Several new studies this week add to the evidence base for machine learning risk and prognostic models. The PERISCOPE project trained mortality-prediction models on a US registry of 38,489 patients, tested them on a separate 119,035-patient dataset, and externally validated them on 6,776 patients in Amsterdam, finding the models outperformed the established GRACE 2.0 score for predicting death after a heart attack across all three cohorts. In Alzheimer's disease, a dual ANN-CNN framework combined clinical data (87.08% accuracy for early risk prediction) with imaging data (97% accuracy for staging), while a Kaiser Permanente model built on EHR data from more than 3 million adults estimated 1-, 3-, and 10-year type 2 diabetes risk with near-ideal calibration, though a full peer-reviewed paper is still pending.
On the LLM side, a cross-sectional benchmark found Gemini 3 Pro reached 89.9% accuracy on bilingual radiology questions, outperforming four other multimodal models, though all models did worse on image-based than text-based items. A safety evaluation of six public chatbots answering COPD questions found every model produced some potentially unsafe advice, with several safety classifications shifting when the test was repeated months later. A randomized trial in rheumatology found an LLM-based diagnostic tool cut diagnosis time by more than half without improving diagnostic accuracy, and noted a tendency toward physician over-reliance.
In agentic AI, El Agente Fármaco, an autonomous drug-discovery agent, proposed chemically distinct, competitively scored candidate molecules when benchmarked against an expert-led medicinal chemistry campaign. Separately, Nature published Robin, a multi-agent system that closed a full discovery loop and identified an approved glaucoma drug as a repurposing candidate for dry age-related macular degeneration. A clinical-workload study also found that orchestrator-based multi-agent systems held accuracy near 90% at scale where single-agent approaches collapsed, while using far fewer tokens.
Across these studies, a consistent theme is that newer architectures show real gains on narrow, well-defined tasks, but safety, reproducibility, and external validation remain the open questions before broader clinical use.
Sources
- PERISCOPE project: ML models outperform GRACE 2.0 for post-heart-attack mortality
- Dual ANN-CNN framework for Alzheimer's risk prediction and staging
- Kaiser Permanente EHR-based type 2 diabetes risk model (>3 million adults)
- Bilingual radiology LLM benchmark (Gemini 3 Pro, GPT-5.2, and others)
- Safety evaluation of six LLMs answering COPD questions
- RCT of an LLM diagnostic tool in rheumatology
- El Agente Fármaco, an autonomous drug-discovery agent
- Robin: a multi-agent system for automating scientific discovery (Nature)
- Orchestrated multi-agent system sustains accuracy under clinical-scale workloads
Prepared with AI assistance from newly published research. See the sources for the original studies.



