AI in Healthcare: Research Roundup — October 7, 2026

Today's LLM additions lean toward patient-facing deployment and evaluation. A randomized controlled trial of PreA, an LLM chatbot handling history-taking and referral generation ahead of specialist visits, cut physician consultation time by 28.7% across 2,069 patients, with outcomes holding up even when the chatbot ran without staff support. In hospital documentation, a Gemini 2.5 Pro-based agentic workflow (MedAgentBrief) generated nightly discharge summaries across 384 admissions; physicians used its output in 57% of cases, rated 88% of unedited drafts as having no harm potential, and reported a significant drop in burnout scores. Elsewhere, CataractBot's expert-verified answers for cataract surgery patients were rated 84.5% accurate and complete, while scoping reviews flagged how early this space still is: only 8 studies exist on multiturn LLM chatbots for cancer patients and caregivers, and a separate review on LLM-based mental health chatbots found safety and privacy risks remain under-studied relative to the pace of deployment.

On the ML/DL side, prognostic nomograms kept targeting high-stakes, lower-profile conditions. A nomogram for acute chlorfenapyr poisoning hit a C-index of 0.988 in training (0.849 on validation) using procalcitonin, ALT, and dose as key predictors, while a readmission-risk nomogram for acute pancreatitis reached AUC 0.836 internally but dropped to 0.704 on external validation — a reminder that generalization remains the harder test. Imaging-based diagnostic models also advanced: a CBAM-ResNet model for Alzheimer's diagnosis reached AUC 0.876, a CNN classifying cervical CT scans distinguished Kikuchi disease, lymphoma, lymphadenitis, and tuberculosis with PR-AUC between 0.73 and 0.92, and an international reader study pitted a market-approved CNN against 96 dermatologists on real-world skin lesion classification.

Across both tracks, the same pattern holds: LLM tools are moving into live clinical workflows with measurable efficiency and burnout gains, while ML/DL diagnostic and prognostic models keep posting strong internal performance but a recurring external-validation gap.

Sources

1. RCT of LLM chatbot (PreA) for primary-to-specialist care transitions
2. Physician-reported safety outcomes of AI-generated hospital course summaries (MedAgentBrief)
3. CataractBot utility for pre/post-operative cataract care
4. Scoping review of LLM chatbots for cancer patients and caregivers
5. Scoping review on mental health harms of LLM-based chatbots
6. Prognostic nomogram for acute chlorfenapyr poisoning mortality
7. Nomogram for unplanned readmission in acute pancreatitis
8. CBAM-ResNet model for Alzheimer's disease diagnosis
9. CNN vs. 96 dermatologists on skin lesion diagnosis, international study

Prepared with AI assistance from newly published research. See the sources for the original studies.