Overview
- Mount Sinai’s graph neural network calculates patient-specific probabilities of stroke and major bleeding and provides a net-benefit treatment recommendation.
- The model was trained on electronic health records from 1.8 million patients spanning 21 million visits, 82 million clinical notes, and 1.2 billion data points.
- Internal validation included 38,642 atrial fibrillation patients within Mount Sinai and external validation used 12,817 patients from Stanford datasets.
- Compared with guideline-based tools, the system recommended against anticoagulation for about half of patients who would otherwise have been treated.
- The results were presented as a Late Breaking Science session at the European Society of Cardiology, and the authors stress the findings are observational pending prospective randomized trials.