.5D deep learning stacking outperforms other MRI models for glioma molecular classification.

Journal article (Journal of magnetic resonance imaging : JMRI)

In a retrospective multicenter study of one thousand four hundred eight adults with diffuse glioma, a 2.5D deep-learning stacking model achieved the highest external AUCs for isocitrate dehydrogenase (IDH) mutation (0.89–0.93) and 1p/19q codeletion (0.92–0.95), significantly outperforming VOI-mean, radiomics, 2D, and 3D deep-learning approaches. The derived 2.5D-IDHwt score combined with clinical factors yielded higher C-indices than a pathology-based model in all three external test cohorts, but the improvement reached statistical significance only in cohorts 2 and 3. [1]

Research & Evidence

Machine learning predicts persistent fever in cancer patients on antibiotics.

Journal article (npj Digital Medicine)

Models using routine EHR data from two academic hospitals achieved AUROC 0.82 internally and 0.78 externally for predicting fever persistence at 48–72 hours after antibiotic initiation, with temperature patterns as dominant predictors. The authors caution that clinical use should remain adjunctive and prospective evaluation is needed before standalone decision-making. [2]

Industry & Products

East Georgia center integrates pharmacy with EHR workflows.

At East Georgia Healthcare Center, the new eClinicalWorks pharmacy module connects pharmacy operations with the existing EHR, synchronizing patient contact details and eliminating duplicate updates, which staff report makes a meaningful difference for 340B program and open-door pharmacy management. [3]

Industry & Products

Clinical AI governance needs version control, not just one-time validation, argues independent researcher.

In an opinion piece, Dr. Michael T. Neely argues that healthcare AI governance must track which version was validated, which is running, what changed, and what evidence supports continued safety — because vendors can replace models, modify prompts, or retrain without changing the product name. He calls for internal version records capturing vendor, deployed model, configuration, approved use, implementation date, validation evidence, and responsible owner; for generative AI, records should also include prompt templates, retrieval sources, and orchestration logic. Material-change notification should be an operational requirement in vendor contracts, with defined processes for suspension, rollback, and incident traceability. [4]

Industry & Products

Google Antigravity adds cheminformatics skills to accelerate drug discovery workflows.

Google DeepMind's blog reports that Antigravity's new Cheminformatics Skills let an agent search thousands of compounds, perform matched molecular pair analysis, and propose rigidified macrocycles with synthetic routes in about an hour instead of multiple days. The company notes that sparse data, inter-lab assay noise, and narrow dynamic ranges limit predictive precision, and computational designs remain testable hypotheses requiring experimental validation. [5]

Clinical Practice

Interview: Google AI models detected cardiovascular risk from retinal images.

Podcast from The AI Cure: How Technology Is Changing Cancer Care

The interview describes early Google models for diabetic retinopathy that identified cardiovascular risk, patient age, and sex from retinal images alone, which clinicians could not discern. The guest recalls this demonstration as a pivotal moment showing AI could produce new scientific insights beyond human perception. [6]

Diagnostics & Pathology

Retinal photograph biomarker reconstructs individualized lifespan trajectories across three continents.

Journal article (Nature biomedical engineering)

Trained on a European cohort and externally validated across independent cohorts in Southeast Asia, Eurasia, and North America, the Retinal Mortality Equivalent (RetiME) deep-learning biomarker showed robust associations with all-cause and cause-specific mortality and predicted incidence of diverse systemic and ocular diseases. Beyond static risk, RetiME re-timed mortality risk by reconstructing personalized lifespan trajectories associated with observed age of death; clustering revealed distinct mortality patterns spanning cancer-predominant, cardiometabolic, and neurodegenerative risks. Genome-wide analyses identified genetic loci linked to RetiME. [7]

Research & Evidence

AI model MANTIS predicts survival, guides reduced radiotherapy for childhood brain tumors.

Journal article (Nature Genetics)

Stefan et al. developed MANTIS, an AI tool using DNA methylation arrays from 2,786 patients to predict individualized 10-year survival for medulloblastoma and ependymoma, achieving concordance indices up to 0.80 ± 0.07 versus current indicators. Extending the model identified ~60% of average-risk medulloblastoma patients with no survival disadvantage (P=0.683) from reduced craniospinal irradiation, linked to better neurocognitive outcomes. [8]