AI in Healthcare: Research Roundup — September 24, 2026

On the LLM side, several new studies examined triage performance across specialties. A pilot study pitting four LLM chatbots against retina specialists on after-hours emergency calls found humans still more accurate overall (95% vs 76.7% diagnosis), though ChatGPT (90%) and Claude (83.3%) were statistically noninferior to human graders. A separate benchmarking study across five rare-disease datasets found LLM specialty-triage accuracy ranging widely from 0.44 to 0.71, while an evaluation of eight contemporary models on 48 clinical vignettes showed structured prompting lifting mean diagnostic accuracy from 89.84% to 91.67% and triage accuracy from 76.82% to 86.20%.

On broader evidence quality, a meta-research analysis of 11,628 PubMed records from 2023-2026 found medical LLM publications growing 45-fold, yet only 2.5% used randomized, controlled, or prospective designs, and the gap between a study's newest cited model and its own publication date widened from 1.33 to 6.08 quarters. A national survey of nearly 36,000 adults in China found moderate public acceptance of LLM-based healthcare (64.3/100), skewing lower among older adults and those with less digital-health experience. In a real-world multimodal evaluation across 539 inpatient cases, ten frontier LLMs all outperformed routine ward diagnoses on accuracy and safety scores.

On the ML/DL prognostic and diagnostic side, several nomogram studies validated risk-prediction tools: an invasive pulmonary aspergillosis model in ICU sepsis patients reached AUC 0.883/0.844 (training/validation), and a postherpetic neuralgia risk nomogram in herpes zoster patients reached AUC 0.758 with external validation. In imaging, a histology-based deep learning model for pituitary neuroendocrine tumors achieved AUC 0.912 for lineage classification across three independent cohorts, and a sequential-learning framework preserved diagnostic accuracy while adding prognostic capability for osteoarthritis, Alzheimer's disease, and breast cancer.

Across both threads, today's studies reinforce a consistent theme: LLMs are being pushed hardest on triage and diagnostic-reasoning tasks with mixed but improving results, while ML/DL prognostic and diagnostic models keep delivering well-validated, disease-specific risk scores ready for clinical use.

Sources

1. LLM triage of retina emergency telephone calls (pilot study)
2. Initial-visit specialty triage in rare diseases using LLMs
3. Diagnosis and triage performance of contemporary LLMs on clinical vignettes
4. The widening evaluation gap in medical LLM research, 2023-2026
5. Public acceptance of LLMs in healthcare in China (national survey)
6. Evaluating multimodal LLMs for inpatient diagnosis across ten frontier models
7. Nomogram for invasive pulmonary aspergillosis risk in ICU sepsis patients
8. Nomogram for postherpetic neuralgia risk in herpes zoster patients
9. Deep learning for pituitary neuroendocrine tumor lineage and subtypes
10. Sequential learning for robust disease prognosis via diagnostic knowledge preservation

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