On the prognostic and diagnostic modeling side, a new framework unifies survival analysis and classification into one early-risk model for five chronic conditions (diabetes, hypertension, CKD, COPD, ischemic heart disease) from EMR data. A breast cancer model combining clinical, transcriptomic, and copy-number features showed consistent discrimination across age, receptor status, and menopausal subgroups. A temporal point process model, PatientTPP, trained on 500,000+ clinical trajectories outperformed BMI at flagging future high-cardiovascular-cost patients, with its top risk quartile capturing 39.3% of costs versus 22.7% for BMI. A study of over 605,000 patients found SMOTE and other resampling methods gave no meaningful discrimination gains and worsened calibration. In Alzheimer's research, a routine-data model externally validated on 741 ADNI trajectories predicted 12-month cognitive and functional decline from baseline assessments alone, with no imaging needed.
On the LLM side, ChatGPT 5 reached 82.1% accuracy automating TNM cancer staging from 552 PET-CT reports, edging out junior radiologists (77.0%) and outpacing ChatGPT 4o (74.3%) and 3.5 (59.6%). A JMIR evaluation found Qwen 3 detected radiology report errors more accurately than radiologists (0.96 vs. 0.45-0.66) and about ten times faster. A blinded hypertension-vignette comparison found Claude Sonnet 4.6 ranked first in nine of ten cases against ChatGPT-5.2 and Gemini 3.0 Pro, with the widest gaps in complex clinical reasoning. A review of 20 studies comparing LLM answers to patient questions against healthcare professionals found LLMs favored in 12, mixed results in 6, professionals favored in only 2.
On agentic AI, Latent Labs' Latent-Y agent ran autonomous antibody design campaigns end-to-end from a text prompt, producing lab-confirmed nanobody binders against six of nine targets (67% success) at single-digit nanomolar affinity, with experts reporting campaigns roughly 56 times faster than manual estimates. A new ACL 2026 survey frames "AI hospitals" as multi-agent clinical systems with defined roles and safety gates, proposing Integration Readiness Levels for future evaluation. A related architecture paper catalogues agentic roles for hospital information systems, piloted so far only on synthetic data.
Today's throughline: the strongest results come from narrowly scoped, well-validated systems (staging, report QA, antibody design), while broader claims about resampling, governance, and hospital-wide agent orchestration still rest on thinner or synthetic evidence.
Sources
- Survival Meets Classification: A Novel Framework for Early Risk Prediction Models of Chronic Diseases
- Patient foundation model for risk stratification in low-risk overweight patients (PatientTPP)
- Tipping the Balance: Impact of Class Imbalance Correction on the Performance of Clinical Risk Prediction Models
- Evaluating large language models for automated TNM staging from PET-CT reports: a multi-cancer comparative study
- Large Language Model-Assisted Radiology Report Quality Assurance: Real-World Multicenter Evaluation
- Large language models versus healthcare professionals in providing medical information to patient questions
- Latent-Y: A Lab-Validated Autonomous Agent for De Novo Drug Design
- LLM-Based Multi-Agent Systems for Clinical Workflows: A Survey of AI Hospitals
Prepared with AI assistance from newly published research. See the sources for the original studies.
Physician, epidemiologist and researcher. DPhil fellow in Clinical Medicine at the University of Oxford.
