In prognostic, diagnostic, and risk-prediction models, several new studies reported strong gains over existing clinical tools. A JAMA Network Open study applied the deep learning model Mirai to mammograms from women screened between 2009 and 2018, finding it predicted 5-year breast cancer risk with an AUROC of 0.71, versus 0.53 for standard BI-RADS breast density assessment. In gastric cancer, a multimodal deep learning model (DeepComp) trained on 5,237 patients across 11 Chinese centers reached an AUC of 0.888 for predicting postoperative complications, outperforming nine established clinical risk scores and raising surgeons' sensitivity for flagging at-risk patients from 47.1% to 87.9%; a separate hybrid model combining gradient-boosted trees with a 1D CNN reached an AUC of 0.902 for postoperative gastric cancer survival. A diabetic retinopathy framework (TIMM-ProRS) combining a Vision Transformer, CNN, and graph neural network hit 97.8% accuracy on standard retinal imaging benchmarks, and at ESMO Congress 2026 researchers presented roughly 500 machine learning models for 20 cancer types, trained on over 460,000 UK Biobank participants and externally validated across four countries.
In large language models in healthcare, a Nature Medicine randomized trial of Google's AMIE system found blinded subspecialists preferred AMIE-assisted cardiology assessments over cardiologist-only assessments 46.7% to 32.7% of the time, with cardiologists alone producing nearly twice the rate of clinically significant errors (24.3% vs. 13.1%); participating cardiologists reported AMIE helped their assessment in 57% of cases. A JMIR simulation study found that confirmation-biased patient framing cut LLM diagnostic accuracy by 10-40 percentage points across six models and 1,273 cases, with prompt-based fixes giving only weak relief, though pairing a conversational model with a separate reasoning model recovered most of the lost accuracy. Two JMIR reviews took stock of the field more broadly: one of 35 clinical-decision-support studies found only a quarter reported external validation, and a second scoping review of 157 studies found retrieval-augmented and graph-retrieval-augmented LLM systems are now evaluated mostly on question answering (57%) and decision support (45%), but with inconsistent methodology.
In agentic AI for health research, a multi-agent oncology framework modeled on tumor board consultations — seven specialized LLM agents for roles from oncologist to patient advocate, reconciled via reinforcement learning — averaged 87.5% accuracy across five medical benchmarks versus 83.8% for the strongest single-agent baseline, with an 89.3% consensus rate among agents. A confidence-aware doctor-and-pharmacist agent pair (MedCoAct) improved both diagnostic and medication-recommendation accuracy by about 7 points over a single-agent system on a new 2,700-case benchmark. On literature screening, an ISPOR 2026 validation study of an agentic LLM system annotating 61,069 publications across five cancer types for a living systematic review reported accuracy of 93.7-99.8% against human reviewers.
Across today's studies, the pattern holds: models and agent systems that are explicitly validated against existing clinical benchmarks or multiple external cohorts — rather than judged on headline accuracy alone — are the ones producing numbers clinicians can act on.
Sources
- Deep learning model (Mirai) vs. breast density for breast cancer risk prediction
- DeepComp: multimodal model for predicting gastric cancer surgical complications
- TIMM-ProRS: multimodal diabetic retinopathy prognosis framework
- ML models for multicancer risk stratification (ESMO Congress 2026)
- A large language model for complex cardiology care (AMIE)
- Patient cognitive bias in LLM-supported health consultations
- Fine-tuning, RAG, and hybrid adaptation of LLMs for clinical decision-making: systematic review
- Multi-Agent Medical Decision Consensus Matrix System for oncology MDT consultations
- MedCoAct: confidence-aware multi-agent collaboration for clinical decisions
- Validation of an agentic LLM system in a living systematic literature review (ISPOR 2026)
Prepared with AI assistance from newly published research. See the sources for the original studies.



