Several new studies this week add to the evidence base on machine-learning prognostic and diagnostic models. A deep learning ensemble combining ConvNeXt, ResNet-50, and Swin Transformer architectures reached an AUC of 0.92 for melanoma detection from 3D total body photography, outperforming individual networks (0.86-0.88) and traditional ML baselines. A JMIR meta-analysis pooling 56 studies and roughly 887,000 participants found deep learning models reach a pooled AUC of 0.93 for COPD detection, though accuracy for grading disease severity was notably weaker and inconsistent. Not every model clears the bar: in a study of 4,135 Australian hospital admissions, plain logistic regression (ROC-AUC 0.64) matched or beat ensemble and deep learning approaches for predicting 30-day readmission, while a separate framework validated on roughly 415,000 MIMIC-IV admissions paired gradient-boosted trees with SHAP explainability and audited fairness across 16 demographic subgroups.

On large language models in healthcare, a randomized trial of a GPT-4o mini-based chatbot (PreA) across 2,069 patients cut specialist consultation time by 28.7% when used independently before visits. A separate preregistered RCT found a sharper gap between promise and practice: LLMs scored 94.9% and 56.3% on standalone medical benchmarks, but when 1,298 participants used them as real patients, accuracy fell below 34.5% and 44.2%. A comparison of AI-generated and physician-written discharge summaries found clinicians trusted 85.3% of LLM summaries versus 95.5% of physician-written ones, and a 157-study scoping review of RAG-based clinical LLMs noted that 89.2% of evaluations remain offline-only.

In agentic AI research, a systematic review of 52 studies found 96% documented system architecture but only 28% addressed accountability, underscoring a governance gap as these systems scale. On the technical side, a process-supervised multi-agent reinforcement learning approach raised agent-call accuracy from 0.105 to 0.455, and Stanford's Biomni, an autonomous biomedical research agent, matched human-expert performance on several of over 400 research tasks.

Together, these studies point to a field maturing from raw predictive capability toward harder questions of real-world generalization, trust, and accountability.

Sources

  1. Deep learning ensemble for melanoma detection via 3D total body photography
  2. Deep learning diagnostic accuracy for COPD: systematic review and meta-analysis
  3. ML models for 30-day general medicine readmission vs traditional approaches
  4. PreA LLM chatbot RCT for primary-to-specialist care transitions
  5. Preregistered RCT on LLM reliability as medical assistants for the public
  6. Scoping review of RAG and GraphRAG LLM evaluation in healthcare
  7. Governing Agentic AI in Healthcare: systematic review
  8. Biomni: autonomous biomedical research agent

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.