Several new prediction models this week moved toward broader transferability rather than single-cohort accuracy. TRisk, a transformer survival model for heart failure, reached a concordance index of 0.845 for 36-month mortality versus 0.728 for the existing MAGGIC-EHR score, and carried over to a US cohort via transfer learning. GRASP used LLM-generated code embeddings to port UK Biobank-trained risk models across FinnGen and Mount Sinai despite different coding systems, and PRIME, a 19-predictor pancreatic cancer risk model trained on nearly 4.9 million patients across 23 health systems, reached a 36-month AUC of 0.75. A multimodal deep learning model combining ultrasound and digital breast tomosynthesis reached an AUC of 0.944 for breast-level risk classification, while a drift-aware transfer framework (ADAPT) targeted the more practical problem of keeping suicide-risk models accurate as coding systems and care patterns shift over time.

On the LLM side, evaluation results stayed mixed. DeepSeek R1 reached 87% accuracy on mechanical thrombectomy eligibility and 78% on thrombolysis decisions in acute ischemic stroke, outperforming other models and clinicians, but a new off-guideline benchmark found leading models correctly answering only 56% of rare-case questions unaided, rising to 82% with retrieval augmentation. A blinded radiology report comparison still favored radiologist-written lumbar spine MRI reports over ChatGPT-4o on structure and relevance, even though no clinically false statements were found in the AI reports.

Agentic systems continued to cluster around literature and workflow automation. A five-agent clinical-note screening workflow (released as the open-source Pythia framework) reached F1 0.93 for flagging cognitive concerns, and a structured multiagent framework improved stroke treatment recommendations by 18.9% over single-LLM baselines. Review-level evidence (a Methods and Protocols survey citing 87.3% accuracy in clinical trial matching) and literature-automation tools like InsightAgent, which reached 79.7% of human-written review quality in about 1.5 hours, point toward multi-agent orchestration being evaluated as much on auditability as on raw accuracy.

Across all three areas, the throughline is portability and validation: models and agents are increasingly being tested for whether they hold up outside their original setting, not just how well they perform on it.

Sources

  1. TRisk heart failure mortality transformer model
  2. PRIME pancreatic cancer risk stratification model
  3. Multimodal ultrasound + tomosynthesis breast risk classification
  4. DeepSeek R1 and LLMs for acute ischemic stroke treatment decisions
  5. OGCaReBench off-guideline clinical question-answering benchmark
  6. Autonomous agentic workflow for cognitive-concern detection (Pythia)
  7. Multiagent LLM framework for acute stroke treatment recommendation
  8. InsightAgent interactive multi-agent systematic review system

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

Share Facebook WhatsApp
End of articleBack to all writing →
Dr. Md Abdullah Saeed Khan

Physician, epidemiologist and researcher. DPhil fellow in Clinical Medicine at the University of Oxford.