AI in Healthcare: Research Roundup — October 2, 2026

On the LLM side, today's additions skew toward diagnostic and decision-support benchmarking rather than new clinical deployments. A conversational diagnostic AI (AMIE) completed text-chat history-taking with 100 primary-care patients and included the correct diagnosis in its differential 90% of the time (75% top-3 accuracy), with no safety-supervisor interventions needed. Comparative studies were more mixed: ChatGPT o1-preview reached 68.6% concordance with pathologist consensus on oral histopathology descriptions (versus 27.5% for Gemini), and ChatGPT-3.5/4.0 outperformed Bard and Bing on neuro-ophthalmology diagnostic questions. A few-shot triage-categorization benchmark found Claude Haiku 4.5 edging out a supervised BioBERT baseline (macro-F1 0.475 vs. 0.378) but still struggling on ambiguous urgent-review cases, and a separate benchmark found generalist models (GPT-5, Gemini 3 Pro, Claude Sonnet 4.5) outperforming two deployed clinical AI tools on combined medical-knowledge and clinician-alignment tasks. A Nature review mapped security and safety hazards across the clinical LLM development pipeline, and a JMIR perspective argued human-in-the-loop review remains an essential accountability guardrail.

On the ML/DL side, today's models spanned imaging, oncology, and endocrinology. A deep learning model trained on over 62,000 UK Biobank retinal photographs predicted Alzheimer's-linked risk factors (AUROC up to 0.95), and a nationwide transformer-based model flagged young-onset type 2 diabetes with the highest-risk decile showing a 118-fold relative risk 3-15 months before diagnosis. In oncology, a weakly supervised deep learning model stratified lung adenocarcinoma recurrence risk from whole-slide pathology images with external validation, a Random Survival Forest model predicted early-stage NSCLC survival across a SEER and an independent Chinese cohort (C-index 0.74-0.78), and an XGBoost model reached over 90% accuracy on multimodal breast cancer survival prediction. Elsewhere, an XGBoost-and-SHAP nomogram detected PCOS with AUCs of 0.92 (internal) and 0.92 (external), and a GLM-based nomogram distinguished thrombocytopenia etiology with an AUC of 0.945.

Together, today's findings show LLM-based diagnostic and triage tools continuing to close the gap with clinician and supervised-model baselines on narrow tasks, while ML/DL risk and diagnostic models keep posting high discrimination (AUCs above 0.90) across a wide range of disease areas.

Sources

1. Conversational diagnostic AI (AMIE) feasibility study in primary care
2. LLMs as decision-support tools in oral pathology
3. Few-shot LLMs for actionable triage categorization
4. Generalist LLMs outperform clinical tools on medical benchmarks
5. Deep learning retinal imaging model for Alzheimer's risk factors
6. Deep learning model for young-onset type 2 diabetes risk
7. Weakly supervised deep learning recurrence model for lung adenocarcinoma
8. XGBoost and SHAP nomogram for PCOS detection

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