Paper to Practice

Unpack the methods behind each paper. Understand how they work and where they fall short before considering how to apply them.

Paper to Practice · 2026-10-14

External Evaluation of 3 Commercial Artificial Intelligence Algorithms for Independent Assessment of Screening Mammograms

An independent team in Stockholm ran 3 commercial mammography AI algorithms blind on 8805 screening exams (739 cancers, 8066 healthy controls). The best one (AI-1) reached an AUC of 0.956 and, at the first readers' specificity of 96.6%, a sensitivity of 81.9% versus 77.4% for first-reader radiologists; pairing AI-1 with the first reader found the most cancers (88.6% sensitivity) but at a lower specificity of 93.0%.

Read the full note ↗

External Evaluation of 3 Commercial Artificial Intelligence Algorithms for Independent Assessment of Screening Mammograms

Paper to Practice · 2026-09-19

Learning the natural history of human disease with generative transformers

Delphi-2M adapts the GPT architecture to health records: trained on 402,799 UK Biobank participants, it reads a person's diagnosis history, predicts the rates and timing of 1,256 diseases plus death, and can simulate future health trajectories. It reached an average AUC of 0.76 internally and 0.67 on 1.93 million Danish registry records without retraining.

Read the full note ↗

Learning the natural history of human disease with generative transformers